<?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[Stories by Aaron Arellano on Medium]]></title>
        <description><![CDATA[Stories by Aaron Arellano on Medium]]></description>
        <link>https://medium.com/@aaronjobarellano?source=rss-557a5fab3061------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*5PFTr01kXDfB5Td1weRlRA.jpeg</url>
            <title>Stories by Aaron Arellano on Medium</title>
            <link>https://medium.com/@aaronjobarellano?source=rss-557a5fab3061------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 15:15:57 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aaronjobarellano/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Harnessing Xenobots to Combat Microplastics]]></title>
            <link>https://medium.com/@aaronjobarellano/harnessing-xenobots-to-combat-microplastics-a1a1b20a5ec2?source=rss-557a5fab3061------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1a1b20a5ec2</guid>
            <category><![CDATA[plastic-pollution]]></category>
            <category><![CDATA[biotechnology]]></category>
            <dc:creator><![CDATA[Aaron Arellano]]></dc:creator>
            <pubDate>Fri, 10 May 2024 03:53:52 GMT</pubDate>
            <atom:updated>2024-05-10T03:53:52.874Z</atom:updated>
            <content:encoded><![CDATA[<p>Picture a day in your life without touching a single plastic item — from the moment you wake up until you lay your head to rest. Can you?</p><p>The ubiquity of plastics in everyday life has been due to its durability, cost-effectiveness, and versatility. But as we know, its common usage has devastated our environment, leaving dumps of plastics in our forests, rivers, and oceans.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Z4UI9w8hRTOTwDD6" /></figure><p>But one of the most damaging products of plastic pollution has been elusive, killing, and indistinguishable from the human eye: <strong>microplastics</strong>.</p><p>Coming from the plastics we use in our everyday lives, microplastics are the small fragments of plastic bottles, textiles, wrappers (and much more) that measure less than five millimetres across.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aw094-5RZM8hBNH5.jpg" /></figure><h4>Why are microplastics a problem?</h4><p>Simply put, microplastics are extremely persistent. Like any plastic, they do not biodegrade, so they accumulate in the environment over time, threatening the health of the entire marine food chain while polluting the ocean. They also act as carriers for other pollutants, concentrating and transporting toxic chemicals.</p><p>But, unlike macroplastics (plastics larger than 5 millimeters), microplastics bring unique problems to the environment:</p><ul><li>Their small size causes them to be extremely more difficult to detect and capture. As a result, the<strong> 460 million tonnes</strong> of microplastics in the oceans of the world cannot be feasibly cleaned up without expensive actions.</li><li>Microplastics have a higher surface area to volume ratio, allowing them to carry and concentrate significantly more toxins and pollutants.</li><li>Microplastics are more bioavailable than larger plastics to a wider range of organisms, from plankton to fish.</li><li>Their tiny scale causes microplastics to be unnoticeable to the human eye, resulting in the average adult consuming <strong>approximately 2,000 microplastics per year. </strong>These microplastics can cause endocrine disruption, weight gain, insulin resistance, decreased reproductive health, cytotoxicity, and cancer.</li><li>Microplastics can negatively impact industries like agriculture, fisheries, and other livelihoods and communities that rely on healthy ecosystems. An estimated<strong> $426 million</strong> dollars could be saved in just the UK if microplastics were not present.</li></ul><p>Although microplastics are minuscule compared to larger macroplastics, they arguably have more of a extensive chokehold on the ocean. Worst part is that <strong>we do not have a feasible and effective way of cleaning up microplastics</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*APFSUCDqhR-ZcS_f.jpg" /><figcaption>Manually using sifts is the current way to recover microplastics, which is an extremely slow an inefficient process.</figcaption></figure><p>Microplastics isn’t like trying to catch fish in a net; it’s more like trying to sift sand from water with a strainer. Without an efficient and intelligent way of cleaning our oceans of microplastics, they will continue to threaten the marine food cycle, human health, and much, much, more.</p><h4>Xenovation Lab’s solution</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*39dE9VT28IDghdeSIItMTw.png" /></figure><p>At Xenovation, we’re developing a cutting-edge solution towards cleaning microplastic pollution.</p><p>Our solution leverages the remarkable capabilities of <strong>xenobots</strong>, whch are <strong>bioengineered robots crafted from living cells.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1pkzrUJHBDvhC1hm.jpg" /><figcaption>Xenobots in a field of particulate matter.</figcaption></figure><h4>What are xenobots?</h4><p>(pronounced zee-noh bots)</p><p>These xenobots are synthetic, programmable, living cells. Xenobots are genetically engineered cells that use living frog embryos from <em>Xenopus Laevis</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7sdiNX8Yql5RLjui.jpg" /><figcaption>Different designs of xenobots</figcaption></figure><p>Taking the heart and skin stem cells from these frog embryos, we can attach different combinations of these cells together to create specific designs of xenobots. The spontaneous contraction and relaxation of the heart cells within the xenobots act as a “miniature engine” to propel the movement of the overall structure. The skin cells help the xenobot feel objects near them, and will help us collect microplastics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*R7FFFwOGHdJ4GeSA.gif" /><figcaption>The outer layers of cells are the skin cells, and the inner cells propelling the xenobots are the heart cells.</figcaption></figure><p>The xenobots’ behaviour and capabilities emerge from the collective interactions of their cells, and we can observe and control them through the design process. We use evolutionary algorithms to physically construct precise designs of heart and skin cells to genetically modify xenobot behaviours to do whatever we want — push payloads:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*gmF1X_Bfh_fkmCi1.gif" /></figure><p>self-repair,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/816/1*wVpo1aLNOdJEivldzzXsnA.png" /><figcaption>A xenobot repairing itself.</figcaption></figure><p>and even self-replication.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/288/0*0OLzyyU7qCLytNv7" /><figcaption>A parent xenobot (red) with its spherical offspring (green)</figcaption></figure><p>Crazy, right?</p><p>Like any living animal, xenobots demonstrate group behaviors, such as the spontaneous aggregation of particles from joint movement patterns. We use this behaviour to allow a group of xenobots to move together, similar to a school of fish, to collect microplastics from a concentrated area.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JoqYpxyvQjOX6dk2oVvsUw.png" /><figcaption>A group of Xenobots displaying collective behaviour by sweeping out areas of debris.</figcaption></figure><p>Being made of <strong>completely </strong>biological components, xenobots are completely biodegradable and do not harm marine environments. They also do not require external energy.</p><p><strong>Xenovation’s usage of Xenobots</strong></p><p>Xenovation’s solution to eradicate microplastics with xenobots involves this 3 step process:</p><p><strong>1) Data Collection</strong></p><p>Before we can deploy the xenobots, we need to know where the microplastics are located. This is where Xenovation’s USVs and apps come in.</p><p><strong><em>USVs — Unmanned Surface Vehicles</em></strong></p><p>Unmanned surface vehicles, or USVs, are robotic boats that can operate on the ocean’s surface without a human crew. At Xenovation Labs, our USV’s operate autonomously, meaning they can perform its tasks without direct human intervention. It follows pre-programmed routes or can be remotely controlled by operators to navigate to specific areas of interest. The vessel is equipped with various sensors to detect and avoid obstacles, as well as to collect environmental data. With advanced microplastic detectors like hyperspectral imaging, or simply a fine mesh, Xenovation’s USV’s can quickly scoop samples of water within the area it is surveying to detect locations of high concentrations of microplastics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*c7dJ81Z7eQgajijX" /><figcaption>The above image shows an USV called “Waste Shark”, which is the design for which Xenovation Lab’s USV’s are based on.</figcaption></figure><p>Our USV’s operates quietly and with minimal disturbance to marine life and ecosystems. Its electric propulsion system reduces noise and emissions, making it environmentally friendly.</p><p>They can be deployed individually or as part of a fleet of autonomous surface vessels, depending on the scale of the waste management operation and the size of the water body being monitored.</p><p><strong><em>App</em></strong></p><p>But another powerful tool that we can leverage is the monitoring capabilities citizens. Similar to<em> </em>apps like <em>Pollution Tracker</em>, Xenovation Lab’s prospect app has the capacity for the detections of microplastics with caring human participants. For less regularly monitored areas of water such as ponds, small lakes and creeks, citizens can directly inform Xenovation Labs of high concentrations of microplastics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hZYWtRK2bzUoDM-6" /><figcaption>Pollution Tracker’s UI, displaying marked areas of pollution.</figcaption></figure><p>If users find plastic pollution near a body of water, which is a prime indicator for microplastics, user can tag the type of litter they’ve found, instantly notifying Xenovation Labs.</p><p>The app automatically geotags the location where the photo was taken. This information helps build a database of litter hotspots and trends, allowing for targeted cleanup efforts and policy interventions.</p><p>With USV’s and our app, Xenovation Labs can be aware of the locations of microplastics.</p><p><strong>2) Xenobots</strong></p><p>Once the areas of highly concentrated microplastics are detected, this is where the xenobots are brought in. With our USV’s, thousands of xenobots can be deployed in the area, for up to 2 weeks at a time. With this amount of xenobots, we can extract tons of microplastics from the area.</p><p>Altogether, “schools” of xenobots can be used to sweep out areas of pollution, carrying microplastics until recollection. In a larger group, xenobots can greatly resist the tides of water and other environmental threats.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*IYp-hF4Tvmsp3nF7.jpg" /><figcaption>Like a school of fish, Xenovation Lab’s solution involves schools of xenobots.</figcaption></figure><p>Xenobots have been shown to display strong group coordination skills, showing the ability to clean areas from debris.</p><p>These Xenobot schools will follow the behaviours of schools of fish, including:</p><p><strong><em>Shoaling</em></strong></p><p>The Xenobots can align their bodies and swim in the same direction, maintaining a specific spatial arrangement. The Xenobots coordinate their movements, with individual Xenobots focusing on and following the actions of their closest neighbours. This behaviour would be embeded within the Xenobots during the evolutionary design process, where Xenobots that display greater coordination with other Xenobots will be used.</p><p>However, unlike schools of fish, the schools of Xenobots will be held significantly tighter. In order to collectively resist ocean tides and other forces, holding a tight formation (and thus, a larger mass) can allow collective Xenobots to direct their trajectory more steadily.</p><p>Maintaining a shoaling behaviour is essential for the Xenobots towards foraging effectively, being able to repair themselves, self-replication, and avoiding predators/obstacles.</p><p><strong><em>Foraging</em></strong></p><p>By swimming together in a coordinated manner, the school can cover more area and share information about the location of microplastics, a behavior known as “diffusion adaptation.” The collective abilities of the school allow them to make more accurate choices about where to find and access food compared to individual Xenobots. Essentially, schools of Xenobots will navigate areas of microplastics, holding onto microplastics as they pass through them. (similar to how they carry payloads)</p><p>This behaviour expands on the Xenobots pile-making capability, which displays them as being able to collect debris into a concentrated pile.</p><p>However, in a school of Xenobots, they can navigate bodies of water while bringing the piles with them. In doing so, the microplastics can be carefully extracted at a later time.</p><p><strong><em>Obstacle Avoidance</em></strong></p><p>Like schools of fish, schools of Xenobots have key behaviors and adaptations that allow them to effectively avoid predators and navigate through their environment. The large number of Xenobots in a school can create a “sensory overload” for predators such as sea turtles or fish, making it difficult for them to target and capture individual prey.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sJAofpvwqjyLkAxp" /><figcaption>Together, schools of Xenobots can avoid predators and obstacles more effectively.</figcaption></figure><p>Currently, Xenobot’s darker complexion can make them harder to spot from predators, and can cover the microplastic’s shimmering reflections of light.</p><p>As well, Xenobot’s have strong coordination navigating obstacle-filled environments, adjusting their behaviors to effectively maneuver around obstacles.</p><p>A prospective area of research in Xenobots is developing further behaviour to avoid persistent predators, including synchronized evasive maneuvers, like U-turns or splitting into multiple groups, to confuse and deter predators.</p><p><strong><em>Self-Repair, Self-Sustaining, Self-Replication</em></strong></p><p>But, unlike schools of fish, Xenobot’s possess some unique behaviours.</p><p>One of which is self-repair and self-replication. Given the skin and heart cells that they are made of, Xenobot’s can push them into controlled designs, repairing themselves or even creating more Xenobots.</p><p>This self-repair capability allows Xenobots to maintain their structural integrity and functionality, even when faced with disruptions or injuries to their body.</p><p>As well, the Xenobot’s do not require food or energy during their short lifespan.</p><p><strong>3) Recollection</strong></p><p>Eventually, the schools of Xenobot’s will either collect the maximum capacity of microplastics, or, the Xenobots will reach the extent of their lifespans.</p><p>At that point, the USV’s monitoring the bodies of water will recover them. With technologies such as predictive machine learning (to predict Xenobot movements), computer vision, and routine sample analysis, the USV’s can predict endpoints of schools of Xenobots.</p><p>Compared to traditional cleanup methods like manual collection or filtration systems, Xenovation Labs offers many economic advantages. Xenobots can access hard-to-reach areas, and in the next few years, will be able to navigate through these complex environments in the ocean. We estimate that in the coming years, approximately 50,000 Xenobots working together can collect 1 ton of microplastics in about a day. Their biodegradable material eliminates the need for disposal procedures and also minimizes the long-term environmental impact usually associated with cleanup costs.</p><p>According to National Geographic, there are 5.25 trillion pieces of plastic debris in the ocean. Of that mass, 269,000 tons float on the surface. That is a staggering amount, which needs cleaning. With our algorithm, Xenobots’ collective effort will be 10x more efficient than if they work individually.</p><p>Let’s assume that Xenovation targets a large area, like the Great Pacific Garbage Patch, which covers approximately 1.6 million square kilometers. Our Xenobots can cover an area of about 10 square kilometers together. Regarding costs, producing each Xenobot costs approximately $50, considering factors such as materials, labor, and equipment. With 5,000 Xenobots needed for this operation, the production cost would be around $2.5 million.</p><p>This process will take about 6 months to complete. Overall, while there are upfront costs associated with producing and deploying Xenobots, their efficiency and effectiveness in microplastic cleanup make them a cost-effective solution compared to traditional methods. Plus, their biodegradable nature minimizes the long-term environmental impact, which is both economically and environmentally beneficial.</p><p>The removal of microplastics creates a future of improved marine life, human health and economic benefits.</p><p>Marine animals, from tiny plankton to large whales, would suffer less harm from ingesting or becoming entangled in plastic debris. This would lead to healthier ecosystems and more stable marine food webs.</p><p>Additionally, with fewer microplastics in the ocean, there would be reduced risk of these particles entering the human food chain through seafood consumption. This would potentially lower the risk of health issues related to ingesting microplastics, although more research is needed to fully understand the health impacts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0PyYf9EVonj9pFWM.jpg" /></figure><p>Lastly, coastal economies that depend on tourism and fisheries would likely benefit from cleaner waters and healthier marine ecosystems. Cleaner waters mean the revival of coastal communities and the tourism industry.</p><p>For example, national fisheries in Mozambique lose<strong> $88,892,000 a year</strong> from specifically microplastics damaging their quality of goods.</p><p>At Xenovation Labs, we are leading the charge against microplastic pollution. Don’t miss out on this opportunity to invest in a cleaner, brighter future. Join us and be at front of the solution.</p><p>Thanks for reading!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1a1b20a5ec2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Coding a personal AI trainer!]]></title>
            <link>https://medium.com/@aaronjobarellano/coding-a-personal-ai-trainer-8e4a752dad57?source=rss-557a5fab3061------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e4a752dad57</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[fitness]]></category>
            <dc:creator><![CDATA[Aaron Arellano]]></dc:creator>
            <pubDate>Sat, 17 Feb 2024 05:07:46 GMT</pubDate>
            <atom:updated>2024-02-17T05:07:46.595Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*EJ_5KLwvhbdMMuxA" /></figure><p>A lingering feeling I’ve constantly had when I’m working out is if my exercise form is the most effective it can be. I’d hate for me to be arching my back a little too much or bending my elbows not enough, because this leaves a sub-optimal workout. As anyone would, I’d like for my form to be perfect.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*c-SNEcI242GB9X2-.gif" /></figure><p>Naturally, if I feel I need some correcting, I’d either ask a friend or watch a video to correct my form, but a superior option is becoming more readily available in the near future — artificial intelligence!</p><h4><strong>AI? Like ChatGPT or QuillBot? How?</strong></h4><p>Two AI applications can make this possible <strong>— computer vision</strong>, and <strong>machine learning</strong>.</p><p><strong>Computer vision</strong> — Artificial intelligence that enables machines to interpret and make sense of visual information from the world, allowing them to perceive and understand their surroundings through images or videos.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/0*YHrqc1GTwikh4gj9.jpeg" /></figure><p><strong>Machine Learning</strong> — Machine learning is a subset of artificial intelligence that empowers computers to learn from data for a specific output.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eN_G64ggEnw_6xp9.jpeg" /></figure><p>Keep on reading this article to understand how I used computer vision and machine learning to build <strong>PrecisionFit</strong>, a personal fitness AI trainer!</p><p>With only a camera, PrecisionFit users can workout with real-time quality scores evaluating the qualities of their exercise reps. Using intelligent pre-trained AI models, a webcam’s video data is all PrecisionFit needs to provide quick evaluations. Here’s how:</p><p><em>Note: I’m gonna be referring to my code (which I simplified here) throughout this article, so follow along with this </em><a href="https://github.com/AArellano101/PrecisionFit"><em>Github repo</em></a></p><h4><strong>Step 1 — Data Creation</strong></h4><p><a href="https://github.com/AArellano101/PrecisionFit/blob/main/ml/data_creation.py"><em>data_creation.py</em></a></p><p>Artificial intelligence gets so smart by learning (hence, machine learning) and recognizing patterns over large amounts of data. So, PrecisionFit needs to gain insights on large datasets of example exercise reps to be able to analyze users reps in real time.</p><p>Through some extensive searching, I found a variety of videos of certain exercises (push-ups, squats, sit-ups) and stored them in a Firebase database, along with the person’s physical data (height, weight, gender).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KBrgeqHidbW1iVUl" /></figure><p>These videos each had varying qualities of reps, which would help the eventual AI to have a greater understanding of patterns in rep quality during training. In machine learning, having a diverse dataset is essential to generalize well to possible scenarios and to recognize broad patterns and features.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/0*k0HqlXAXKKj0ppUE.gif" /><figcaption>Bad form</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*CJKj03wROETQKKgB.gif" /><figcaption>Good form</figcaption></figure><p>However, the video frame pixels aren’t the datapoints that the AI will be training on, (it could be, but this would be an extremely expensive procedure), rather, the AI will be trained on a metric focusing on the exercise movements.</p><p>I chose this metric to be the angles of the joints taken frame-by-frame. This can accurately represent the exercise movements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*A9k17b8ANUNcWNyYxXM8ng.png" /><figcaption>Angles of variations of squats</figcaption></figure><p>To find the joint angles, I used <a href="https://developers.google.com/mediapipe"><strong>Mediapipe</strong></a> to analyze the videos to give the joint angles in each frame. One feature of Mediapipe is the <a href="https://developers.google.com/mediapipe/solutions/vision/pose_landmarker">pose landmark detection</a>, which detects certain parts of the body and returns its position on the screen with its x and y coordinates.</p><blockquote><strong>Mediapipe</strong> — An open-source computer vision framework</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*CdI4qG1T-IzIsuoR.gif" /><figcaption>Mediapipe in use</figcaption></figure><p>The specific joint angles given the coordinates of different landmarks of the body can be found with a function found in the <strong>NumPy</strong> library, <strong>arctan2</strong>.</p><pre>import numpy as np<br><br>def find_angle(a,b,c):<br>  # a, b, c, are coordinates in the form of tuples  <br>  # with x &amp; y being its first &amp; second elements.<br>  radians = np.arctan2(c[1]-b[1], c[0]-b[0]) - np.arctan2(a[1]-b[1], a[0]-b[0])<br>  return radians</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/687/0*q_HRNIbOfxygb2iT" /><figcaption><em>Wrist, elbow and shoulder coordinates give elbow joint angle.</em></figcaption></figure><p>The joint angles of a video can then be found frame-by-frame, and then saved in the following code:</p><pre>import cv2 as cv<br># cv2, or OpenCv is another computer vision library that&#39;ll help with mediapipe<br>import mediapipe as mp<br><br># MediaPipe pose landmarks tools<br>mp_pose = mp.solutions.pose<br>... <br># Each video from Firebase dataset is temporarily saved in this directory.<br># The following code is ran for each video:<br>...<br><br>cap = cv.VideoCapture(&#39;path/to/temporarily/saved/video&#39;) # stores video in variable<br><br># Pose instance<br>pose = mp_pose.Pose(min_detection_confidence=0.5, min_tracking_confidence=0.5)<br><br>frame_number = 0<br><br>txt_data = [] # Where the joint angles will be stored<br><br># This ensures all frames of the video are read<br>while True:<br>    # cap.read() reads one frame of the video, giving ret and frame<br>    # ret gives whether the function was successful, and frame gives the frame<br>    # Once the entire video frame has been iterated over, ret is False<br>    ret, frame = cap.read()<br>    if not ret: <br>        break<br>    <br>    # converts the frame into a readable format for Mediapipe<br>    image = cv.cvtColor(frame, cv.COLOR_BGR2RGB)<br>    <br>    ...<br>    <br>    results = pose.process(image) # processes image<br>    <br>    lms = results.pose_landmarks # gives list of landmarks<br>    <br>    frameAngles = {}<br>    # angles is the list of angles needed to be found for the given exercise<br>    # joints is a list that shows which landmarks are used in each angle<br>    for angle in angles: <br>        frameAngles[angle] = find_angle(*[lms.landmark[i] for i in joints[angle]])<br>  <br>    # append_angles is not shown, but it&#39;s a function that appends the the frames angles to txt_data (in proper format)<br>    txt_data = append_angles(frameAngles, frame_number, txt_data)<br>    frame_number += 1<br><br><br>...<br># txt_data is converted into a .txt file, which is then stored into Firebase</pre><h4><strong>Step 2 — Model Training</strong></h4><p><a href="https://github.com/AArellano101/PrecisionFit/blob/main/ml/neural.py"><em>neural.py</em></a></p><p>In <strong>neural.py</strong>, I used these formatted joint angles to be trained on by a <strong>Tensorflow</strong> AI neural network:</p><blockquote><strong>Tensorflow</strong> — An open source machine learning library</blockquote><p>Firstly, I iterated over the .txt files stored in the Firebase database, temporarily saving them in a CSV file.</p><pre>import csv # library containing functions to write csv files<br><br>...<br># data stores all txt files from Firebase<br>for file in data:<br>    # file is saved into directory<br>    ...<br><br>    with open(&#39;path/to/current/txt/file&#39;, &#39;r&#39;) as f:<br>        csv_line = []<br>        csv_line.append(f.readline().split(&#39;,&#39;))<br><br>        with open(&#39;path/to/csv/file&#39;, &#39;a&#39;) as c:<br>            w = csv.writer(c)<br>            w.writerow(csv_line)  <br>  <br>    # file is removed from directory</pre><p>With the <strong>Pandas</strong> library, the function <strong>read_csv</strong> translated the CSV file into tabular data that <strong>Keras</strong> can read.</p><blockquote><strong>Pandas — </strong>A Python data analysis library</blockquote><blockquote><strong>Keras — </strong>The TensorFlow neural network interface</blockquote><pre>import pandas as pd<br><br>ds = pd.read_csv(&#39;path/to/csv/file&#39;)</pre><p>With this data, I arbitrarily split them into training and testing data.</p><blockquote><strong>Training data</strong> — The data the neural network will find patterns in.</blockquote><blockquote><strong>Testing data</strong> — The data the neural network will use to see if the patterns it found were correct.</blockquote><pre># splits the tabular data into x (input) and y (output) values<br>dsx = ds.iloc[:, :-1].values <br>dsy = ds.iloc[:, -1].values<br><br># randomly splits data into 20% test data and 80% training data <br>X_train, X_test, Y_train, Y_test = train_test_split(dsx, dsy, test_size=0.2, random_state=42)</pre><p>Finally, this data can be used to be trained on a <strong>Keras</strong> model.</p><p>Firstly, my neural network will be using Keras’ sequential model, set to a variable here.</p><pre>import tensorflow as tf<br><br>model = tf.keras.Sequential()</pre><p>A sequential model is a type of neural net that has layers of nodes in sequence. Visualized, it looks like this:</p><figure><img alt="Building A Deep Learning Model using Keras | by Jaz Allibhai | Towards Data  Science" src="https://cdn-images-1.medium.com/proxy/1*eJ36Jpf-DE9q5nKk67xT0Q.jpeg" /></figure><p>The method <strong>model.add</strong> adds a hidden layer with a certain number of input nodes, described by the variable <strong>inputNodes </strong>(which varies from exercise to exercise). This gives an input layer and hidden layer, shown by the red and blue nodes in the graph (of course, it has a different size).</p><pre>model.add(tf.keras.layers.Dense(25, input_dim=inputNodes, activation=sin_activation))</pre><p>For the other hidden layers, the following code uses the same method to add another 3 layers of 25 hidden nodes.</p><pre>model.add(tf.keras.layers.Dense(25, activation=sin_activation))<br>model.add(tf.keras.layers.Dense(25, activation=sin_activation))<br>model.add(tf.keras.layers.Dense(25, activation=sin_activation))</pre><p>In each of these layers, they have a custom <strong>sin </strong>activation function. An activation function is an operation that the nodes inputs are applied to, determining how the node responds to it. A <strong>sin </strong>activation function is one that applies the trigonometric <strong>sin </strong>function on the values.</p><p>The output layer is shown by this following layer, giving one final node. It has a <strong>sigmoid</strong> activation function, which keeps the output between 0 and 1.</p><pre>model.compile(loss=&#39;mean_squared_error&#39;, optimizer=&#39;adam&#39;, metrics=[&#39;mean_squared_error&#39;])</pre><p>Next, I trained the data using a <strong>mean squared error</strong> loss metric, which finds the average squared difference between the estimated value and the actual value.</p><blockquote><strong>Loss metric </strong>— A metric used to evaluate a model’s performance, comparing predictions and actual results.</blockquote><pre>model.compile(loss=&#39;mean_squared_error&#39;, optimizer=&#39;adam&#39;, metrics=[&#39;mean_squared_error&#39;])<br>model.fit(X_train, Y_train, epochs=EPOCHS, verbose=1)</pre><p>I evaluated the model using the testing data I got earlier, returning the value of the loss metric.</p><pre>test_loss = model.evaluate(X_test, Y_test)<br>print(f&quot;Test Mean Squared Error: {test_loss[1]}&quot;)</pre><p>Finally, I saved the model somewhere in the directory where the app can later use it in real-time.</p><pre>model.save(&quot;path/to/usable/directory&quot;)</pre><p><strong>Step 3 — Live Usage (simplified)</strong></p><p>Now that I have a trained model, I can use this in real-time in an app.</p><p>I used a <strong>Flask</strong> app, which is a Python-based web framework.</p><pre>from flask import Flask, render_template, Response, request, jsonify<br>from webcam import generate_frames<br><br>app = Flask(__name__)<br><br>@app.route(&quot;/&quot;)<br>def index():<br>    return render_template(&quot;index.html&quot;) # the function render_template, obviously renders HTML pages<br><br>@app.route(&quot;/workout&quot;)<br>def workout():<br>    return render_template(&quot;workout.html&quot;)<br><br>@app.route(&#39;/video&#39;, methods=[&quot;GET&quot;]) # this function is reserved for GET HTTP methods<br>def get_video():<br>    print(&#39;Turning on webcam.&#39;)<br>    return Response(generate_frames(app_info), mimetype=&#39;multipart/x-mixed-replace; boundary=frame&#39;)</pre><p>In the above snippet, each route specified in the <strong>@app.route</strong> decorator runs the function below it whenever it is reached. The routes of<strong> “/”</strong> and <strong>“/workout”</strong> call the function <strong>index()</strong> and <strong>workout(),</strong> each rendering HTML pages.</p><p>In <strong>get_video()</strong>, however, rather than returning a HTML page, this function returns the frames of the computer’s camera, using a <strong>generate_frames</strong> function. I’ll be using this route to display the app’s webcam.</p><pre>import cv2 as cv<br>import mediapipe as mp<br><br># A lot of this code is similar to the snippet in Step 1<br>def generate_frames(data):<br>    # MediaPipe pose landmarks tools<br>    mp_drawing = mp.solutions.drawing_utils<br>    mp_pose = mp.solutions.pose<br><br>    ...<br>   <br>    cap = cv.VideoCapture(0)<br>    pose = mp_pose.Pose(min_detection_confidence=0.5, min_tracking_confidence=0.5)<br>    frame_number = 0<br><br>    while True:<br>        ret, frame = cap.read()<br><br>        if not ret:<br>            break<br><br>        image = cv.cvtColor(frame, cv.COLOR_BGR2RGB)<br>        image.flags.writeable = False<br><br>        results = pose.process(image)<br>            <br>        image.flags.writeable = True<br>        image = cv.cvtColor(image, cv.COLOR_RGB2BGR)<br><br>        lms = results.pose_landmarks<br>  <br>        # if the workout is active, woActive = True, and is False otherwise<br>        if woActive:<br>            frameAngles = {}<br>            for angle in angles: <br>                frameAngles[angle] = find_angle(*[lms.landmark[i] for i in joints[angle]])<br>            txt_data = append_angles(frameAngles, frame_number, txt_data, printInfo=False)<br>            frame_number += 1<br>        <br>        # draws landmarks onto webcam frames <br>        mp_drawing.draw_landmarks(image, lms, mp_pose.POSE_CONNECTIONS)<br><br>        # formats frame into streaming data<br>        ret, buffer = cv.imencode(&#39;.jpg&#39;, image)<br>        image = buffer.tobytes()<br><br>        # instead of returning, the function yields each frame to continuously stream frames<br>        yield (b&#39;--frame\r\nContent-Type: image/jpeg\r\n\r\n&#39; + image + b&#39;\r\n&#39;)</pre><p>In <strong>generate_frames</strong>, I use the <strong>Mediapipe</strong> library again to place the pose landmarks on the screen.</p><p>As well, in this function, when the workout is active the function will record the joint angles in each frame, similar to the way I did in Step 1.</p><p>The <strong>/video</strong> route will also handle POST requests. Whenever a POST request is sent, it starts or completes the rep to let the <strong>generate_frames</strong> function know whether to record the data or not.</p><p>Whenever the rep is completed, the recorded data is used in the trained model to output an accuracy score for the rep.</p><p>In some functions found in the Javascript written in the HTML page, this POST request is sent to complete or start reps, making up a complete workout. Then, accuracy scores are received through a GET request.</p><pre>async function toggleRep(start=true, exercise) {<br>    const postData = {<br>        method: (start) ? &quot;start-workout&quot; : &quot;end-workout&quot;,<br>        exercise: exercise,<br>    }<br><br>    // Sending fetch get request to change method.<br>    const response = await fetch(&quot;link-of-website/video&quot;, {<br>      method: &quot;POST&quot;,<br>      mode: &quot;cors&quot;,<br>      headers: {<br>        &quot;Content-Type&quot;: &quot;application/json&quot;,<br>      },<br>      body: JSON.stringify(postData),<br>    })<br><br>    if (!start) {<br>        // getting value of accuracy score<br>        let response = await fetch(&quot;link-of-website/get-model-response&quot;, {<br>            method: &quot;GET&quot;,<br>            mode: &quot;cors&quot;,<br>        })    <br>        <br>        let data = await response.json()<br>        if (data[&#39;response&#39;] != NaN) {<br>            let prediction = Number(data[&#39;response&#39;])<br>             // using jQuery, show accuracy score on HTML page<br>            $(&#39;#accuracy-score&#39;).text(`${prediction}%`)        <br>        }<br>    }<br>}</pre><p><strong>Step 4 — Outcome</strong></p><p>The final outcome will look like this:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F-x1EQg9IAEE&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D-x1EQg9IAEE&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2F-x1EQg9IAEE%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c622add175da8e626104b0662be626a6/href">https://medium.com/media/c622add175da8e626104b0662be626a6/href</a></iframe><p>When the workout begins, in every rep, the POST request is sent to notify <strong>generate_frames</strong> to record the joint angles. After every rep, another POST request is sent to end the rep, applying the recorded joint angles on the trained model, and the output is displayed on the webpage. Then, another rep begins until the workout is over.</p><p>This project took a few months and some hard dedication! This project replicates <a href="https://github.com/shreyanshsatvik/personal_trainer-using-mediapipe">this repo</a> and <a href="https://www.youtube.com/watch?v=06TE_U21FK4">tutorial</a> I found that similarly uses Mediapipe for personal training, but my project also uses AI to evaluate performance, similar to <a href="https://www.youtube.com/watch?v=tPYj3fFJGjk">this tutorial</a>.</p><p><strong>Next Step — Feedback Feature?</strong></p><p>One limitation of PrecisionFit is that it provides only an accuracy score. During the data creation process, it was much less consuming to label each video with only an accuracy score. If I were to expand PrecisionFit to include more outputs, such as feedback to improve form, I would have to manually label each video for a re-train.</p><p>Thanks for reading!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e4a752dad57" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Should boxing and MMA fans look to AI as judges?]]></title>
            <link>https://medium.com/@aaronjobarellano/should-boxing-and-mma-fans-look-to-ai-as-judges-b0c3e909fb57?source=rss-557a5fab3061------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0c3e909fb57</guid>
            <category><![CDATA[boxing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[mma]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Aaron Arellano]]></dc:creator>
            <pubDate>Thu, 02 Nov 2023 16:02:16 GMT</pubDate>
            <atom:updated>2023-11-02T16:02:16.949Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Aaron Arellano</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/770/0*TItLsbGIAv_65Yjc" /><figcaption><em>Vasily Lomachenko (left), Devin Haney (right)</em></figcaption></figure><p>Earlier this year, boxing lightweight contender Vasily Lomachenko challenged current undisputed lightweight champ Devin Haney.</p><p>Throughout the fight, myself and thousands of fans watched a historic clash of two highly skilled fighters. The bout was nothing short of a brillant showcase of talent, with both Lomachenko and Haney giving it their all in the octagon. As the final bell rang, the spotlight shifted from the fighters to the judges’ decision. Unanimous decision for the official winner, Devin Haney.</p><p>Although the fight was very close, myself, the majority of fans, and even fighters seemed to be in agreement: the decision was just plain wrong. We found ourselves wondering if we had watched the same fight as the judges, as the unanimous decision in favor of Haney felt undeserved (despite Haney’s efforts, Lomachenko’s aggression, pressure, and landed punches earned him the fight in our hearts).</p><p>It was disheartening to see such an incredible fight overshadowed by a bad decision. Although, that feeling is not unfamiliar — throughout the history of MMA and boxing, corrupt, biased, and straight out incorrect judges have given decisions to the wrong fighter, ruining the reputations of the sports, every. single. time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LyJ2YsMseJvBfFrx" /><figcaption><em>During the 1988 Olympics, boxing legend Roy Jones Jr. (left) lost to Park Si-Hun (right) for the gold medal despite outlanding him 86–32.</em></figcaption></figure><p>After the fight, I wondered how combat sports should handle bad judges — more fines? better training? reforms? Then, something came to mind — <strong>should AI be used as judges? </strong>After all, artificial intelligence could provide analysis, objectivity, and data processing on another level compared to human judges. Can combat sports be the next industry AI will conquer?</p><p><strong>AI key concepts explained</strong></p><p>Over the course of the past year, the artificial intelligence industry has grown by <strong>37% </strong>(CAGR). With ChatGPT, Dall-E, and many other AI innovations, it’s been hard to avoid the usage of AI in our everyday lives. Although it may seem complex, the way AI analyzes data and produces a sophisticated answer can be abstracted into these concepts:</p><p><em>Machine Learning:</em></p><p>To provide intelligent answers, AI-powered tech has to learn. The machine learning process involves feeding the system with lots of data, showing it countless examples of things we want it to understand. This then teaches the system about patterns about the data. Through understanding these patterns, the system can make predictions related to the data, much like how we improve our skills by learning from our experiences or practicing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W35Lii6sVb9LYePf" /></figure><p>For example, machine learning systems leaning about traffic might recognize driver routes and behaviours by monitoring a road over a long period of time. Over time, the system will find patterns in the drivers and will be able to make predictions about the drivers.</p><p><em>Neural Networks:</em></p><p>The systems that use machine learning may also use neural networks, which allow the system to understand and distinguish patterns in data similar to how the human brain does. These use nodes, similar to neurons in our brains, which slowly develop “rules” towards data to help understand them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/982/0*YLy4EQno-i_Ajco_" /></figure><p>The image above is a diagram of neural networks: each circle represents a node, where different data is processed. The blue nodes represent the input data to be processed, with the green nodes indicating the data interpreted from the input data (this is called the hidden layer of nodes, because they aren’t accessible to the user). Each node’s data entails different information for the next node, all eventually leading to output data or information (red circle).</p><p>For example, if a neural network were programmed to distinguish images of different fruits, it would start by learning about the fruits’ different qualities, including the colours, shapes, and textures.</p><p>Eventually, the neural network would start forming rules like, “round and red could be an apple,” or “curved and yellow could be a banana.” Overtime, these rules would become more and more accurate.</p><p>In terms of the diagram above, the blue input nodes would each represent the fruit’s different qualities, the hidden layer would be where the neural net would develop rules, and the output data is the type of fruit it has determined.</p><blockquote><a href="https://www.ibm.com/topics/neural-networks">More info</a></blockquote><p><em>Deep Learning:</em></p><p>“Deep” Learning means that the neural network uses significantly more layers of nodes, allowing the system to have more rules to understand more intricate patterns in data.</p><p>On the diagram from earlier, a deep learning model would simply have more layers of green hidden layers.</p><p>In our fruit example, more rules allow subtle distinctions to be recognized between different fruits through deep learning, including unique textures or variations of shades of colour.</p><p><strong>Implementation in the sports</strong></p><p>Using these concepts, AI could be implemented in combat sports.</p><p>Like any other job, judges are subject to human error. This includes bias, subjectivity, inattentiveness, and many, many more. Through the past year, artificial intelligence has reduced such errors in practices such as <a href="https://www.fiercebiotech.com/medtech/philips-unveils-new-ai-powered-platform-improve-diagnoses-imaging-scans">medical diagnosis</a>, <a href="https://www.techopedia.com/precision-farming-how-ai-and-drones-are-reshaping-agriculture">agriculture</a>, and even <a href="https://learn.tryq.ai/blog/portfolio-protection?utm_campaign=Forbes&amp;utm_medium=article&amp;utm_source=Forbes">investing</a>.</p><p>So why not judging fights?</p><p>For one, lifelong fans and combat sports experts might not understand the capabilities of AI. AI won’t just tally punches and kicks to determine the winner from whichever fighter landed more; deep learning and neural networks allows the AI to completely understand complex fight factors based on historical and current data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/0*v1XU3d3wkkomLZmY.gif" /></figure><p>The neural net would interpret millions of clips/videos like the one above by creating rules such as “controlling the center of the octagon shows higher control of the match”, or “causing an opponent to miss shots shows higher technical and possible superior ability.” (These rules are like the ones made in our fruit example earlier)</p><p>These neural nets would consider the fights’ input data and use the rules that it has developed over large databases of fights to determine the winner.</p><p><strong>Jabbr — Deepstrike</strong></p><p>This idea is being implemented by <a href="https://jabbr.ai/about">Jabbr</a>, a Danish tech startup building a computer vision AI designed for boxing called DeepStrike.</p><p>With DeepStrike, visual interpretation of the head, body, legs, and gloves while detecting their movements provides DeepStrikes neural network input data. Interpretation of this data provides punches, blocks, and 48 other factors that are used to determine 5 main factors: landed punches, damage, thrown punches, pressure and aggression.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Kfs7EqpI1f-EplkA" /><figcaption><em>Deepstrike being used in the Canelo vs GGG fight.</em></figcaption></figure><p>Deepstrike then evaluates the fight to determine the winner, using an objective scoring system trained on deep learning and neural networks over thousands of datasets. These include sparring footage, professional matches, and more. Deepstrike exhibits the current and future capabilities of AI in combat sports:</p><ul><li>Consistency</li><li>Objectivity</li><li>Data-driven insights</li></ul><p>However, one flaw Deepstrike has is its understanding of the evaluated factors: the AI developers themselves have deemed certain factors to be more/less important by varying degrees, which would rather be more accurate from the input of commissions, judges, and experts.</p><p>For example, Deepstrike may value high impact punches the most to determine the winner, where professional judges may think pressure and aggression is more important. It is essential to value the inputs of professionals in the industry, rather than relying on the developers.</p><p>With an improved system, AI judging fights could be determined through this process:</p><p><strong>1. Computer vision of video footage</strong></p><p>One convenience that Deepstrike possesses is its ability to process data with any video footage.</p><p>This is made possible through computer vision, which is a usage of neural networks for computers to determine certain patterns and objects in a picture or video.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L_Ui0fc6TuRWYDSS" /></figure><p>In this picture, computer vision in Deepstrike is used to distinguish body position, head movement, punches and other actions of the fighters. Computer vision is also used in other sports such as tennis and soccer to track players movement and ball movement.</p><p>Computer vision would be a staple towards developing a convenient judge for AI, as data about fights can be derived directly from video information.</p><p><em>Possible technologies: OpenCV, Viso Suite</em></p><p><strong>2. Deep Learning using extended datasets</strong></p><p>The data gathered from computer vision can then be used towards a deep learning model.</p><p>Similar to Deepstrike, a deep learning model would allow the AI system to understand the patterns of combat sports. Punches, blocks, effective damage, and many other parameters would be evaluated for the primary factors in the fight: landed punches, damage, thrown punches, pressure and aggression. These factors would be used as the main indicators of a win or loss in a fight.</p><p>However, the extent to which these factors contribute towards a win would be unlike Deepstrike. This will use a separate dataset of articles, videos, and interviews of combat sport experts, providing data towards what factors should be considered more than another. This dataset would establish a model towards the weighting of each factor, rather than relying on the AI developer (as Deepstrike does) to manually input their weightings.</p><blockquote><a href="https://www.youtube.com/watch?v=JhfAScxmrGA">Interviews such as this</a> or <a href="https://www.theguardian.com/sport/queensberry-rules-boxing-blog/2019/jan/02/boxing-judge-scoring-528-fights">this</a> would be part of this dataset.</blockquote><p>As a result, the parameters of the fight can be fine-tuned to be like an actual professional judge.</p><p><em>Possible technologies: TensorFlow, Caffe</em></p><p><strong>3. Professional input on model</strong></p><p>The system can be subjected to AI bias, a controversy spreading throughout the machine learning industry. AI bias is a phenomenon where an AI model produces results that are prejudiced due to incorrect assumptions in the machine learning process.</p><p>In 2018, Amazon’s recruiting engine used machine-learning on job applications to determine acceptable applicants. However, the engine developed a preference for male candidates due to the historical data it was trained on, which largely consisted of resumes from male applicants. The engine largely discriminated against women, as the data was misrepresentative towards acceptable job applications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lyU5Ludr0C5NSBhF" /><figcaption><em>Because there were more male employees in Amazon, the engine deemed male applicants to be more successful.</em></figcaption></figure><p>Our system may be just as susceptible to prejudiced or incorrect AI bias, so the necessary maintenance and adjustment of the model is necessary. Beyond possible model bias, the AI can also make models that combat sports experts may disagree with. If either patterns are made, the datasets and models would be intensely reviewed.</p><p><strong>4. Rinse and repeat</strong></p><p>Repeat steps 1–3 over, and over, and over, until perfection.</p><p>Our AI system won’t be perfect, but through more data will the AI be more fine-tuned. Over time, with the correct adjustments and larger datasets, our AI judge will become more and more perfect.</p><p>Through a longer period of time, our AI will also develop certain beneficial habits, such as correcting its own biases or handling fringe data cases.</p><p><strong>Should AI be used as judges?</strong></p><p>I think so! With our AI system, the AI industry can expand its borders to the combat sports industry providing a fresh sense of objectivity, data-driven analytics, and consistent behaviour. AI judges could revolutionize the way we assess combat sports, reducing the impact of human biases and subjectivity. With a fresh outlook and a self-correcting system, bad judging decisions can be made extinct.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5eBWFz-pWXf2apLz" /></figure><p><strong>Sources</strong></p><p>Thanks for reading!</p><p><a href="https://www.ibm.com/topics/neural-networks">What are Neural Networks? | IBM</a></p><p><a href="https://jabbr.ai/about">About | Jabbr.ai</a></p><p><a href="https://www.reuters.com/video/watch/how-ai-could-revolutionize-boxing-idRW604622082023RP1">How AI could revolutionize boxing | Reuters Video</a></p><p><a href="https://hbr.org/2019/10/what-do-we-do-about-the-biases-in-ai">What Do We Do About the Biases in AI?</a></p><p><a href="https://d1wqtxts1xzle7.cloudfront.net/63223371/Essay_On_Ethics___AI_Hiring20200506-124352-1cpydva-libre.pdf?1588823614=&amp;response-content-disposition=inline%3B+filename%3DAn_overview_of_ethical_issues_in_using_A.pdf&amp;Expires=1698907852&amp;Signature=IuliFxpwI-4dhlnFJ7RiA7WfaC0zmOM7YU4DelCmVzZH39-sD5KnoRYjq-uzNm2NtT~66GF6cci5fTa~jr-gHLVV4JrpIndLjcgL2HZ~Jy2a9Jm~XNMWSjaxQ6b5Fz0P2W-5HhaQroZZvt4QXSKTcfEMDAIFg8txy0nUckZuHV2eUVXFZ~Jkyi5Tn2yaXRlfleK7ypilulMTQrTBpZBUDPIeR7L~82IV-4tUKocUBs1WGVe5v9g5EFWSicFIK~p3a8~lT4MOSjmW9h4EQA~eRPMNlV0aCP8Aaw9E-xbx54ZbGQzSokCRko5NuftH04h9zomheARl4amMgxBBVDrHUw__&amp;Key-Pair-Id=APKAJLOHF5GGSLRBV4ZA">An overview of ethical issues in using AI systems in hiring with a case study of Amazon’s AI based hiring tool</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0c3e909fb57" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>