<?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 Aydin Javadov on Medium]]></title>
        <description><![CDATA[Stories by Aydin Javadov on Medium]]></description>
        <link>https://medium.com/@ajavadov99?source=rss-1375b2d0c883------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*tciZIGBN-XO79VgUfEI1Aw.jpeg</url>
            <title>Stories by Aydin Javadov on Medium</title>
            <link>https://medium.com/@ajavadov99?source=rss-1375b2d0c883------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 01:35:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ajavadov99/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[PointNet ++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space [Paper Go-Through]]]></title>
            <link>https://medium.com/@ajavadov99/pointnet-deep-hierarchical-feature-learning-on-point-sets-in-a-metric-space-paper-go-through-b9121e559549?source=rss-1375b2d0c883------2</link>
            <guid isPermaLink="false">https://medium.com/p/b9121e559549</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[3d]]></category>
            <category><![CDATA[pointnet]]></category>
            <dc:creator><![CDATA[Aydin Javadov]]></dc:creator>
            <pubDate>Tue, 30 Aug 2022 18:04:35 GMT</pubDate>
            <atom:updated>2023-02-28T14:39:58.751Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>PointNet++ [1] is an extension of its predecessor PointNet [2], which has been the pioneering work in the sphere of deep learning on point sets. PointNet++ addresses the issue of capturing the local structures within the space of points — an aspect that PointNet has been falling short. As a toolset to build further on PointNet, PointNet++ introduces the concept of hierarchical neural networks. PointNet is then recurcively applied to the point groups generated from the nested partitioning of input points. PointNet++ utilizes the metric space distances between the points and enables the local feature learning with incremental contextual scaling. Moreover, novel set learning layers are introduced to address the poor performance of the network arising uniform density point set inputs, thus enabling adaptive and arbitrary combination of various scales. According to experimental results, PointNet++ is surpassing the state-of-the-art methods by a large margin in the context of deep point set feature learning.</blockquote><h4><strong>Motivation &amp; Introduction</strong></h4><p>Nowadays, many new approaches are emerging in the field of computer vision thanks to technological advancements as well as ongoing research in the artificial intelligence domain. As a consequence of enhancements in large-scale 3D geometry datasets, 3D sensors as well as the improvements in 3D deep learning, 3D vision attains a progressively vital role in plentiful application areas [3]. In this writing, we will be looking at one of the major methods presented in the field of 3D vision, PointNet++.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*guTcWZdxEKqr4kJ-TqQZng.png" /><figcaption>Figure 1: Partitioning of Point Sets (example). Colors indicate the depth (distance from the viewer).</figcaption></figure><p>The main goal of the paper is to analyze and elaborate on the 3D input data that is in the form of geometric point sets. By definition of the set, those point sets are expected to be invariant to permutations — that is, the order in which we would arrange the points in the Euclidean set should not matter for the performance of the algorithm. PointNet++ surpasses its predecessor PointNet by considering the local context. This means that PointNet++ is aware of some specific part of the overall object, whereas PointNet falls insufficient in this aspect. In order to function properly and achieve what it strives for, PointNet++ has a number of steps to follow. First of all, PointNet++ requires the data to be in the form of points as described above. Next, the points are partitioned for further potential feature extractions, as can be seen in Figure 1.</p><p>Let us start with discussing the way partitioning is handled. PointNet++ considers each partition as a mathematical ball. Each ball represents a particular neighborhood of the object, e.g. right corner of the table. We need two pieces of information to be able to form a ball. Each ball needs to have a centroid coordinate, defining its location, and a scale, defining its size. The procedure for centroid coordinate assignment is done with Farthest Point Sampling (FPS) algorithm. In FPS, the first centroid is a random point that is chosen among the set. Then the second farthest point to the first centroid becomes the second centroid. The third centroid is the point that is farthest from the first two centroids. This process continues for up to k centroids. Once the centroids are identified, the next step is to assign a scale to each of them. Generally, a fixed scale for all centroids is not favored, because all the neighborhoods of an object do not have to be of the same size. Another reason to avoid fixed scaling is that the input density of the points usually is not uniform. As a solution for that, dynamic scaling for the neighborhoods can be obtained via random input dropout during the training phase. Indeed, the ability to handle neighborhoods with various scales is considered to be a significant contribution by PointNet++. Acquiring both centroid and scaling information completes the partitioning step.</p><h4>Problem Statement</h4><p>In the essence, given a 3D object, PointNet++ extracts the semantic information about the object. This semantic information can be about classification or segmentation. While doing classification we aim to identify the class/category to which the object belongs, whereas in segmentation we aim to identify the different parts within the object.</p><blockquote><strong>Definition.</strong> Suppose X = (M,d) is a discrete metric space whose metric is inherited from an n-dimensional real valued Euclidean space, where X is the set of points and d is the distance metric. We are then looking for f such that: <br>f : X <strong>→ </strong>Semantic information about X<br>It is important to note that the density of does X not have be uniformly spread in the given space.</blockquote><h4><strong>Method</strong></h4><p>After the partitioning described above, the actual local feature extraction takes place. Local features are abstracted out from each of the partitions. PointNet++ chooses its predecessor, PointNet, to be the local feature learner. That is, PointNet++ is applying PointNet recursively for each of the pre-selected partitions (neighborhoods). To understand how this step is done in detail, we must remind ourselves of the working principle of PointNet. Figure 2 depicts the architecture of PointNet. PointNet is a universal continuous point set function approximator.</p><blockquote><strong>Definition.</strong> Given an unordered point set of {x₁,x₂,…,xₙ}, we define a set function such that:<br> f(x₁,x₂,…,xₙ)=γ(max{h(xᵢ)}), for i=1,..,n, where h and γ are multilayer perceptrons.</blockquote><p>It is important to note that PointNet is able to arbitrarily approximate any continuous set function and h(xᵢ) can be interpreted as the spatial encoding of the xᵢ point in the partition. PointNet, indeed, is a powerful tool for 3D image classification and segmentation. It demonstrates an impressive performance on a few benchmarks, but it lacks information about the local context as we discussed earlier. Hence, PointNet++ leverages PointNet’s power by applying it to each of the local neighborhoods.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eP6pA-7Ou_ttIiY9L8QX9w.jpeg" /><figcaption>Figure 2: The architecture of PointNet</figcaption></figure><p>Now we can delve deeper to talk about how the features are learned in PointNet++ and how exactly PointNet is helping us with that. When the topic is feature learning, convolutional neural networks (CNN) come in handy. PointNet++ does not use CNN but it uses CNN’s idea of hierarchy in features, e.g. herds have cows, cows have eyes, eyes have pupils, etc. Thus, PointNet++ uses hierarchical point set feature learning. Figure 3. demonstrates the overall structure of this learning scheme. As can be seen in Figure 3, the structure consists of several set abstraction levels (SAL). A set abstraction level is formed by a) sampling layer, b) grouping layer, c) pointnet layer. Let us have a closer look at each of these layers. Given the input set of points, all the abovementioned sampling of centroids via the FPS algorithm takes place in the sampling layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/927/1*FbauLYfpkMJ1dysoDf280g.jpeg" /><figcaption>Figure 3: Hierarchical Point Set Feature Learning architecture of PointNet++</figcaption></figure><p>The subsequent grouping layer uses a ball query to group the points that are within a particular ball radius. Naturally, each of the groups may contain different numbers of points. The grouping layer defines the scale of the neighborhood. Finally, in the pointnet layer PointNet algorithm is applied to each of the neighborhoods. As an output for each neighborhood, pointnet layer returns the centroid and a local feature encoding. Then PointNet++ takes those neighborhoods and their feature encoding and applies set abstraction levels repeatedly to extract the features hierarchically. It is crucial to mention that in the grouping layer we used a fixed ball radius. However, in the beginning, we mentioned that having a fixed scale for a neighborhood is not favored, because features learned in dense data may not generalize well for sparse regions and vice versa. Table 1. Shows the input and output shapes of each layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pn8h4QY_if9awGwD8uuOcw.png" /></figure><p>In order to address the issue of fixed partition scale, PointNet++ uses density adaptive pointnet layers and comes up with two types of grouping layers: a) multi-scale grouping (MSG), b) multi-resolution grouping (MRG). Figure 4. depicts the structure of MRG and MSG. MSG applies grouping layers with different scales followed by pointnet layer for each scale. Features at different scales are glued together to form a multi-scale feature. Alternatively, MRG glues the previous layer’s feature summary to the feature vector obtained by directly processing all raw input points through a single pointnet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/623/1*mPawT5BBCEIWhrIJq9fdAA.jpeg" /><figcaption>Figure 4: (a) MSG: Multi-scale grouping (b) MRG: Multi-resolution grouping</figcaption></figure><p>All in all, we are done with discussing the feature learning part of PointNet++. The next and final step would be to choose the goal of the task — classification or segmentation. Figure 5 demonstrates the two different architectures for segmentation and classification, whereas Figure 5 shows the final overall structure of PointNet++. As a result of classification, PointNet++ returns the class scores that it outputs for the given 3D object. Alternatively, as a result of segmentation, PointNet++ uses skip links to reconstruct the original point set back with the scores attached for each of them in addition, because in segmentation we aim to identify the different parts within the object. Once we have done the classification or segmentation, we are done with PointNet++.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/841/1*uBsGdN1gXh9it0MVU9JfkQ.png" /><figcaption>Figure 5: Segmentation and Classification architectures of PointNet++</figcaption></figure><h4>Experiments &amp; Results</h4><p>The experiments for evaluating the PointNet++’s performance were done on 4 datasets of different nature (2D, 3D, and real 3D scenes). The categorization of the experiments is as follows:</p><blockquote><strong>a.</strong> Point set classification in Euclidean Metric Space — the aim is to classify point clouds sampled from both 2D (MNIST) and 3D (ModelNet40) Euclidean spaces. Figure 6.a shows the experiment results.</blockquote><blockquote><strong>b.</strong> Point set segmentation for semantic scene labeling — the aim is to predict semantic object labels for points in indoor scans. Figure 6.b. shows the experiment results.</blockquote><blockquote><strong>c.</strong> Point set classification in Non-Euclidean Metric Space — the aim is non-rigid shape is classification. As a result of this experiment, the generalizability of PointNet++ to non-euclidean space has been shown. Figure 6.c shows the comparative results between the approaches. Comparing the first and second settings of PointNet++ (<em>ours</em>) shows that the intrinsic features are very important for non-rigid shape classification. Also, comparing the second and third settings of PointNet++ (<em>ours</em>) shows that the use of geodesic neighborhoods yields better results than those of the euclidean neighborhoods.</blockquote><blockquote><strong>d.</strong> Feature visualization — Figure 6.d. visualizes what has been learned by the first level kernels of the hierarchical network.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MMhG3sOdzndq4I3PGKsqFA.png" /><figcaption>Figure 6: Experiment results.</figcaption></figure><h4>Related Works</h4><p>There have been several related works in this particular area. Eventually, the idea of hierarchical feature learning proved itself to be successful. Some related works include:</p><blockquote><strong>1.</strong> Convolutional Neural Networks — hierarchical structure but do not fit to unordered point sets with distance metrics.</blockquote><blockquote><strong>2.</strong> PointNet — fits to unordered point sets with distance metrics but ignores the underlying distance metric and does not support information about the local context.</blockquote><blockquote><strong>3.</strong> Geometry Processing community, Photogrammetry and Remote Sensing community –try to handle the selection of proper scale for point feature design but PointNet++ learns to extract point features and balances multiple feature scales in an end-to-end way.</blockquote><blockquote><strong>4.</strong> Volumetric grids and geometric graphs — popular deep learning representations but none of them consider the issue of non-uniform sampling density.</blockquote><h4>Conclusion</h4><p>To sum up, PointNet++ strikes as a robust neural network architecture for processing point sets sampled in a metric space. The functioning principle of PointNet++ is composed of recursively nested partitioning of the input point set, and effective learning of hierarchical features with respect to the distance metric. Furthermore, PointNet++ comes up with two novel set abstraction layers, which in turn allow handling the non-uniform point sampling issue. With all these contributions, PointNet++ achieves state-of-the-art performance on challenging benchmarks of 3D point clouds.</p><h3>Let’s Connect</h3><p>If you want to reach out to me or to share any feedback, feel free to drop by and say <em>hi </em>on LinkedIn — <a href="https://www.linkedin.com/in/aydin-javadov/">https://www.linkedin.com/in/aydin-javadov/</a>. Happy learning :)</p><p><strong>References</strong></p><p>[1] — Qi, C. R., Yi, L., Su, H., &amp; Guibas, L. J. (2017, June 7). <em>PointNet++: Deep hierarchical feature learning on point sets in a metric space</em></p><p>[2] — Qi, C. R., Su, H., Mo, K., &amp; Guibas, L. J. (2017, April 10). <em>PointNet: Deep Learning on point sets for 3D classification and segmentation</em></p><p>[3] — Fang, Y. I. (2021). 3D Computer Vision | Frontiers. Frontiers. <a href="https://www.frontiersin.org/research-topics/23013/3d-computer-vision#overview">https://www.frontiersin.org/research-topics/23013/3d-computer-vision#overview</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b9121e559549" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[But how do machines ‘learn’?]]></title>
            <link>https://medium.com/@ajavadov99/but-how-do-machines-learn-47341753fa12?source=rss-1375b2d0c883------2</link>
            <guid isPermaLink="false">https://medium.com/p/47341753fa12</guid>
            <category><![CDATA[supervised-learning]]></category>
            <category><![CDATA[binary-classification]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Aydin Javadov]]></dc:creator>
            <pubDate>Sun, 12 Jun 2022 19:40:35 GMT</pubDate>
            <atom:updated>2022-06-12T20:11:27.122Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N4Cw8Pydi3S7b_n7-AytLQ.jpeg" /></figure><p>The term “machine learning” is attaining more and more popularity, especially in the last couple of decades. It has become a mere routine to hear or read about the advancements in technology, such as state-of-the-art face recognition software, voice agents, intelligent robots, and so on. Hypothetically, there could be several reasons lying behind such hype. One obvious reason could be the fact that such advancements are meant to play facilitating roles in people’s daily lives. Thus, the excitement of people for such a helping hand should come as no surprise. Another potential reason for the prevalence of the “machine learning” term could be, indeed, its naming. The name somehow suggests that machines are learning the dedicated tasks, an intriguing yet counter-intuitive suggestion. Hence, this writing aims to discuss as well as clarify the process of learning for machines and attempts to answer the question “but how do machines learn?”. Depending on the nature of the data, whether labeled or not, and the nature of the task the learning algorithms can be roughly divided into three main categories: supervised learning, unsupervised learning, and reinforcement learning. This writing delves into the supervised use case, where the assumed data would be labeled.</p><p>There can be numerous tasks for the machine to learn, such as the prediction of medical issues, categorization of business processes, intelligent visual recognition, and so on. In general, the learning process for machines is divided into three essential parts: training, validation, and testing. Although all the tasks essentially boil down to the same steps, we will discuss the task of visual recognition, as it is arguably more intuitive to imagine.</p><p>To begin with, the training process is quite analogous to human learning processes where learning is achieved as a consequence of repetitive trials and failures. Consider a vision recognition task where the machine needs to be able to differentiate between dogs and cats. For specific tasks, there are several machine learning methods and approaches. Without providing further technical nuances, we say that artificial or convolutional neural networks might be suitable tools for our example case of dog-cat differentiation. For the training process to be robust, the amount of data that will be provided to the machine is vital. If there is a lack of data, the machine will be deprived of the opportunity to learn more. Thus, say, we have 20,000 images of dogs and cats combined. Furthermore, it is important to have a balanced data set, that is, fairly equal proportions of each class. Otherwise, say, if we provided the machine with 18,000 dog images and 2,000 cat images, then the machine would be biased and would think that there are many more dogs in the world than cats. Thus, assume that, ideally, we have 10,000 dog and 10,000 cat images. Once we are settled with our initial data content, we further split our data into roughly 80%, 10%, and 10% ratios for the actual training, validation, and test sets, respectively. We end up with a training set of 16,000 images, including 8,000 cats and 8,000 dogs.</p><figure><img alt="An example of Neural Networks" src="https://cdn-images-1.medium.com/max/1024/1*iFmq169lT1vXr0N1ZrGuVA.png" /><figcaption>An example of a Neural Network</figcaption></figure><p>Moving on, we are still yet to discuss how training takes place. For that, we need to promptly mention that neural networks are nothing but very complex functions, sometimes composed of millions or even billions of parameters! So, like every other function, neural networks also have the parameters to use. In fact, those unique combinations of parameters make the neural networks good or bad classifiers. Having that in mind, we are initially assigning random values to those parameters. From that, we obtain our first version of the neural network function. We shuffle the 16,000 images and ask our first version of the neural network to classify them. Since the parameters were purely randomly generated, the provided initial classifications are random as well. We call one complete pass through all the images in the training set an epoch. At the end of the first epoch, we compare the random predictions of the neural network with ground truth values. The difference between the predictions and the ground truth is called “loss”. Usually, the accuracy after the first epoch for the classification of two classes (dog and cat) is around 50%, if not less. Mathematically, the lower the accuracy, the greater the loss between the prediction and ground truth. What we aim to achieve is as little loss as possible. As we have said before, the training process is quite similar to human learning processes, in which learning occurs as a result of repeated trials and failures. Thus, we are sending our feedback to the neural network back and mentioning the correct as well as wrongly predicted cases. The network considers this feedback and makes appropriate tunings to its parameter values; it punishes (decreases) the ones who did a negative job and rewards (increases) the ones who did a positive job. Mathematically speaking, this process is called backpropagation, where the parameter values are updated by the loss function’s partial derivatives concerning those particular parameters.</p><p>After all the parameter updates are made, we end up with the second version of the neural network. Then, we again shuffle the 16,000 images and ask our second version of the neural network to classify them. Since the parameters were adjusted, we should hopefully see a decreased loss at the end of the second epoch. Our algorithm has already been run twice. We will repeat this process multiple times until we are convinced enough that the loss is sufficiently small and that our network is capable of distinguishing cats from dogs. One crucial detail is that if we repeat our algorithm more than enough, the network will start to overfit. In other words, the network would rather memorize, not learn, those 16,000 images. However, in the world, there are millions of other dogs and cats carrying other features, so we want to rather generalize our predictions. This is the point where we see the importance of the validation set.</p><p>Next, the validation set gets a 10% share of all the data. Hence, we have 1,000 images of dogs and cats, 500 of each. One reason we use the validation set is to test the generalizability of the network. The predictions of these new 1,000 unseen images provide useful information about how well the network is leveraged in general. To test the generalizability, we can plot the loss vs. epoch number for the training and validation sets. Until some point in the epoch number, both the training and validation loss will decrease together. However, after a certain point, although the training loss continues to decrease, the validation loss will either stop decreasing or start to increase. This is the point where generalization is maximized and the memorization of the training set starts. Hence, that value of the epoch is selected as the early-stopping point for the algorithm to train. We also use the validation set for hyperparameter tuning. Hyper-parameters are the implicit parameters such as the number of epochs, depth and structure of the network, number of parameters, learning rate, and so on. We can test the different combinations of those hyper-parameters to come up with an even enhanced version of the network.</p><p>Finally, once we are done with our training and validation, we should have a robust enough model (network). To test the final performance of the model, we are using the last portion of the data, which is the last unseen 10% share, composed of 1,000 images. It is strictly important to keep in mind that the test set is used only to evaluate the final performance of the model, and any further tweaks from this set are prohibited. The evaluation of the test set provides the final performance of the model, and it will hopefully provide fair prediction results. If not, then the process should be revised from scratch, starting from the training phase.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OClemWlNwXXCKKqtoJzU5A.jpeg" /></figure><h4>Why should we care about Machine Learning?</h4><p>The Artificial Intelligence sphere is subject to continuous advancements. Machine learning, a subset of artificial intelligence, is a notion where machines are ‘learning’ to perform dedicated tasks. In this ongoing flow of improvements, machine learning offers a helping hand for every age or job category, whether directly or indirectly. We observed a toy example to get a feeling of machine learning, whereas, of course, machine learning is not only about vision recognition, yet vision recognition can also be much more complicated than a dog-cat distinction task. Nevertheless, such a simple example can sufficiently encompass the basic steps of the learning processes of complicated designs and tasks for other use cases.</p><p>There are many application domains of machine learning useful for us, which makes the awareness of the subject even more worthwhile. For instance, machine learning-powered home robots can be one solution for improving the quality of elderly care. Medicine is also benefitting from machine learning, as now deeper networks can help health experts to diagnose diseases more accurately, minimize potential errors, predict the unseen side effects of treatments, or even come up with new drugs. Machine learning utilizations can lessen dangers for humans through bomb disposal or welding. The potential applications for daily life include personalized digital media, innovations in financial sectors, self-driving cars, environmental protection, and so on. All these machine learning applications relate to all of us somehow. Thus, it is essential not to envision machine learning as black-box magic. By gaining a more common understanding from society, perhaps, machine learning can be improved and welcomed more conveniently.</p><p>To conclude, throughout this writing we have concisely yet effectively highlighted how the process of learning is actually performed, where we have assumed a particular scenario of dog-cat image classification. We have discussed key terminologies, common practices, and important points that one needs to be aware of regarding machine learning, whereas we have also highlighted the reasons why would someone want to pay attention to machine learning.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=47341753fa12" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>