<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[machinevision - Medium]]></title>
        <description><![CDATA[Machine Learning, Computer Vision, and Augmented Reality techniques explored - Medium]]></description>
        <link>https://medium.com/machinevision?source=rss----cc93edaeae9f---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>machinevision - Medium</title>
            <link>https://medium.com/machinevision?source=rss----cc93edaeae9f---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 21:28:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/machinevision" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Implement SLAM from scratch]]></title>
            <link>https://medium.com/machinevision/implement-slam-from-scratch-b1fb599f40c8?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/b1fb599f40c8</guid>
            <category><![CDATA[mapping]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Tue, 23 Apr 2024 18:44:20 GMT</pubDate>
            <atom:updated>2025-01-14T04:35:43.836Z</atom:updated>
            <content:encoded><![CDATA[<p>There are many ways to implement a solution for SLAM (Simultaneous Localization and Mapping), but the simplest algorithm to implement is Graph SLAM. <em>(If you need a refresher on SLAM and what it’s for, read my </em><a href="https://medium.com/machinevision/overview-of-slam-50b7f49903b7"><em>Overview of SLAM</em></a><em>)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z-r3ye1mudN_n4UvfOWUXA.png" /><figcaption>This is the trajectory of the robot over three time steps. Nearby, there’s a landmark (e.g., tree).</figcaption></figure><p>Graph SLAM is a technique used in robotics to simultaneously estimate a robot’s trajectory over time and estimate the positions of landmarks in the environment by representing them as nodes and constraints in a graph. The graph consists of nodes representing robot poses at different time steps and landmark positions, and edges representing the constraints between them, such as initial location, relative motion, and relative measurement constraints. By optimizing the graph, Graph SLAM aims to find the most likely trajectory and landmark positions that best explain the sensor measurements.</p><h4>Example</h4><p>Before we dive into Graph SLAM, I’d like to illustrate an example that will help us with the discussion as we continue to explore how to implement Graph SLAM. In this example, there is a robot and the robot is traversing in a one dimensional world. The robot’s first pose is at time step t0, where the robot pose is x=2. From this pose, the robot sees a landmark L0 (e.g., tree) and it’s 9 units away. Then, the robot moves forward by 5 units. At this point the robot should be at x=7 and the landmark should be 4 units away; however, the robot does not see/measure the distance to the landmark at time step t1. The lack of landmark measurement at time step t1 could be due to sensor error, occlusions, or something else. Finally, the robot moves 3 units forward and sees the landmark 1 unit away. At this point, the robot should be at x=10 and the landmark should be at x=11.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*ZklqO_mgJ3-1o-F3bU_exw.png" /><figcaption>This is a graph representation of robot poses and landmarks in a one dimensional space. First, the robot sees a landmark 9 units away. Second, the robot moves forward 5 units. Finally, the robot moves 3 units forward and sees the same landmark 1 unit away.</figcaption></figure><h4>Constraints</h4><p>In Graph SLAM, there are three important types of constraints. Let’s dive deeper into each of these constraints:</p><ol><li><strong>Initial Location Constraint:</strong> The initial location constraint represents our knowledge about the robot’s initial position in the environment. It provides a fixed reference point for the robot’s trajectory estimation. This constraint can be represented as <em>&lt;x,y,θ&gt;</em> to capture the initial position and orientation estimate. By incorporating this constraint into the graph, we can anchor the trajectory estimation and provide a starting point for further constraint optimization.</li><li><strong>Relative Motion Constraints:</strong> Relative motion constraints capture the information about how the robot’s pose changes between consecutive time steps. These constraints are typically obtained from odometry sensors, such as wheel encoders or IMUs. Odometry sensors provide estimates of the robot’s motion, such as the change in position and orientation. By comparing the odometry measurements between consecutive time steps, we can derive relative motion constraints that describe the robot’s movement. These constraints are typically represented as Gaussian distributions that capture the uncertainty in the motion estimates.</li><li><strong>Relative Measurement Constraints:</strong> Relative measurement constraints capture the information about the relative positions or distances between different landmarks or features in the environment. These constraints are obtained from sensor measurements, such as laser range finders or cameras. For example, if the robot observes a landmark and measures the distance to the landmark, this information can be used as a relative measurement constraint. These constraints help in estimating the positions of landmarks relative to the robot’s trajectory.</li></ol><p>In Graph SLAM, all of these constraints are used together to build a graph representation of the environment and the robot’s trajectory. The graph consists of nodes representing robot poses at different time steps and landmark positions, and edges representing the constraints between them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*OHZJdtSFFnZ3dfE07V1Djw.png" /><figcaption>This is the pure graph representation. The relative motion constraints are defined by the t0-t1 and t1-t2 edges. The relative measurement constraints are defined by the t0-L0 and t2-L0 edges.</figcaption></figure><p>In our example, we have 5 total constraints: 1 initial location constraint, 2 relative motion constraints, and 2 relative measurement constraints. The 4 relative constraints are represented by the edges in the graph.</p><h4>Graph Representation</h4><p>In Graph SLAM, the environment and the robot’s trajectory are represented using a graph structure. The graph consists of nodes and edges, where nodes represent the robot’s poses (positions and orientations) at different points in time, and edges represent the constraints or measurements between these poses.</p><p>In addition to representing the robot’s poses, the graph also includes nodes that represent the landmarks or features in the environment. These landmarks can be objects, points of interest, or any other distinctive features that the robot can perceive and use for localization and mapping.</p><p>The graph representation connects the robot’s poses and landmarks through edges, which represent the measurements or constraints obtained from sensors. These measurements can include range measurements, bearing measurements, or any other type of sensor data that provides information about the relative positions and orientations of the robot and the landmarks.</p><p>For example, let’s say the robot observes a landmark from its current pose. This observation creates an edge between the robot’s pose node and the landmark node in the graph. The edge represents the measurement obtained from the sensor, which provides information about the relative position and orientation of the robot with respect to the landmark.</p><p>By incorporating these measurements into the graph, the SLAM algorithm can estimate the most likely trajectory of the robot and the map of the environment that best satisfies the constraints imposed by the measurements. The graph optimization process involves adjusting the poses and landmark positions in the graph to minimize the errors between the predicted measurements and the actual measurements obtained from the sensors.</p><h4>Matrix and Vector Representation</h4><p>In Graph SLAM, we use matrix and vector representations to model the relationships between robot poses and landmarks in a map. These representations help us solve the SLAM problem.</p><p>Let’s start with the matrix representation. In Graph SLAM, we create a matrix called the information matrix, denoted as Ω (Omega). This matrix represents the constraints or relationships between different variables in the SLAM problem. Each variable corresponds to a robot pose or a landmark in the map.</p><p>The information matrix is a square matrix, and its size depends on the number of variables in the SLAM problem. A Graph SLAM problem with <em>n </em>nodes in the graph, would have an <em>n x n</em> information matrix. The elements of this matrix encode the information about the relationships between variables. For example, if two variables are highly correlated, the corresponding elements in the information matrix will have higher values.</p><p>Now, let’s move on to the vector representation. In Graph SLAM, we also create a vector called the information vector, denoted as ξ (Xi). This vector represents the measurements or observations we have made in the SLAM problem. Each element of the vector corresponds to a specific measurement or observation. A Graph SLAM problem with <em>n</em> nodes in the graph, would have an <em>n x 1</em> information vector.</p><p>The information vector contains the information about the measurements and their relationships with the variables in the SLAM problem. It helps us incorporate the measurements into the SLAM problem and update our estimates of the robot poses and landmarks.</p><p>In our example, we would initialize a <em>4 x 4</em> matrix, since there are <em>4</em> nodes in the graph. This is the information matrix:</p><pre>// 4x4 matrix filled with zeroes<br>     + -- + -- + -- + -- +<br>     | t0 | t1 | t2 | L0 |<br>+ -- + -- + -- + -- + - -+<br>| t0 | 0  | 0  | 0  | 0  |<br>| t1 | 0  | 0  | 0  | 0  |<br>| t2 | 0  | 0  | 0  | 0  |<br>| L0 | 0  | 0  | 0  | 0  |</pre><p>In our example, we would also initialize a <em>4 x 1</em> vector, since there are <em>4</em> nodes in the graph. This is the information vector:</p><pre>// 4x1 vector filled with zeroes<br>+---+<br>| 0 |<br>| 0 |<br>| 0 |<br>| 0 |<br>+---+<br></pre><h4>Graph SLAM Algorithm</h4><p>Once we declared our information matrix and vector, then we need to apply the initial location constraint to the matrix and vector. For example, to update the information matrix with an initial location of 2, we would create a simple linear equation:</p><blockquote>t0 = 2</blockquote><p>Now, let’s take our simple linear equation and its coefficients <em>&lt;1,0,0,0;2&gt;</em>, and add it the row corresponding to <em>t0</em>:</p><pre>// Omega matrix (result)<br>     + -- + -- + -- + -- +<br>     | t0 | t1 | t2 | L0 |<br>+ -- + -- + -- + -- + - -+<br>| t0 | 1  | 0  | 0  | 0  |<br>| t1 | 0  | 0  | 0  | 0  |<br>| t2 | 0  | 0  | 0  | 0  |<br>| L0 | 0  | 0  | 0  | 0  |</pre><pre>// Xi vector (result)<br>+---+<br>| 2 |<br>| 0 |<br>| 0 |<br>| 0 |<br>+---+</pre><p>To generalize this, here’s some initial pseudo-code:</p><pre>void GraphSLAM(G, startPose) {<br>    // Declare Omega and Xi<br>    Omega = new Matrix(n,n)<br>    Xi = new Vector(n)<br><br>    // Initial Location Constraint<br>    Omega[&#39;t0&#39;,&#39;t0&#39;] = 1<br>    Xi[&#39;t0&#39;] = startPose<br><br>    // Graph Optimization<br>    Mu = GraphOptimization(Omega, Xi, G)<br>    return Mu<br>}</pre><p>From here, we need to discuss Graph Optimization.</p><h4>Graph Optimization</h4><p>Once we have have the initial matrix and vector representations of the graph, we need to perform graph optimization. Graph optimization in Graph SLAM is the process of refining the estimates of robot poses and landmark positions by iteratively updating the graph based on sensor measurements. It involves two main steps: the measurement update and the state update.</p><ol><li><strong>Measurement Update</strong>: In the measurement update step, we iterate over the edges of the graph and add constraints into the information matrix. These constraints represent the measurements obtained from sensors, such as range measurements or bearing measurements.</li><li><strong>State Update</strong>: In the state update step, we solve a system of linear equations to estimate the optimal robot poses and landmark positions that minimize the error in the graph. This is done by taking the inverse of the information matrix and multiplying it by the information vector.</li></ol><p>Here is some high-level pseudo-code for Graph Optimization:</p><pre>void GraphOptimization(Omega, Xi, G) {<br>    Omega, Xi = MeasurementUpdate(Omega, Xi, G);<br>    Mu = StateUpdate(Omega, Xi);<br>    return Mu;<br>}</pre><p>Let’s go into more detail for the measurement and state updates.</p><h4>Measurement Update</h4><p>In the Measurement Update, it uses the Graph data to define the information matrix and vector data. Omega is the information matrix that represents the coefficients of the linear equations, Xi is the information vector that represents the constant terms in those equations, and G is the graph that contains the edge weights representing the measurements (e.g., distance).</p><p>Each of the edges helps to define a pair of linear equations. For example, to update the information matrix with the edge <em>t0-t1</em>, we would create two linear equations:</p><blockquote><em>Lin Eqn 1: t0 - t1 = -5</em></blockquote><blockquote><em>Lin Eqn 2: t1 - t0 = 5</em></blockquote><p>Now, let’s take the first linear equation and its coefficients <em>&lt;1,-1,0,0;-5&gt;</em>, and add it the row corresponding to <em>t0</em>:</p><pre>// Omega matrix (result)<br>     + -- + -- + -- + -- +<br>     | t0 | t1 | t2 | L0 |<br>+ -- + -- + -- + -- + - -+<br>| t0 | 2  |-1  | 0  | 0  |<br>| t1 | 0  | 0  | 0  | 0  |<br>| t2 | 0  | 0  | 0  | 0  |<br>| L0 | 0  | 0  | 0  | 0  |</pre><pre>// Xi vector (result)<br>+---+<br>|-3 |<br>| 0 |<br>| 0 |<br>| 0 |<br>+---+</pre><p>Now let’s take the coefficients for the second linear equation <em>&lt;-1,1,0,0;5&gt;</em> and add it to the row corresponding to <em>t1</em>:</p><pre>// Omega matrix (result)<br>     + -- + -- + -- + -- +<br>     | t0 | t1 | t2 | L0 |<br>+ -- + -- + -- + -- + - -+<br>| t0 | 2  |-1  | 0  | 0  |<br>| t1 |-1  | 1  | 0  | 0  |<br>| t2 | 0  | 0  | 0  | 0  |<br>| L0 | 0  | 0  | 0  | 0  |</pre><pre>// Xi vector (result)<br>+---+<br>|-3 |<br>| 5 |<br>| 0 |<br>| 0 |<br>+---+</pre><p>To see how this generalizes, let’s show you the pseudo-code:</p><pre>void MeasurementUpdate(Omega, Xi, G) {<br>    for each edge:<br>        // LinEq 1: src - dst = -weight<br>        Omega[edge.src, edge.src] += 1<br>        Omega[edge.src, edge.dst] += -1<br>        Xi[edge.src] += -edge.weight<br>        // LinEq 2: dst - src = weight<br>        Omega[edge.dst, edge.dst] += 1<br>        Omega[edge.dst, edge.src] += -1<br>        Xi[edge.dst] += edge.weight<br>    return Omega, Xi<br>}</pre><p>As you can see, the Measurement Update process involves iterating over each edge in the graph G. For each edge, the code performs two steps. First, it updates the Omega and Xi values for the row corresponding to the source node of the edge. Second, it updates the Omega and Xi values for the row corresponding to the destination node of the edge. In both steps, the edge weight represents the measurement between both nodes (e.g., distance).</p><p>These steps ensure that the constraints imposed by the measurements are correctly represented in the Omega and Xi matrices. After iterating over all the edges, the function returns the updated Omega and Xi matrices.</p><h4>State Update</h4><p>At this point, we have fully defined Omega and Xi. So we just need to solve a system of equations to solve for Mu:</p><blockquote>Omega * Mu = Xi,</blockquote><p>where Mu represents the updated state estimate. To solve for Mu, we merely need to invert Omega:</p><blockquote>Mu = Omega^{-1} * Xi</blockquote><pre>void StateUpdate(Omega, Xi) {<br>    Mu = Omega.invert() * Xi<br>    return Mu<br>}</pre><p>The provided pseudo code performs the state update by multiplying the inverse of the covariance matrix (Omega) with the measurement vector (Xi). The result, Mu, represents the state estimate of the robot’s pose and landmark positions. It is a vector that contains the values of all the variables that define the state of the system. In our example, this is the expected value of Mu:</p><pre>// Mu vector (result)<br>+----+<br>|  2 |<br>|  7 |<br>| 10 |<br>| 11 |<br>+----+</pre><p>In this example, the first element is the location of the robot at t0, the second element is the location of the robot at t1, the third element is the location of the robot at t2, and the fourth element is the location of the landmark (L0).</p><h4>Discussion</h4><p>Note that the Graph SLAM algorithm might not give you exactly the correct answer, but it will be close to it. The result of the SLAM algorithm depends on the quality of measurements that you feed the algorithm.</p><p>Also note that this was just a simple one-dimensional example. This can be easily extended into 3-dimensional space and also include additional directional dimensions (to explain which direction the robot is looking at).</p><p>If any SLAM jargon did not make sense, please review my <a href="https://medium.com/machinevision/overview-of-slam-50b7f49903b7">Overview of SLAM</a>.</p><p><em>If you enjoyed this article, please </em><a href="http://heartsymbol.love/symbols/heart">❤</a><em> to help others find it!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b1fb599f40c8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/implement-slam-from-scratch-b1fb599f40c8">Implement SLAM from scratch</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Overview of SLAM]]></title>
            <link>https://medium.com/machinevision/overview-of-slam-50b7f49903b7?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/50b7f49903b7</guid>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[mapping]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Tue, 16 Apr 2024 22:18:45 GMT</pubDate>
            <atom:updated>2024-04-23T18:49:51.099Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_pnmBA0lo-X9Uf4iSkN8sw.png" /><figcaption>SLAM can be used to create a map of the world. [Credit: Firefly]</figcaption></figure><p>SLAM stands for Simultaneous Localization and Mapping. It is a technique used in robotics to solve the problem of building a map of an unknown environment while simultaneously localizing the robot within that map.</p><p>The main goal of SLAM is for a robot to navigate and explore an unknown environment, create a map of that environment, and at the same time, determine its own position within that map. This is done in real-time, without any prior knowledge of the environment.</p><p>SLAM is a fundamental problem in robotics because it enables robots to operate autonomously in unknown or dynamic environments. By using SLAM, robots can navigate, explore, and perform tasks in environments where pre-built maps are not available or may be outdated.</p><p>SLAM is used in various applications, including autonomous vehicles, drones, mobile robots, and even augmented reality systems. For example, in autonomous vehicles, SLAM allows the vehicle to navigate and build a map of its surroundings, enabling it to plan safe and efficient paths. In drones, SLAM helps in mapping the environment and maintaining stable flight. In augmented reality systems, SLAM is used to overlay virtual objects onto the real world accurately.</p><p>To achieve SLAM, robots typically use a combination of sensor data, such as cameras, lidars, or range finders, to perceive the environment. They also utilize algorithms that process this sensor data, estimate the robot’s position, and update the map as the robot moves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FhrX-4UgWyfwt32HbD4lXA.png" /><figcaption>A robot uses its sensors to detect its surrounding landmarks. It’s able to create a map based on its sensor measurements, and it’s also able to identify its location on the map. [Credit: Firefly]</figcaption></figure><h4>What is the necessary vocabulary to understand SLAM?</h4><p>In the context of SLAM (Simultaneous Localization and Mapping), there are several key terms and concepts that are important to understand. Let’s break them down:</p><ol><li><strong>Localization:</strong> Localization refers to the estimated poses of the robot. It contains the robot’s position and orientation at different time steps. These poses are typically represented as (x, y, theta) coordinates, where (x, y) represents the position and theta represents the orientation.</li><li><strong>Mapping:</strong> Mapping refers to the estimated positions of the landmarks in the environment. Landmarks are distinct features in the environment that the robot can perceive, such as walls, corners, or objects.</li><li><strong>Robots:</strong> Robots are the autonomous agents that are equipped with sensors and actuators to interact with the environment. In the context of SLAM, the robot’s main task is to explore the environment, collect sensor data, and estimate its own position and the positions of landmarks. In the context of augmented reality (AR) systems on smartphones, there is typically no physical robot involved; instead, the smartphone itself can be considered as the “robot”.</li><li><strong>Landmarks:</strong> Landmarks are distinct features or points of interest in the environment that the robot can perceive. They can be objects, corners, walls, or any other feature that the robot can use to navigate and localize itself. Landmarks serve as reference points for the robot’s position estimation.</li><li><strong>Sensors:</strong> Sensors are the devices that the robot uses to perceive the environment. They provide information about the robot’s surroundings, such as range measurements, images, or depth data. Common sensors used in SLAM include range finders (such as laser range finders or ultrasonic sensors), cameras, lidars, and odometry sensors. In the context of a smartphone as a “robot”, the smartphone acts as a platform that integrates various sensors, such as cameras, gyroscopes, accelerometers, and depth sensors, to perceive the real-world environment.</li></ol><p>In addition to these terms, there are a few other vocabulary words that are commonly used in SLAM:</p><ul><li><strong>Odometry:</strong> Odometry refers to the estimation of the robot’s motion based on its internal sensors, such as wheel encoders, accelerometers, or gyroscopes. It provides information about the robot’s velocity, rotation, and displacement.</li><li><strong>Loop Closure:</strong> Loop closure is the process of detecting and correcting loops in the robot’s trajectory. It occurs when the robot revisits a previously observed landmark or location. Loop closure is important for correcting accumulated errors and improving the accuracy of the SLAM solution.</li></ul><p>These are some of the key terms and jargon used in SLAM. Understanding these concepts will help you navigate through the SLAM algorithms and techniques.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7MDhJGcVqhDxy2cf3hziXQ.png" /><figcaption>SLAM can also be used for Augmented Reality. [Credit: Pexels]</figcaption></figure><h4>What are the use cases for SLAM?</h4><p>SLAM technology is becoming increasingly prevalent in our daily lives, even though we may not always be aware of it. Here are a few examples of SLAM applications that we encounter in our daily lives:</p><ol><li><strong>GPS Navigation Systems:</strong> Many outdoor GPS navigation systems utilize SLAM techniques to provide accurate and real-time positioning information. These systems combine GPS data with other sensor inputs, such as accelerometers and gyroscopes, to estimate the vehicle’s position and orientation. Examples include Google Maps, Waze, and Apple Maps.</li><li><strong>Mobile Augmented Reality (AR) Apps:</strong> AR apps on smartphones and tablets often employ SLAM algorithms to track the device’s position and overlay virtual objects onto the real world. SLAM helps to precisely align virtual content with the user’s surroundings, creating immersive AR experiences. Examples include Pokémon Go, SnapChat, and Instagram.</li><li><strong>Autonomous Vacuum Cleaners:</strong> Robotic vacuum cleaners, such as Roomba, utilize SLAM to navigate and clean a room efficiently. They build a map of the environment and use localization techniques to determine their position within that map, enabling them to autonomously navigate and avoid obstacles.</li><li><strong>Indoor Navigation Systems:</strong> SLAM is used in indoor navigation systems to provide real-time location information within large buildings like shopping malls, airports, or museums. By using sensors like cameras or depth sensors, SLAM algorithms can map the indoor environment and assist users in navigating through complex spaces. One example of SLAM-based indoor navigation is Google’s “Indoor Maps” feature, which is available in their Google Maps application. It uses SLAM alongside Wi-Fi, Bluetooth, and other sensor data to provide real-time location information and directions within these buildings.</li><li><strong>Self-driving Cars:</strong> SLAM is a fundamental technology for autonomous vehicles. Self-driving cars use various sensors, such as LiDAR, cameras, and radar, to perceive their surroundings and build a detailed map of the environment. SLAM algorithms help the vehicle localize itself within the map and navigate safely. Examples include Waymo, Tesla, Cruise, and Uber.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7Zj1KLBx1U9fUZk6WT83Qw.png" /><figcaption>Robotic vacuum cleaners, such as Roomba, utilize SLAM to navigate and clean a room efficiently. [Credit: Pexels]</figcaption></figure><p>These are just a few examples of how SLAM technology is integrated into our daily lives. As the field of robotics and AI continues to advance, we can expect to see even more applications of SLAM in various domains.</p><p>Looking for an implementation? This is <a href="https://medium.com/machinevision/implement-slam-from-scratch-b1fb599f40c8">The Simplest SLAM Algorithm</a>.</p><p><em>If you enjoyed this article, please </em><a href="http://heartsymbol.love/symbols/heart">❤</a><em> to help others find it!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50b7f49903b7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/overview-of-slam-50b7f49903b7">Overview of SLAM</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Overview of MLOps]]></title>
            <link>https://medium.com/machinevision/overview-of-mlops-a07053fc2a80?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/a07053fc2a80</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Tue, 12 Mar 2024 18:09:29 GMT</pubDate>
            <atom:updated>2024-03-12T18:09:29.665Z</atom:updated>
            <content:encoded><![CDATA[<p>As Machine Learning (ML) models are increasingly incorporated into software, a nascent subfield called MLOps (short for ML Operations) has emerged. MLOps is a set of practices that aim to deploy and maintain ML models in production reliably and efficiently. It’s widely agreed that MLOps is hard. At the same time, it’s unclear why MLOps is hard. In this article, we aim to bring clarity to MLOps, specifically in identifying what MLOps typically involves — across organizations and ML applications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hjIkNQNq2C_97U5X1WtxcA.png" /><figcaption>The most common MLOps tasks (Left). The core tech stack for MLOps (Right).</figcaption></figure><h3>MLOps Stack</h3><p>First, we take a look at the MLOps tech stack that Machine Learning Engineers (MLEs) need to work with to develop the machine learning models and deploy them. These are the four layers of the MLOps stack.</p><ol><li><strong>Run Layer:</strong> This layer manages hyper-parameters, data, and experiments. A run is a record of an execution of an ML or data pipeline (and its components). Run-level data is often managed by data catalogs, model registries, and training dashboards. Examples include Weights &amp; Biases, MLFlow, Hive metastores, AWS Glue.</li><li><strong>Pipeline Layer:</strong> This layer is designed to facilitate the development, training, deployment, and management of ML models. Finer-grained than a run, a pipeline further specifies the dependencies between artifacts and details of the corresponding computations. Pipelines can run ad-hoc or on a schedule. Pipelines change less frequently than runs, but more frequently than components. Examples include Sagemaker, Papermill, DBT, Airflow, TensorFlow Extended.</li><li><strong>Component Layer:</strong> This layer allows you to create the ML model, including model training, feature generation and selection. A component is an individual node of computation in a pipeline, often a script inside a managed environment. Some ML teams have an organization-wide “library of common components” for pipelines to use, such as feature generation and model training. Examples include PyTorch, TensorFlow, Python, Spark.</li><li><strong>Infrastructure Layer:</strong> This layer provides computing resources to store and use the ML model. It provides the underlying infrastructure for running ML pipelines. It also includes cloud storage (e.g., S3) and GPU-backed cloud computing (AWS, GCP). It changes less frequently than other layers but with more significant consequences. Examples include AWS, Google Cloud Platform (GCP), Docker.</li></ol><p>Each layer serves a specific purpose in managing the ML lifecycle, from execution and dependency management to computation and infrastructure provision. As the stack gets deeper, changes become less frequent: training jobs run daily but Dockerfiles are modified only occasionally.</p><h3>MLOps Tasks</h3><p>Across the MLOps tech stack, successful Machine Learning Engineers (MLEs) are needed to operationalize Machine Learning. The process of operationalizing Machine Learning consists of a continual loop of these four routine tasks.</p><ol><li><strong>Data Collection:</strong> This involves gathering data from various sources, organizing it into a centralized repository, and cleaning it. Labeling data involves assigning correct labels or annotations to the data points, which can be done either in-house or outsourced to annotators.</li><li><strong>Model Experimentation:</strong> ML engineers focus on improving the performance of machine learning models by experimenting with different features and model architectures. This can include creating new features, modifying existing ones, or changing the model architecture itself. The performance of these experiments is typically measured using metrics such as accuracy or mean-squared-error.</li><li><strong>Model Evaluation:</strong> Once a model is trained, it needs to be evaluated to ensure its performance meets the desired standards. This involves computing metrics such as accuracy on labeled data points that were not seen during training. If the model performs satisfactorily, it can then be deployed into production. Deployment involves careful review of the proposed changes, possibly staging them to a portion of the user base, conducting A/B testing, and maintaining records of changes for potential rollbacks.</li><li><strong>Model Maintenance:</strong> After deployment, it’s crucial to monitor the ML pipelines for any issues or anomalies. This involves tracking live metrics through queries or dashboards, investigating the quality of predictions for different sub-populations, and addressing any failures or bugs that arise. This might include patching the model with non-ML heuristics for known failure modes and updating the evaluation set with real-world failures.</li></ol><p>The next three sections discuss common strategies from ML Engineers for successful Model Experimentation, Evaluation, and Maintenance.</p><h3>Model Experimentation</h3><p>ML Engineering is very experimental and iterative in nature, especially compared to software engineering. Many experiments don’t ever make it into production. It’s important that experiment ideas can be prototyped and validated quickly. These are some common strategies that ML engineers use to generate successful experiment ideas.</p><ol><li><strong>Collaboration and Idea Validation:</strong> Good project ideas often stem from collaboration with domain experts, data scientists, and analysts who have conducted exploratory data analysis. Collaborative discussions, asynchronous communication, and cross-team cooperation help validate and refine project ideas.</li><li><strong>Iterating on Data:</strong> ML engineers emphasize the importance of iterating on the data rather than solely focusing on the model. Experimentation often involves adding new features, improving feature engineering pipelines, and iterating on the dataset to enhance model performance.</li><li><strong>Accounting for Diminishing Returns:</strong> ML projects typically undergo staged deployment, where ideas are validated offline and gradually deployed to production traffic. Engineers prioritize experiment ideas with the largest performance gains in the earliest stages of deployment, considering the diminishing returns of later stages.</li><li><strong>Preferential Treatment for Small Changes:</strong> ML engineers follow software best practices by making small, incremental changes to the codebase. Changes are kept small to facilitate faster code review, easier validation, and fewer merge conflicts. Config-driven development is emphasized to minimize bugs and ensure reproducibility.</li></ol><p>These strategies collectively contribute to the success of ML experiments by fostering collaboration, optimizing data utilization, prioritizing high-impact ideas, and maintaining code quality and reproducibility.</p><h3>Model Evaluation</h3><p>Model evaluation efforts need to keep up with changes in data and business requirements. It’s important to keep bad models from making it to production while maintaining velocity. These are some common organizational efforts to effectively evaluate models.</p><ol><li><strong>Dynamic Validation Datasets:</strong> Engineers analyze live failure modes and update validation datasets to prevent similar failures from occurring. This process involves addressing shifts in data distribution and subpopulation performance, often through systematic categorization of data points based on error patterns.</li><li><strong>Standardized Validation Systems:</strong> Organizations aim to standardize validation processes to maintain consistency and reduce errors stemming from inconsistent evaluation criteria. This standardization helps ensure that models are evaluated comprehensively and consistently before deployment.</li><li><strong>Multi-stage Deployment and Evaluation:</strong> Many organizations employ multi-stage deployment processes, where models are progressively evaluated at each stage before full deployment. Stages such as testing, shadow deployment, and limited user rollout allow for early detection of problems and validation of model performance.</li><li><strong>Tying Evaluation Metrics to Product Metrics:</strong> It’s crucial to evaluate models based on metrics that are critical to the product’s success, such as click-through rate or user churn rate, rather than solely relying on ML-specific metrics. This alignment with product metrics ensures that models are evaluated based on their real-world impact and value to the organization.</li></ol><p>These efforts collectively aim to ensure that models are thoroughly evaluated, validated, and aligned with the organization’s objectives and requirements before deployment to production.</p><h3>Model Maintenance</h3><p>Sustaining high performance models requires deliberate software engineering and organizational practices. These are some common strategies that ML engineers use during monitoring and debugging phases to sustain model performance after deployment.</p><ol><li><strong>Creating New Versions:</strong> ML engineers frequently retrain models on live data to ensure that model performance does not suffer from data staleness. This involves automatically retraining models on a regular cadence, ranging from hourly to every few months, and frequently labeling live data to support retraining.</li><li><strong>Maintaining Old Versions as Fallback Models:</strong> To minimize downtime when a model is known to be broken, engineers maintain fallback models, either old versions or simpler versions, to revert to until the issue with the production model is resolved.</li><li><strong>Maintaining Layers of Heuristics:</strong> Engineers augment models with rule-based layers to stabilize live predictions. These heuristics help filter out inaccurate predictions based on domain knowledge, ensuring that only reliable predictions are served to users.</li><li><strong>Validating Data Going In and Out of Pipelines:</strong> Engineers continuously monitor features and predictions for production models, implementing various data validation checks to ensure data quality and consistency. These checks include enforcing constraints on feature values, monitoring data completeness, and verifying schema adherence.</li><li><strong>Keeping it Simple:</strong> ML engineers prefer simplicity, relying on simple models and algorithms whenever possible to streamline post-deployment maintenance. While some choose deep learning models for their ease of use and interpretability, others opt for simpler models to avoid overfitting and facilitate debugging.</li><li><strong>Organizational Support:</strong> Organizations implement various processes to support ML engineers in sustaining models, including on-call rotations for supervising production models, central queues for tracking and prioritizing production ML bugs, and Service Level Objectives (SLOs) to ensure minimum performance standards for pipelines.</li></ol><p>These strategies collectively help ML engineers maintain the performance and reliability of production ML pipelines, ensuring that models continue to deliver accurate predictions and meet the organization’s requirements and standards.</p><h3>Conclusions</h3><p>Across the four layers of the MLOps tech stack (run, pipeline, component, infrastructure) and across the four routine tasks for MLOps (data collection, experimentation, evaluation, maintenance), there are three common variables that dictate how successful the model deployment will be when developing and pushing ML models to production.</p><ol><li><strong>Velocity:</strong> This refers to the speed at which ML engineers can prototype, iterate, and develop new ideas into trained models. High experimentation velocity allows for rapid testing of hypotheses and quicker development cycles.</li><li><strong>Validation:</strong> It involves testing changes, identifying and pruning bad ideas, and proactively monitoring pipelines for bugs as early as possible. Early validation helps to catch errors before they become more expensive to handle, particularly when users encounter them.</li><li><strong>Versioning:</strong> This involves storing and managing multiple versions of production models and datasets. By maintaining version control, ML engineers can query, debug, and minimize production pipeline downtime. It also enables the quick switch to alternative versions in response to bugs or issues in the current production model.</li></ol><p>ML Tools at each layer of the MLOps tech stack should aim to enhance user experiences across the Three Vs: Velocity, Validation, and Versioning. For example, experiment tracking tools increase the velocity of iterating on feature or modeling ideas. In another example, feature stores (i.e., tables of derived features for ML models) help debug models because they offer reproducibility of historical versions of features used in training such models.</p><h3>Improvements</h3><p>Despite all of these best practices for MLOps, there are still many areas that need improvement. We need additional tooling to address the pain points within MLOps. These are some of the common pain points within MLOps.</p><p><strong>1. Mismatch Between Development and Production Environments:</strong> This includes issues such as data leakage, differing philosophies on the use of Jupyter notebooks, and non-standardized code quality practices.</p><ul><li><strong><em>Data leakage</em></strong> occurs when assumptions made during model training are not valid at deployment time, leading to incorrect predictions. It ranges from issues like using the same data for training and downstream tasks to ignoring feedback delays during training, impacting model performance in production.</li><li><strong><em>Opinions on Jupyter Notebooks.</em></strong> There are divergent opinions among practitioners regarding the use of Jupyter notebooks in ML workflows, with some favoring them for their quick prototyping abilities. However, concerns about reproducibility and code quality lead others to avoid using notebooks in production, emphasizing the importance of balancing velocity with environment consistency.</li><li><strong><em>Code quality standards</em></strong> and review practices in ML development vary widely, with some organizations lacking ML-specific coding guidelines. While code review can enhance deployment reliability, it may be perceived as a hindrance to agility in experimentation, leading to inconsistencies in maintaining code quality across different stages of ML projects.</li></ul><p><strong>2. Handling a Spectrum of Data Errors:</strong> ML engineers struggle with hard errors, soft errors, and drift errors in data, leading to false-positive alerts, alert fatigue, and challenges in creating meaningful data alerts.</p><ul><li><strong><em>Hard errors</em></strong> are obvious and result in clearly “bad predictions”, such as when mixing or swapping columns or when violating constraints (e.g., a negative age).</li><li><strong><em>Soft errors</em></strong>, such as a few null-valued features in a data point, are less pernicious and can still yield reasonable predictions, making them hard to catch and quantify.</li><li><strong><em>Drift errors</em></strong> occur when the live data is from a seemingly different distribution than the training set; these happen relatively slowly over time.</li></ul><p><em>False-positive alerts</em><strong><em> </em></strong>happen most often. They’re triggered even when ML performance is acceptable and leads to alert fatigue among engineers. This causes them to ignore or silence alerts and potentially miss actual performance issues. <em>Alert fatigue</em> management becomes crucial during on-call rotations, with initiatives aimed at reducing noise in alerts and improving the precision of alerting criteria. <em>Creating Meaningful Data Alerts is Challenging.</em><strong><em> </em></strong>One method might generate too many false positives, and another method might not catch enough of the errors. Finding a ‘Goldilocks’ method is difficult.</p><p><strong>3. ML Bugs Are Unique:</strong> ML debugging is different from debugging during standard software engineering, where one can write test cases to cover the space of potential bugs. ML Debugging is difficult to categorize bugs effectively because every bug feels unique.</p><p><strong>4. Multi-Staged Deployments Seemingly Take Forever:</strong> End-to-end experimentation in ML deployments takes too long, leading to uncertainties and the potential abandonment of promising ideas.</p><h3>Even More Improvements</h3><p>As you can imagine, there’s even more areas that need further improvement. These are some common anti-patterns within MLOps.</p><ol><li><strong>Industry-Classroom Mismatch:</strong> There is a gap between the skills learned in academic settings versus those required in industry. This leads to a lack of preparation for real-world ML challenges.</li><li><strong>Keeping GPUs Warm:</strong> There is a compulsive need to utilize all computational resources and run random experiments in parallel, instead of prioritizing the most impactful experiments and running those in sequence as insights continue to grow.</li><li><strong>Retrofitting an Explanation:</strong> There is pressure to provide explanations for successful ML experiments for successful collaboration and customer satisfaction. So explanations are created after the experiment is run, instead of having principled reasons as to why the experiment was run in the first place.</li><li><strong>Undocumented Tribal Knowledge:</strong> Pain points arise due to undocumented knowledge about ML experiments and pipelines. Team members are learning faster than documentation can be updated. Automated documentation assistance for ML pipelines would be helpful.</li></ol><p>Overall, addressing these anti-patterns requires a combination of educational initiatives, tool development, and process improvements to enhance the effectiveness and efficiency of MLOps practices.</p><p><em>If you enjoyed this article, please </em><a href="http://heartsymbol.love/symbols/heart">❤</a><em> to help others find it!</em></p><p><em>I would like to thank Shreya Shankar for conducting the interviews that this introductory article was based on. To see more details of these interviews, see the original paper: </em><a href="https://arxiv.org/abs/2209.09125"><em>[Operationalizing Machine Learning: An Interview Study]</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a07053fc2a80" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/overview-of-mlops-a07053fc2a80">Overview of MLOps</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fix ChatGPT Hallucinations]]></title>
            <link>https://medium.com/machinevision/fix-chatgpt-hallucinations-cbc76e5a62f2?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/cbc76e5a62f2</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Tue, 05 Mar 2024 18:06:33 GMT</pubDate>
            <atom:updated>2024-03-05T18:23:48.974Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IPipk0Xmo-oBsQwC5qswlw.png" /></figure><p>ChatGPT doesn’t always give you the response that you want or even information that is correct. Sometimes the response can be outright nonsense. Other times, ChatGPT can give you a response which appears to make sense but is factually inaccurate. These nonsensical, irrelevant, or incorrect responses are called hallucinations. Hallucinations might be welcome if you’re looking for creative output, but not if you’re looking for factual consistent information.</p><h3><strong>Causes of Hallucinations</strong></h3><p>ChatGPT is powered by LLMs (Large Language Models). ChatGPT takes a prompt as input and produces a response as output. However, the response often contains hallucinations. This may happen for a variety of reasons:</p><ul><li><strong>Lack of context:</strong> The prompt does not provide enough information.</li><li><strong>Unclear Prompt:</strong> The prompt does not make sense or is ambiguous.</li><li><strong>Training Data:</strong> The LLM does not have enough knowledge on the topic.</li></ul><h3><strong>Fixes for Hallucinations</strong></h3><p>Many experts have stated that hallucinations cannot be fixed or removed from LLMs. However, there are many strategies to mitigate hallucinations:</p><p><strong>1. Prompt Engineering:</strong> Make the prompt clearer and add more details.</p><p>Prompt Engineering is straightforward, but it requires knowledge of best practices for prompts and it takes more time to write the lengthier prompt.</p><p><strong>2. Human Feedback:</strong> Tell ChatGPT if the response is good or bad.</p><p>ChatGPT asks for user feedback via a thumbs up or thumbs down for each response. Also, OpenAI hires people full-time to let ChatGPT know whether a response is good or bad. To leverage the human feedback, ChatGPT uses an algorithm called RLHF (Reinforcement Learning with Human Feedback). <a href="https://medium.com/machinevision/overview-of-reinforcement-learning-58fbb905dbe0">[Learn about Reinforcement Learning]</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*zNzImw3uI8gG_ps2g_7YZQ.png" /><figcaption>Prompt Engineering and Human Feedback help mitigate hallucinations.</figcaption></figure><p><strong>3. External Resources:</strong> Ask to search the internet or upload PDFs.</p><p>You can ask ChatGPT to leverage external data or databases. A simple way to do this is to ask ChatGPT to search the internet for a particular question or topic inside your prompt. Also, you can upload PDF’s and other documents alongside your prompt.</p><p>Looking to the future, a popular technique that is being researched is RAG (Retrieval Augmented Generation). When an LLM is generating a response, RAG helps by simply searching the internet or external databases, without expecting the user to <strong><em>ask</em></strong> for external resources (e.g., internet) or upload external resources (e.g., PDFs). This makes the user experience simpler. However, this algorithm is not used in ChatGPT yet.</p><p><strong>4. Reduce Randomization:</strong> Decrease the temperature in OpenAPI.</p><p>OpenAI offers ChatGPT to non-technical users and it offers OpenAPI to technical users who can code. OpenAPI is an API which lets you use the ChatGPT technology via Python programming. OpenAPI lets you change one of the LLM hyper-parameters called <em>temperature</em>. Higher temperatures increase the selection probabilities for words that are less likely, making the model’s output more diverse and creative but potentially less grounded. Conversely, a lower temperature results in the model sticking more closely to words that are more likely, generally yielding more reliable outputs that are grounded in reality. However, this option is not available via ChatGPT yet.</p><p><strong>5. Fine-tune Model:</strong> Train the LLM on domain-specific data.</p><p>ChatGPT leverages a pre-trained Large Language Model (LLM), but this pre-trained model is trained on a broad set of domains and is not an expert in all domains. OpenAPI lets you take the pre-trained model and further train it on a smaller targeted dataset. This is called fine-tuning the model. If you have data related to a specific topic or domain, then you can fine tune the model using this data. This fine-tuning process helps the LLM have a better understanding of a particular topic or domain. However, this option is not available via ChatGPT yet.</p><p>There are many more ways to mitigate hallucinations, but there are no ways to completely remove hallucinations given the nature of LLMs.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cbc76e5a62f2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/fix-chatgpt-hallucinations-cbc76e5a62f2">Fix ChatGPT Hallucinations</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introduction to Bias in AI]]></title>
            <link>https://medium.com/machinevision/introduction-to-bias-in-ai-5058429ba0e?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/5058429ba0e</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[diversity]]></category>
            <category><![CDATA[diversity-in-tech]]></category>
            <category><![CDATA[bias]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Mon, 01 Mar 2021 20:38:52 GMT</pubDate>
            <atom:updated>2021-03-03T04:56:11.468Z</atom:updated>
            <content:encoded><![CDATA[<p>To understand AI Bias, we need to understand Dataset Bias. Collecting, labelling, and organizing data is a time consuming and expensive effort. Many popular datasets in the artificial intelligence community can take years to produce and publish. This effort requires a large amount of resources, and does not make dataset creation a small or efficient task. Since it’s impractical to create a dataset with all possible permutations and domains, all datasets have some form of bias in them. This limitation in data causes lower performance and decreased generalization across unrepresented domains.</p><p>The simple answer is to create more data, but that’s not easy. A better solution is to improve existing machine learning models with existing solutions (e.g., domain adaptation). But before we dive into solutions, let’s review the problem itself.</p><h3>What is the necessary vocabulary to understand AI Bias?</h3><p>I’ll go over a few terms that come up often, but there are many more.</p><h4><strong>Dataset Bias</strong></h4><p>Dataset bias corresponds to properties that are seen frequently in a dataset. For instance, in the COCO dataset, “person” is the most frequent object category across images. The bias in the COCO dataset is “person”. Biases can make it easier for a human to distinguish between datasets, but they can also often result in decreased model performance (due to overfitting), which can hinder learning reliable features.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*NKa66dsK73AF1OL-Ig1XiA.png" /><figcaption>COCO dataset has 66,000 labels for “person”. It only has 2,000 labels for “elephant”. The data bias is “person”.</figcaption></figure><h4>Domain Shift</h4><p>Domain shift refers to the situation where training data and test data have different domains. For instance, domain shift happens when daytime images are used for training data and nighttime images are used for test data. The bias in the training data is “daytime”. This domain shift leads to lower performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mZzgSyzQ486-TXP7rB34DQ.png" /><figcaption>Train the model on daytime data (left). Test the model on nighttime data (right).</figcaption></figure><h3>What is an example of AI Bias?</h3><p>Suppose you have a neural network that takes images and predicts object bounding boxes and labels. You could deploy it on a self-driving car to identify pedestrians. Let’s assume that the model was trained on a dataset from sunny California.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MqsehmGrS2m-Na63olbDUA.png" /><figcaption>The detection system can detect pedestrians from California (left), but not from Boston (right).</figcaption></figure><p>Let’s say that you deploy the trained network in Boston during the Winter. Since the network hasn’t seen this type of data during training time, it’s very likely that it’s going to miss quite a few objects and quite a few pedestrians in this example. This is the problem of domain shift which is caused by dataset bias. Where the source data is from the source domain, and the target data from deployment is called the target domain. And it has distribution shift over the input data. The distribution has shifted from training to test time.</p><h3><strong>What are other examples of AI Bias?</strong></h3><p>Domain shift happens in applications when the models are deployed in the real world.</p><h4><strong>Datasets</strong></h4><p>In general, domain shift can happen when the model is trained on one dataset and the trained model is applied to a different dataset. Specifically, a dataset of objects can come from product images with white backgrounds and canonical poses, and the model can be deployed on a dataset collected from a mobile robot where the backgrounds are cluttered and the viewpoints and lighting conditions are very different.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HvWj0_1UAM6rHx9NVaJsMw.png" /><figcaption>Product images with white backgrounds (left). Robot images with cluttered backgrounds (right).</figcaption></figure><h4><strong>Skin Tones</strong></h4><p>Here is a very important problem where we are training a face detection system on data with a bias for light skinned faces. This results in poor predictive performance on images of darker skinned faces. We’d like to improve the performance for people with darker skin tones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zFWYxksDWPOmHt9AHnKMHA.png" /><figcaption>Face detection works on lighter-skinned people, but less on darker-skinned people. [Credit: Joy Buolamwini]</figcaption></figure><h4><strong>Modalities</strong></h4><p>Another example of domain shift is when its training data is biased to a particular modality. For example, the training data is RGB images, and the test data is depth images. Using domain adaptation, we can improve the detection performance on the Depth images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9s8Pj_UOos1kMSXeOkfLFg.png" /><figcaption>This uses domain adaptation to improve performance from RGB images (left) to Depth images (right).</figcaption></figure><h4><strong>Sim2Real</strong></h4><p>Another very common example is training robots in simulation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nqsXsbwaDw8a1KDEmASHrA.png" /><figcaption>Training an object manipulator in simulation and deploying it in the real setting.</figcaption></figure><p>In this example, a robot arm is picking up an object. Ideally, we want to train these policies in simulation, because it’s a cheaper source of data that doesn’t damage the robot. However, at test time, we get real images. Therefore, we’d like to be able to adapt to this bias from simulation to reality. <em>(For more information on Sim2Real, check out my Overview on Sim2Real. Coming Soon.)</em></p><h3><strong>Why is AI Bias bad?</strong></h3><p>Dataset Bias in the training data causes poor performance and poor generalization to future test data. The distinction in the kind of images during train and test can render a model useless during evaluation. It causes a significant drop in performance and makes our models inaccurate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PI6_9Pu2gOGhTbc05Nu0cA.png" /><figcaption>If you train on MNIST data, but test on a different domain (e.g., USPS, SVHN), then you have a significant drop in performance.</figcaption></figure><p>For example, if we train on MNIST and test it on images from MNIST, then it should have 99% accuracy. However, if we train on the MNIST domain and test it on the Street View House Numbers (SVHN). We’ll see that the performance will be around 67.1%. This is much lower than it should be, so this is a serious issue. In fact, even between two similar looking domains of MNIST and USPS, there is still a very significant drop in performance.</p><h3><strong>Existing Solutions</strong></h3><p>Solutions for AI Bias exist today. These solutions improve model performance. We get a high level sense of some of them here.</p><h4>Domain Adaptation</h4><p>Domain adaptation is one solution for domain shift. It’s a method for adapting a model to a novel dataset with no or few labels from the novel dataset. We assume access to a small set of target domain images during training. During training, you can get a sense of what the target distribution looks like.</p><h4>Domain Generalization</h4><p>Domain Generalization is another solution for domain shift. It doesn’t assume that target data exists in the training data. Instead, it assumes no target data at all during training. It’s a strictly zero-shot scenario and relies heavily on model generalization.</p><h4>Latent Domain Discovery</h4><p>Latent Domain Discovery helps with Domain Generalization. Some aspects of both dataset bias and domain shift can be easy for humans to infer by observation. But there can be some aspects which are inherently latent and may not be immediately obvious. Latent variables define a dataset but are not strictly observed. These latent domains exist but are not labeled in datasets. For instance, images found on the web can be thought of as a collection of many hidden domains. Discovering latent domains can significantly improve generalization and performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l7eKGlva9J5lQWjhHMAA7A.png" /><figcaption>Image results for “person” consist of latent domains (e.g., “groups”, “silhouettes”, “line drawings”, “close-ups”).</figcaption></figure><p>Other solutions not mentioned here include Transfer Learning and Representation Learning. We will cover more solutions in further depth for future articles. In the meantime, feel free to reach out if you have any specific topics you’d like me to cover. Also, if you’re new to AI, this brief <a href="https://medium.com/machinevision/overview-of-artificial-intelligence-buzz-adb7a5487ac8">Intro to AI</a> should help.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5058429ba0e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/introduction-to-bias-in-ai-5058429ba0e">Introduction to Bias in AI</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Overview of Embodied Artificial Intelligence]]></title>
            <link>https://medium.com/machinevision/overview-of-embodied-artificial-intelligence-b7f19d18022?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/b7f19d18022</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[facebook]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[nlp]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Tue, 09 Feb 2021 10:37:18 GMT</pubDate>
            <atom:updated>2021-05-17T21:19:03.145Z</atom:updated>
            <content:encoded><![CDATA[<p>Recent research trends in Artificial Intelligence, Machine Learning, and Computer Vision have led to a growing research space called Embodied AI. Facebook AI Research (FAIR) and Intel Labs has been spearheading new projects in the space of Embodied AI. “Embodied” is defined as “giving a tangible or visible form to an idea.” Simply put, “Embodied AI” means “AI for virtual robots.” More specifically, Embodied AI is the field for solving AI problems for virtual robots that can move, see, speak, and interact in the virtual world and with other virtual robots — these simulated robot solutions are then transferred to real world robots.</p><h3>History</h3><p>Linda Smith proposed the “embodiment hypothesis” in 2005 as the idea that intelligence emerges in the interaction of an agent with an environment and as a result of sensorimotor activity. They argue that starting as a baby grounded in a physical, social, and linguistic world is crucial to the development of the flexible and inventive intelligence that characterizes humankind. Furthermore, the Embodiment Thesis states that many features of cognition are embodied in that they are deeply dependent upon characteristics of the physical body of an agent, such that the agent’s beyond-the-brain body plays a significant causal role, or a physically constitutive role, in that agent’s cognitive processing. While the initial hypothesis comes from Psychology and Cognitive Science, the recent research developments of Embodied AI has come largely from Computer Vision researchers.</p><p>AI subfields have been largely separated since the 1960s, subject to various limitations. However, Embodied AI brings together interdisciplinary fields, such as Natural Language Processing (NLP), Computer Vision, Reinforcement Learning, Navigation, Physics-Based Simulations, and Robotics. While Embodied AI requires multiples AI subfields to succeed, the growth of Embodied AI as a research area has largely been driven by Computer Vision researchers.</p><p>Computer Vision researchers define Embodied AI as artificial agents operating in 3D environments that base their decisions off of egocentric perceptual inputs that in turn change with agent actions. Embodied AI enables training of embodied AI agents (virtual robots and egocentric assistants) in a realistic 3D simulator, before transferring the learned skills to reality. This empowers a paradigm shift from “Internet AI” based on static datasets (e.g., ImageNet, COCO, VQA) to “Embodied AI” where agents act within realistic simulated environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IzP6ru0VN7z5AdQM.png" /><figcaption>Internet AI. These images from the COCO dataset do not provide a 3D realistic environment.</figcaption></figure><h3>Motivation</h3><p>Many of the AI advances in the last decade have been because of Machine Learning and Deep Learning (e.g., Semantic Segmentation, Object Detection, Image Captioning). Machine Learning and Deep Learning have been successful because of the increasing amounts of data (e.g., Youtube, Flickr, Facebook) and the increasing amounts of computing power (e.g., CPUs, GPUs, TPUs). However, this type of “Internet data” (images, video, and text curated from the internet) does not come from a real-world first-person human perspective. Data is shuffled, randomized, coming from satellites, coming from selfie photos, coming from Twitter feeds, and none of this is how a human perceives the world. However, Machine Learning methods are trying to feed this data to NLP, CV, and Navigation problems. While there has been much progress in these fields due to “Internet data” and “Internet AI”, it is not the most suitable data or the most suitable methods. The methods for Machine Learning do not map to the ways that humans learn. Humans learn by seeing, moving, interacting, and speaking with others. Humans learn from sequential experiences, not from shuffled and randomized experiences. The thesis of Embodied AI is to have embodied agents (or virtual robots) learn in the same way that humans learn. Which is why insight from Cognitive Science and Psychology experts is essential. This means that virtual robots should learn by seeing, moving, speaking, and interacting with the world — just like humans.</p><p>While “Embodied AI” has a different methodology than “Internet AI”, Embodied AI can benefit from many of the successes from Internet AI. Computer Vision and Natural Language Processing actually work pretty well for some things now (if there is plenty of labelled data). These advances in CV and NLP greatly increase the potential for success in Embodied AI.</p><p>Additionally, there now exist plenty of realistic 3D scenes which can serve as simulated environments for Embodied AI training. These environments include SUNCG, Matterport3D, iGibson, Replica, Habitat, and DART. These scenes are much more realistic than the environments that have been used in previous research simulators. The widespread and public availability of these datasets greatly increase the potential for success in Embodied AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a_1a2Bsos1k1h8sE.png" /><figcaption>This is a 3D Map view of the realistic environment. The embodied agent only sees a first person view.</figcaption></figure><p>Embodied AI lends itself to applications, such as personal robotics, virtual assistants, and even autonomous vehicles. The combination of NLP, Computer Vision, and Robotics actually make the problem tasks of Embodied AI easier.</p><h3>Agent and Environment</h3><p>An <strong>agent</strong> is merely an abstraction that can take actions in an environment. We refer to this agent as virtual robot, simulated agent, virtual agent, or egocentric agent (due to its first-person views/sensors/interactions). You can think of an agent as a player of a game. On the other hand, an <strong>environment</strong> is just an abstraction that represents a 3D map with multiple locations, rooms, and objects in the world. It is a simulated environment that represents the physical world. You can accomplish many goals in an environment, such as interaction, navigation, and language understanding. For more information on Agents and Environments, feel free to read our <a href="https://medium.com/machinevision/overview-of-reinforcement-learning-58fbb905dbe0">Overview of Reinforcement Learning</a>.</p><h3>Egocentric Perception</h3><p>Embodied AI is egocentric as opposed to allocentric. Egocentric perception can be though of as a first-person view. Egocentric perception encodes objects with respect to the agent. Allocentric perception encodes objects with respect to another object (e.g., the front door, the center of the room, etc.). An allocentric perception might be useful if the AI knows the whole map of the environment; however, an embodied agent only knows what they have seen from egocentric perception. An embodied agent does not have access to a map, unless they create it as they navigate the different rooms and locations in the environment. Recall that the Embodiment Thesis focuses on the self, so it would only be appropriate for Embodied AI to also focus on the self.</p><h3>Internet AI</h3><p>Embodied AI has additional challenges compared to Internet AI. Internet AI learns from static images from Internet datasets (e.g., ImageNet). These static images are high quality and nicely framed; on the other hand, Embodied AI uses egocentric perception, which produces images or videos that might be shaky and not well composed. The inherent properties of Egocentric Perception create additional changes for Embodied AI as compared to Internet AI. Additionally, the focus of Internet AI is pattern recognition in images, videos, and text on datasets typically curated from the internet; on the other hand, the focus of Embodied AI is to enable action by an embodied agent (e.g. robot) in an environment. Ultimately, the goal is to take all of the advances that Machine Learning and Computer Vision have made in Internet AI and apply it to Embodied AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KSoHnePFYMxBo1nn2vEenQ.png" /><figcaption>The goal in this example is to find the find the car and perceive its color. It needs to understand the question, find the car, and then answer.</figcaption></figure><h3>Active Perception</h3><p>The agent may be spawned anywhere in the environment and may not immediately ‘see’ the pixels containing the answer to its visual goal (i.e. the car/goal may not be visible). Thus, the agent must move to succeed — controlling the pixels that it will perceive. The agent must learn to map its visual input to the correct action based on its perception of the world, the underlying physical constraints, and its understanding of the question. The observations that the agent collects are a consequence of the actions that the agent takes in the environment. The agent is controlling the data distribution that is coming in. The agent controls the pixels it gets to see. This is unlike static datasets, which have been curated online and there’s less control over viewpoint variations of objects, etc. One of the challenges of active perception is to be generally robust to visual variation.</p><h3>Sparse Rewards</h3><p>Unlike object detection or image recognition (supervised learning), these agents do collect immediate rewards for each action. Agents in an environment often experience sparse rewards (reinforcement learning). The aim of a reinforcement learning (RL) algorithm is to allow an agent to maximize the rewards from the environment. In some environments, the rewards are supplied to the agent continuously. In others, a positive reward is only provided when the agent completes the goal (e.g., “walk to car”), but it leads to sparse rewards. Sparse rewards can make learning the intended behavior more challenging. It can also make exploration more challenging. For more information on Reinforcement Learning, feel free to read our <a href="https://medium.com/machinevision/overview-of-reinforcement-learning-58fbb905dbe0">Overview of Reinforcement Learning</a>.</p><h3>Tasks</h3><p>There are several tasks that can be accomplished in the field of Embodied AI. Here are some of the existing tasks.</p><p>1) <strong>Visual Odometry.</strong> Odometry is using any sensor to determine how much distance has been traversed, so visual odometry is just clarification that the particular sensor to be used for odometry is visual (e.g., camera). Traversed distance in odometry is relative to the starting position. So visual odometry assumes the initial position is known. Visual odometry (VO), as one of the most essential techniques for pose estimation, has attracted significant interest in both the computer vision and robotics communities over the past few decades. It has been widely applied to various robots as a complement to GPS, Inertial Navigation System (INS), wheel odometry, etc. In the last thirty years, enormous work has been done to develop accurate and robust VO systems.</p><p>2) <strong>Global Localization.</strong> Localization is the problem of estimating the position of an autonomous agent given a map of the environment and agent observations. The ability to localize under uncertainty is required by autonomous agents to perform various downstream tasks such as planning, exploration and navigation. Localization is considered as one of the most fundamental problems in robotics. Localization is useful in many real-world applications such as autonomous vehicles, factory robots and delivery drones. The global localization problem assumes the initial position is unknown (as compared to VO which assumes that the initial position is known). Despite the long history of research, global localization is still an open problem.</p><p>3) <strong>Visual Navigation.</strong> Navigation in three-dimensional environments is an essential capability of robots that function in the physical world (or virtual robots in a simulated environment). Animals, including humans, can traverse cluttered dynamic environments with grace and skill in pursuit of many goals. Animals can navigate efficiently and deliberately in previously unseen environments, building up internal representations of these environments in the process. Such internal representations are of central importance to Artificial Intelligence. For more information on Visual Navigation, feel free to read our Overview of Embodied Navigation. (Coming Soon)</p><p>4) <strong>Grounded Language Learning.</strong> We are increasingly surrounded by artificially intelligent technology that takes decisions and executes actions on our behalf. This creates a pressing need for general means to communicate with, instruct and guide artificial agents, with human language the most compelling means for such communication. To achieve this in a scalable fashion, agents must be able to relate language to the world and to actions; that is, their understanding of language must be grounded and embodied. However, learning grounded language is a notoriously challenging problem in artificial intelligence research.</p><p>5) <strong>Instruction Guided Visual Navigation.</strong> The idea that we might be able to give general, verbal instructions to a robot and have at least a reasonable probability that it will carry out the required task is one of the long-held goals of robotics, and artificial intelligence. Despite significant progress, there are a number of major technical challenges that need to be overcome before robots will be able to perform general tasks in the real world. One of the primary requirements will be new techniques for linking natural language to vision and action in unstructured, previously unseen environments. It is the navigation version of this challenge that is referred to as Vision-and-Language Navigation (VLN).</p><p>6) <strong>Embodied Question Answering.</strong> EmbodiedQA is where an agent is spawned at a random location in a 3D environment and asked a question (e.g., ‘What color is the car?’). In order to answer (e.g., ‘Orange!’), the agent must first intelligently navigate to explore the environment, gather information through first-person (egocentric) vision, and then answer the question. This challenging task requires a range of AI skills — active perception (e.g., agent must move to perceive the car — controlling the pixels that it will perceive), language understanding (e.g., what is the question asking?), goal-driven navigation, commonsense reasoning (e.g., where are cars generally located in the house?), and grounding of language into actions (e.g., associate entities in text with corresponding image pixels or sequence of actions).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2iOk-91VkYeD8LKGKyq3lA.png" /><figcaption>Egocentric view of embodied agent navigating through a realistic environment.</figcaption></figure><h3>Datasets and Simulators</h3><p>Datasets have been a key driver of progress in Internet AI. With Embodied AI, simulators will assume the role played previously by datasets. <strong>Datasets</strong> consist of 3D scans of an environment. These datasets represent 3D scenes of a house, or a lab, a room, or the outside world. These 3D scans however, do not let an agent “walk” through it or interact with it. <strong>Simulators</strong> allow the embodied agent to physically interact with the environment and walk through it. The datasets are imported into a simulator for the embodied agents to live in and interact with. With the simulator, the agent can see, move, and interact with its environment. The agent can even speak to other agents or humans with the power of the simulator.</p><p>Environments are realized through simulators. However, there are different types of environment representations.</p><p>1) <strong>Unstructured.</strong> An environment can keep everything and maybe compress it. An example of this is Habitat. It will be able to handle long-horizon tasks. This is the predominant model of current environments and usually has metric representations of the environments.</p><p>2) <strong>Topological.</strong> An environment can be represented with a graph of key nodes. Topological environments consist of a graph with nodes corresponding to locations in the environment and a system capable of retrieving nodes from the graph based on observations. The graph stores no metric information, only connectivity of locations corresponding to the nodes. Topological environments allow for navigation strategies, such as “landmark navigation” as opposed to navigating with metric representations.</p><p>3) <strong>Spatial Memory / Cognitive Map.</strong> A mapper can build a spatial map of its environment based on the agent’s egocentric views. The spatial memory captures the layout of the environment. The cognitive map is fused from first-person views as observed by the agent over time to produce a metric/semantic egocentric belief about the world in a top-down view. At each time step, the agent updates the belief of its environment. This allows the agent to progressively improve its model of the environment as it moves around.</p><h3>Sim2Real</h3><p>A straightforward way to train embodied agents is to place them directly in the physical world. This is valuable, but training robots in the real world is slow, dangerous (robot can fall over and break), resource intensive (robot and environment demand resources and time), and difficult to reproduce (especially rare edge cases). An alternative is to train embodied agents in realistic simulators and then transferring the learned skills to reality. Simulators can help overcome some of the challenges of the physical world. Simulators can run orders of magnitude faster than real-time and can be parallelized over a cluster; training in simulation is safe, cheap. Once an approach has been developed and tested in simulation, it can be transferred to physical platforms that operate in the real world.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/738/1*_Vm385LPqpXVWjjHn1yzZQ.png" /><figcaption>Embodied AI robot can work with fire fighters.</figcaption></figure><h3>Applications</h3><p>We want to apply Embodied AI to the real world. We want a physical agent that is capable of taking actions in the real world and can talk to humans with natural language. For instance, a search and rescue scenario. A fire fighter asks: “Is there smoke in any room?” First, the robot has to understand what the question is asking. Then, the language learning and comprehension needs to be grounded in terms of the environment, so that it can understand what it means by “room”. If it the agent is confused or needs assistance, it needs to be able to ask the firefighter what to do. For example, the agent might ask “What subset of rooms are you interested in?” And the firefighter could respond with “The rooms in the burning house.” Or perhaps this particular agent could have learned enough “common sense” to know that the rooms that need to be searched for smoke are the rooms in the burning building. Once the agent understand its objective, it needs to be able to navigate to each room. After it arrives at each room, it needs to be able to comprehend whether there is smoke in the room or not base on its egocentric perception. Finally, it can navigate back to the firefighter and respond: “Yes, in one room.”</p><p>This firefighting scenario shows multiple examples of Embodied AI tasks.</p><h3>Conclusion</h3><p>Recent advancements in Internet AI have fueled the initial progress of Embodied AI. Once Embodied AI transcends the developments of Internet AI, it will be a major leap forward in enabling robots to effectively learn how to interact with the real world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b7f19d18022" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/overview-of-embodied-artificial-intelligence-b7f19d18022">Overview of Embodied Artificial Intelligence</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Overview of Reinforcement Learning]]></title>
            <link>https://medium.com/machinevision/overview-of-reinforcement-learning-58fbb905dbe0?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/58fbb905dbe0</guid>
            <category><![CDATA[luis-bermudez]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Wed, 19 Dec 2018 01:01:22 GMT</pubDate>
            <atom:updated>2018-12-21T16:38:26.180Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KCLG3VkqwWXdgV2CXwp3Kg.jpeg" /></figure><p>You might have heard of Google DeepMind beating several Atari games, beating a professional Go player, or even teaching a humanoid simulation to walk. This has been achieved with the help of Reinforcement Learning.</p><p>Reinforcement Learning is a subcategory of Machine Learning where an agent learns to behave in an environment. Other popular subcategories of Machine Learning are Supervised Learning and Unsupervised Learning. Reinforcement Learning is different than supervised learning and unsupervised learning, in that the sequence of actions you take to achieve your goal is important to the problem at hand.</p><p>Supervised learning is done using ground truth. Thus, the goal of supervised learning is to learn a function or mapping between sample data and the ground truth outcomes. Popular uses of Supervised Learning include Regression and Classification.</p><p>Unsupervised learning, on the other hand, does not have labeled ground truth. Thus, the goal of unsupervised learning is to infer the inherent structure of the sample data without using explicitly provided ground truth labels. A popular use of Unsupervised Learning is Clustering.</p><p>Reinforcement Learning does not require preexisting ground truth mappings. It learns a behavior (a set of actions) by experiencing an environment through trial and error. This can be a real-world environment or a simulated world. The agent tries different actions in the environment until he starts learning the optimal actions to take. As such, the goal of reinforcement learning is to learn an optimal behavior to achieve a goal.</p><p>Examples of successful use cases for Reinforcement Learning include: Robotic Arm Manipulation, Google DeepMind beating a professional Alpha Go Player, Google DeepMind beating several Atari games, Google DeepMind training a humanoid to walk on its own, Unity teaching a puppy how to walk and play fetch, and the OpenAI team beating a professional DOTA player.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/201/1*iy1yxG_nsjNwX7NC9zoTFQ.png" /><figcaption>This humanoid agent is teaching itself to walk. [<a href="https://www.youtube.com/watch?v=imOt8ST4Ejc">Source</a>]</figcaption></figure><h3>Agent and Environment</h3><p>The first concept to try to understand when you are doing Reinforcement Learning is that a lot of it takes place as a conversation between an agent and an environment. You can imagine that you are an agent in a video game environment. You are the agent. The video game environment is the environment. The conversation is what is going back and forth between the agent and the environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tTsUFmyRR7tRoRu3KYm7IQ.png" /></figure><p>An <strong>agent</strong> is merely an abstraction that can take actions, exist in a current state, transition between states, and receive rewards. You can think of an agent as a player of a game. On the other hand, an <strong>environment</strong> is just an abstraction that is represented via a set of states and rewards. The environment also has a goal that it wants accomplished. You can think of an environment as a board game or a sports game — really any activity with a goal.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*IhjRVJAfQ5outRnoOFM_ag.png" /></figure><p>The environment is going to reveal itself in the form of states. The agent can then influence the environment by taking actions. After taking an action, you (the agent) can get some kind of reward before moving onto the next state. This reward is based on the most recent state-action combination. This interaction between the agent and environment allows the agent to learn more and more about the environment by experiencing the states and rewards as consequences of its actions. Really, all of the computation is happening in the head of the agent. And the information about the environment is really only available through the course of this interaction between the agent and the environment. The environment is not fully known to the agent.</p><h3>Unknown Environment</h3><p>Remember to keep in mind that the agent does not know all of the environment. You (the agent) only experience the environment through the set of interactions with the environment. If you interact with the environment often enough, then you can build some kind of model of the environment in your head. But note that the model of the environment that you build, and the actual environment are not necessarily the same thing. In other words, the agent does not know the environment, the agent is just experiencing the environment by interacting with it. All of the computation is happening in the brain of the agent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*BWrC7ibIKMpOMNd5q9VKuQ.png" /></figure><p>If all of the environment was already known, then the agent would not need to do any learning and instead the problem would just be to plan what to do based on the existing known environment. This brings us to the difference between <strong>planning</strong> and <strong>learning</strong>. Learning happens when you do not know the model/environment. Planning happens when you do know the model/environment. Planning might be able to happen in your own home or in your local streets, since you know them. Learning might happen in a garden maze or a new theme park that you have not been to before. You can only do planning when you know the environment and have a model of it.</p><h3>Learning Through Experience</h3><p>When we put an agent in an environment, this environment will have states that the agent can exist within and the environment will also have actions that the agent can take. This environment can be a game or the real world. Taking an action can transition you to another state. Taking an action can also result in a reward. However, as opposed to a regular game, you do not know the rules. You do not know how things work. You do not even know what you’re supposed to do. But what you can do is to start playing and use that experience to start discovering how things work.</p><p>If you choose an action, then the environment will tell you if you gain a reward and also if you move to another state. Some actions might give you a reward and some might not (depending on your state). Some actions might change your state and some might not (depending on your state). For example, “forward action” might move you forward; however, “forward action” might not move you forward if you are in front of a wall. Another example is that “pickup action” might pickup an object underneath you; however, if there is no object underneath you, then the “pickup action” might not do anything.</p><p>These rules of the environment are unknown to you when you begin the game. So, as an agent, you just have to try different actions and see what happens. Also, keep in mind that while “forward action” or “pickup action” might mean something to you (as a human with life experience), the agent has no idea what “forward action” or “pickup action” does and it might as well be called “Action 1” or “Action 2”. You can think of an agent as starting out as a baby with no life experience. Another example is that an agent does not initially know what a “wall” is, so it has to learn what happens when you bump into a wall.</p><p>As you can see, an agent has an even harder time learning what a state-action pair might do than a human. Eventually, the agent might understand how the environment works. This understanding might be correct, it might be incorrect, or more likely somewhere in between. All you know is that you want to maximize the reward that you’ve been collecting along the way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/732/0*RUXpkumDw9kLyzUc.jpg" /></figure><h3>Maximizing Rewards</h3><p>For each action you take in a state, you receive a reward. You can think of rewards as positive rewards, negative rewards or neutral rewards. The higher the reward the better. If you win the game, you might receive an extremely high positive reward. If you fall into a pit, you might lose the game and receive an extremely negative reward. Along the way, you can receive smaller positive and negative rewards. For example, if you get closer to the goal you might get a small positive reward every time you get closer. On the other hand, if you lose health in your video game or you lose a pawn in your chess match, then you might receive a small negative reward. There might be actions that do not affect your goal at all, so you might receive a neutral reward of zero.</p><p>During your experience in the environment, you gain a sum of rewards which is called the <strong>return</strong>. The goal of interacting with the environment is to maximize the total return. So you might have one run of a video game where you lose the game if walk into a pit, this will likely lead a low total return. Another run of the same video game might try another action (such as a jump), which might lead to not falling into the pit and receiving a higher total return. It is by trial and error that the agent learns what are good actions and what are bad actions in certain states.</p><h3>Stochastic Environment</h3><p>So far we have been considering a deterministic environment, where the same thing happens for any state-action pair. But what if different things happen sometimes for a given state-action pair? This is what we call a Stochastic Environment. Most real-life environments are stochastic. If you cross an intersection and there is incoming traffic, then you will not move ahead, and instead your car will collide with another car. So in a given state (e.g., an intersection), the same thing doesn’t always happen with one action. To account for this, we create a probabilistic model, which might say that there is oncoming traffic 20% of the time and there is no oncoming traffic 80% of the time. This part of the stochastic environment, is what you call a probabilistic transition function.</p><p>A good way to model the environment is through a Markov Decision Process (MDP). We go further into MDP’s in the next article.</p><h3>Stationary Policy</h3><p>With reinforcement learning algorithms, we are trying to learn behavior. We are trying to learn a way to interact with the environment that collects high reward. The behavior that you might be represented by a <strong>plan</strong> or a <strong>conditional plan</strong>, but reinforcement learning algorithms prefer to represent behavior as a <strong>universal plan </strong>(aka, stationary policy).</p><p>A plan is a fixed sequence of actions. Assuming that there is no traffic and all of the stop lights are green, then you have to do is follow a plan to get from work to home (e.g., “left”, “left”, “up”, “left”, “left”, “down”). However, these assumptions will not always hold in real life situations. There might be a traffic at one intersection, so it is better to have a conditional plan. A conditional plan allows you to pick different actions at a particular state. Most of the plan is fixed, but there might be a one or two conditional states that provide different actions based on a condition.</p><p>A universal plan, on the other hand, is similar to a conditional plan except that it has a condition for every single state. This is different than a plan or conditional plan, which only needs to worry about the states that go from point A to point B. A universal plan will have a conditional set of actions at every single state in your environment. If your environment is your city, and your goal is to go from your work to your home, then you will have a conditional set of actions at every single intersection in your city. You will know what to do at every single state in your environment, instead of just a list of states which achieve your goal.</p><p>Reinforcement Learning focuses on stationary policies because they allow us to express an optimal behavior. This is very powerful, but is also very large which makes it difficult for large state spaces. Just imagine a continuous state space, which is infinitely large. Deep Reinforcement Learning is one solution to this largeness problem. We will go further into Deep Reinforcement Learning in another article.</p><h3>Exploitation vs Exploration</h3><p>A general theme that comes up in Reinforcement Learning Algorithms is exploitation vs exploration. If your algorithm is always exploiting its knowledge of the model, then it might only find a local optimum, since it is disregarding other optimal points which could yield higher returns. This is where exploration comes in. If you want to discover if there are other local optimums which are better than what you have already found, then you need to go exploring. This doesn’t mean that you should explore all the time or that you should exploit all the time, but you should find a balance between the two.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/0*cb1UnGDORLkBv0h9" /></figure><h3>Conclusion</h3><p>Reinforcement Learning isn’t the buzz word that you have been hearing. There are plenty of other popular buzzwords when it comes to Machine Learning. But this is the technique behind some of the most interesting examples of Machine Learning that you’ve heard about. If you’d like an overview of the trendier Machine Learning concepts, then read my other article on <a href="https://medium.com/machinevision/overview-of-artificial-intelligence-buzz-adb7a5487ac8">Machine Learning Buzz</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=58fbb905dbe0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/overview-of-reinforcement-learning-58fbb905dbe0">Overview of Reinforcement Learning</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Overview of Neural Networks]]></title>
            <link>https://medium.com/machinevision/overview-of-neural-networks-b86ce02ea3d1?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/b86ce02ea3d1</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[luis-bermudez]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Tue, 14 Nov 2017 17:04:52 GMT</pubDate>
            <atom:updated>2017-11-15T05:39:46.457Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7gSjrTgqw3NDTrnrTHZYPA.png" /><figcaption>Neural Networks are modeled after the human brain.</figcaption></figure><p>If you’ve heard about Artificial Intelligence, Machine Learning, or Deep Learning recently, then you might have heard of a Neural Network.</p><p>Neural Networks are a key piece of some of the most successful machine learning algorithms. The development of neural networks have been key to teaching computers to think and understand the world in the way that humans do. Essentially, a neural network emulates the human brain. Brains cells, or neurons, are connected via synapses. This is abstracted as a graph of nodes (neurons) connected by weighted edges (synapses).</p><p>So let’s dive in. What is a neural network? The human brain consists of 100 billion cells called neurons, connected together by synapses. If sufficient synaptic inputs fire to a neuron, that neuron will also fire. We call this process “thinking”. We can model this process by creating a neural network on a computer. A neural network has input and output neurons, which are connected by weighted synapses. The weights affect how much of the forward propagation goes through the neural network. The weights can then be changed during the back propagation — this is the part where the neural network is now learning. This process of forward propagation and backward propagation is conducted iteratively on every piece of data in a training data set. The greater the size of the data set and the greater the variety of data set that there is, the more that the neural network will learn, and the better that the neural network will get at predicting outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aoJHHuA3Qd5SycyC9_hXww.png" /><figcaption>This neural network has one layer, three inputs, and one output. Any neural network can have any number of layers, inputs, or outputs.</figcaption></figure><p>Simply put, a neural network is a connected graph with input neurons, output neurons, and weighted edges. Let’s go into detail about some of these components:</p><p>1) <strong>Neurons</strong>. A neural network is a graph of neurons. A neuron has inputs and outputs. Similarly, a neural network has inputs and outputs. The inputs and outputs of a neural network are represented by input neurons and output neurons. Input neurons have no predecessor neurons, but do have an output. Similarly, an output neuron has no successor neuron, but does have inputs.</p><p>2) <strong>Connections and Weights</strong>. A neural network consists of connections, each connection transferring the output of a neuron to the input of another neuron. Each connection is assigned a weight.</p><p>3) <strong>Propagation Function</strong>. The propagation function computes the input of a neuron from the outputs of predecessor neurons. The propagation function is leveraged during the forward propagation stage of training.</p><p>4) <strong>Learning Rule</strong>. The learning rule is a function that modifies the weights of the connections. This serves to produce a favored output for a given input for the neural network. The learning rule is leveraged during the backward propagation stage of training.</p><h3>Deep Neural Networks</h3><p>So now that we know what a Neural Network is. What is a Deep Neural Network? A Deep Neural Network simply has more layers than smaller Neural Networks. A smaller Neural Network might have 1–3 layers of neurons. However, a Deep Neural Network (DNN) has more than a few layers of neurons. A DNN might have 20 or 1,000 layers of neurons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0UovTUIDiixK3z5ewLMlfg.png" /><figcaption>This neural network has two layers, three inputs, and one output. Any neural network can have any number of layers, inputs, or outputs. The layers between the input neurons and the final layer of output neurons are hidden layers of a deep neural network.</figcaption></figure><h3>Conclusion</h3><p>That’s basically it for a Neural Network. A neural network is just a core architecture. There are different types of neural networks. For example, Convolutional Neural Networks have been very effective for Computer Vision applications. Recurrent Neural Networks are also very popular. If anything here was unclear or you need a broader overview, then feel free to read this <a href="https://medium.com/machinevision/overview-of-artificial-intelligence-buzz-adb7a5487ac8">Overview of Artificial Intelligence, Machine Learning, and Deep Learning</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b86ce02ea3d1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/overview-of-neural-networks-b86ce02ea3d1">Overview of Neural Networks</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Overview of Artificial Intelligence Buzz]]></title>
            <link>https://medium.com/machinevision/overview-of-artificial-intelligence-buzz-adb7a5487ac8?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/adb7a5487ac8</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[luis-bermudez]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Tue, 14 Nov 2017 16:31:26 GMT</pubDate>
            <atom:updated>2017-11-15T03:29:45.529Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JpSmkJ8Yj-vtM1uUJwKypw.jpeg" /></figure><p>If you’re in tech, you’ve been hearing a lot of buzz around Artificial Intelligence, Machine Learning, and even Deep Learning. What’s the right word to be using and when? Do they all mean the same thing? I mean, people are sure using it interchangeably all the time.</p><p>Artificial Intelligence, Machine Learning, and Deep Learning are each a subset of the previous field. Artificial Intelligence is the overarching category for Machine Learning. And Machine Learning is the overarching category for Deep Learning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-1BdsVEJlTwuGkUuvGPH4g.png" /><figcaption>Deep Learning is a subset of Machine Learning, and Machine Learning is a subset of Artificial Intelligence.</figcaption></figure><p>The real hype in recent times should all be credited to Deep Learning. This particular field of Artificial Intelligence and Machine Learning is the one that has been solving a ton of interesting problems in recent years — from automated grocery store purchases to autonomous cars.</p><h3>Artificial Intelligence</h3><p>So why have we been hearing so much about Artificial Intelligence? Some might credit Elon Musk and Sam Altman. Elon Musk has been increasingly speaking on the subject of ethics in Artificial Intelligence. I suppose he’s using Artificial Intelligence as familiar jargon used in science fiction media.</p><p>Artificial Intelligence is a wide field encompassing several sub-fields, techniques, and algorithms. The field of artificial intelligence is based on the goal of making a machine as smart as a human. That is literally the initial overarching goal. Back in 1956, researchers came together at Dartmouth with the explicit goal of programming computers to behave like humans. This was the modern birth of Artificial Intelligence as we know it today.</p><h3>AI Goals</h3><p>To further explain the goals of Artificial Intelligence, researchers extended their primary goal to these six main goals.</p><p>1) <strong>Logical Reasoning</strong>. Enable computers to do the types of sophisticated mental tasks that humans are capable of doing. Examples of solving these Logical Reasoning problems include playing chess and solving algebra word problems.</p><p>2) <strong>Knowledge representation.</strong> Enable computers to describe objects, people, and languages. Examples of this include Object Oriented Programming Languages, such as Smalltalk.</p><p>3) <strong>Planning and navigation</strong>. Enable a computer to get from point A to point B. For example, the first self-driving robot was built in the early 1960’s.</p><p>4) <strong>Natural Language Processing</strong>. Enable computers to understand and process language. One of the first projects related to this, was attempting to translate English to Russian and vice versa.</p><p>5) <strong>Perception</strong>. Enable computers to interact with the world through sight, hearing, touch, and smell.</p><p>6) <strong>Emergent Intelligence</strong>. That is, Intelligence that is not explicitly programmed, but emerges from the rest of the explicit AI features. The vision for this goal was to have machines exhibit emotional intelligence, moral reasoning, and more.</p><h3>AI Fields</h3><p>Even with these main goals, this doesn’t categorize the specific Artificial Intelligence algorithms and techniques. These are just six of the major algorithms and techniques within Artificial Intelligence:</p><p>1) <strong>Machine Learning</strong> is the field of artificial intelligence that gives computers the ability to learn without being explicitly programmed.</p><p>2) <strong>Search and Optimization</strong>. Algorithms such as Gradient Descent to iteratively search for local maximums or minimums.</p><p>3) <strong>Constraint Satisfaction</strong> is the process of finding a solution to a set of constraints that impose conditions that the variables must satisfy.</p><p>4) <strong>Logical Reasoning</strong>. An example of logical reasoning in artificial intelligence is an expert computer system that emulates the decision-making ability of a human expert.</p><p>5) <strong>Probabilistic Reasoning</strong> is to combine the capacity of probability theory to handle uncertainty with the capacity of deductive logic to exploit structure of formal argument. The result is a richer and more expressive formalism with a broad range of possible application areas.</p><p>6) <strong>Control Theory</strong> is a formal approach to find controllers that have provable properties. This usually involves a system of differential equations that usually describe a physical system like a robot or an aircraft.</p><h3>Machine Learning</h3><p>Machine Learning is a subset of Artificial Intelligence. So what is Machine Learning anyway? If Artificial Intelligence aims to make computers smart. Machine Learning takes the stance that we should give data to the computer, and let the computer learn on its own. The idea that computers might be able to learn for themselves was divined by Arthur Samuel in 1959.</p><p>And what makes Machine Learning so important? One major breakthrough led to the emergence of Machine Learning as the driving force behind Artificial Intelligence — the invention of the internet. The internet came with a huge amount of digital information being generated, stored, and made available for analysis. This is when you start hearing about Big Data. And Machine Learning algorithms have been the most effective at leveraging all of this Big Data.</p><p><a href="https://medium.com/machinevision/overview-of-neural-networks-b86ce02ea3d1"><strong>Neural Networks </strong></a>⇱</p><p>If we’re talking about Machine Learning, then it’s worth mentioning a popular component of Machine Learning Algorithms: Neural Networks.</p><p>Neural Networks are a key piece to some of the most successful machine learning algorithms. The development of neural networks have been key to teaching computers to think and understand the world in the way that humans do. Essentially, a neural network emulates the human brain. Brains cells, or neurons, are connected via synapses. This is abstracted as a graph of nodes (neurons) connected by weighted edges (synapses). For more information on neural networks, feel free to read our <a href="https://medium.com/machinevision/overview-of-neural-networks-b86ce02ea3d1">Overview of Neural Networks</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aoJHHuA3Qd5SycyC9_hXww.png" /><figcaption>This neural network has one layer, three inputs, and one output. Any neural network can have any number of layers, inputs, or outputs.</figcaption></figure><h3>Deep Learning</h3><p>Machine learning algorithms have been the driving force behind Artificial Intelligence. And the most effective of all the machine learning algorithms has been Deep Learning.</p><p>Deep Learning involves several layers of computation. In this case, “deep” refers to a “large” number of layers. Deep learning can be 20 layers or 1000 layers, but at least more than 2 or 3 layers. Deep Learning has gained momentum recently, not just because of the vast amount of data provided by the internet, but also because of the rise in computational power in the last decade. Specifically, GPU’s have increased the computational power by enabling parallel computations. As you might’ve guessed, Deep Learning is highly parallelizable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0UovTUIDiixK3z5ewLMlfg.png" /><figcaption>This neural network has two layers, three inputs, and one output. Any neural network can have any number of layers, inputs, or outputs. The layers between the input neurons and the final layer of output neurons are hidden layers of a deep neural network.</figcaption></figure><p>The best showcase of Deep Learning is Deep Neural Networks (DNN). A deep neural network is just a neural network with more than two or three layers. However, deep neural networks are not the only type of Deep Learning algorithm — it’s just the most popular. Another Deep Learning algorithm is a Deep Belief Network (DBN). A Deep Belief Network has undirected connections between some layers. This means that the topology of the DNN and DBN is different by definition. The undirected layers in a DBN are called Restricted Boltzmann Machines.</p><h3>Conclusion</h3><p>So one way to think about all three of these ideas is that Machine Learning is the cutting edge of Artificial Intelligence. And Deep Learning is the cutting edge of the cutting edge.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=adb7a5487ac8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/overview-of-artificial-intelligence-buzz-adb7a5487ac8">Overview of Artificial Intelligence Buzz</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Have any machine learning questions?]]></title>
            <link>https://medium.com/machinevision/have-any-machine-learning-questions-25ec868c455?source=rss----cc93edaeae9f---4</link>
            <guid isPermaLink="false">https://medium.com/p/25ec868c455</guid>
            <category><![CDATA[luis-bermudez]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[augmented-reality]]></category>
            <dc:creator><![CDATA[Luis Bermudez]]></dc:creator>
            <pubDate>Sun, 29 Oct 2017 22:03:28 GMT</pubDate>
            <atom:updated>2017-10-29T22:03:27.800Z</atom:updated>
            <content:encoded><![CDATA[<h4>Feel free to ask anything about computer vision.</h4><p>Hi, my name is Luis Bermudez. I’d be happy to answer any of your questions regarding Artificial Intelligence, Machine Learning, Deep Learning, Computer Vision, and Augmented Reality.</p><p>Have any article ideas? Just want to say hello? Email me at <a href="/service/mailto:beluis3d@gmail.com">beluis3d@gmail.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25ec868c455" width="1" height="1" alt=""><hr><p><a href="https://medium.com/machinevision/have-any-machine-learning-questions-25ec868c455">Have any machine learning questions?</a> was originally published in <a href="https://medium.com/machinevision">machinevision</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>