<?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 Selin Metin on Medium]]></title>
        <description><![CDATA[Stories by Selin Metin on Medium]]></description>
        <link>https://medium.com/@selinmetin?source=rss-3bf0699bda15------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*QfrWFp8YBPRK4kjRkGl90Q.jpeg</url>
            <title>Stories by Selin Metin on Medium</title>
            <link>https://medium.com/@selinmetin?source=rss-3bf0699bda15------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 07:54:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@selinmetin/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[Nonlinear Discriminant Functions]]></title>
            <link>https://medium.com/orion-innovation-techclub/nonlinear-discriminant-functions-ac432cb82e25?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac432cb82e25</guid>
            <category><![CDATA[techclub-tr]]></category>
            <category><![CDATA[machine]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[pattern-recognition]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Thu, 16 Mar 2023 14:58:45 GMT</pubDate>
            <atom:updated>2023-04-12T07:05:38.186Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://medium.com/@selinmetin/how-carl-sagans-apple-expanded-the-universe-2e0b981e7027"><strong>Classifiers</strong></a></p><p><a href="https://medium.com/@selinmetin/linear-discriminant-functions-5b0d91e8e577"><strong>Linear Discriminant Functions</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*-R-gBzeNqiP6a7BdhfI3cw.png" /></figure><p>The red dot easily separates the two classes so we have a one dimensional discriminant in a one dimensional input space. This is equivalent of a linear discriminant function. What if the features are distributed in this space as below? The members of the same class form distant clusters separated by another class. Even if we use two red dots, the boundaries they form will divide the space (the black line) in three regions and classify the two distant yellow circle clusters as different classes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*kunn5VK3zIrsPmzlLcxHCw.png" /></figure><p>Linear discriminators are not useful anymore so we need to think outside the box. The box in this context is our one dimensional input space which seems to confine our capabilities. Let’s see what happens when we introduce nonlinearity in this space:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*wTLARZfIXB1Gffh-C6YBAA.png" /></figure><p>The red circle which is defined with the generalized equation of (x_1-a)²+(x_2-b)²=r² can easily separate the two classes at the expense of increasing the dimension of the input space (now we have to use x1 and x2 to be able to define the discriminant function). Here is a broader vision of what we have done:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/604/1*bf5l4bHqlX_6KluYbZAl-A.png" /></figure><h3>Increase Dimensions of Space for Generalized Discriminant Functions</h3><p>The linear discriminant function g(x) can be written as</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/946/1*pHCQNuWX-chVitFmQx2QQQ.png" /></figure><p>where the coefficients wi are the components of the weight vector w. By adding additional terms involving the products of pairs of components of x, we obtain the quadratic discriminant function:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/946/1*4iceCBD671NdbfmkDQfpXQ.png" /></figure><p>By continuing to add terms, we can obtain the generalized linear discriminant function:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/946/1*vux_zLeUrjXNZTFsyWSM5A.png" /></figure><p>where a is now a d’-dimensional weight vector, and where the d’ functions yi(x) can be arbitrary functions of x. This generalized form is an expansion and the resulting discriminant function is not linear in x, but it is linear in y. The d’-functions yi(x) merely map points in d-dimensional x-space to points in d’-dimensional y-space [4].</p><p>While trying to find the most appropriate features for classification, we wanted to reduce the dimensionality of the input space to focus clearly on the more important data. However, now we are expanding the dimensions of input space to be able to use linear discriminant surfaces between clusters of classes. The linear decision boundary in the new space corresponds to a nonlinear decision boundary in the original space. The mapping from d-dimensional x-space to (d+1)-dimensional y-space is mathematically trivial but nonetheless quite convenient. By adding a constant component to vector representation of data in x, all distance relationships among samples are preserved. The resulting y vectors all lie in a d-dimensional subspace, which is the x-space itself. This mapping enables us to reduce the problem of finding a weight vector w and a threshold weight w0 to the problem of finding a single weight vector a.</p><p>Let’s try to see what we can gain by this expansion in the following example. Below MATLAB code snippet creates random samples of data:</p><pre>% generate 3D clusters of random data<br>clear all, close all, clc<br><br>nPerClust = 50;<br><br>% blur around centroid (std units)<br>n = 1;<br><br>% XY centroid locations<br>A = [1 2 0];<br>B = [-2 1 -2];<br>C = [3 3 2];<br><br>% generate data<br>a = bsxfun(@plus, randn(nPerClust, 3) * n, A);<br>b = bsxfun(@plus, randn(nPerClust, 3) * n, B);<br>c = bsxfun(@plus, randn(nPerClust, 3) * n, C);<br><br>% concatanate into a list<br>data = [a; b; c];<br><br>% show the data<br>figure (1)<br>clf<br>plot3(data(:,1), data(:,2), data(:,3), &#39;s&#39;, &#39;markerfaceColor&#39;, &#39;k&#39;);<br>axis square, grid on<br>rotate3d on</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/354/1*U9tdqmBBb4__A5hiP4HmTw.png" /></figure><p>The 3D plot shows how the data points are distributed. We can use k-means clustering to classify these data points in categories.</p><pre>% how many clusters<br>k = 2;<br><br>[groupidx, cents, sumdist, distances] = kmeans(data, k);<br><br>% draw lines from each data point to the centroids of each cluster<br><br>figure(2)<br>clf<br>% lineColors = &#39;rkbmrkbm&#39;;<br>hold on<br>axis square, grid on<br>rotate3d on<br>PlotClusters(data, groupidx, cents);<br></pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/395/1*tk6juUXHXfvo0eApf-0qkg.png" /><figcaption>k-means clustering of random data points</figcaption></figure><p>Built-in k-means function magically groups data but visual representation shows that groups are mixed with each other. Some samples from red cluster are in blue cluster region and some blue data points seem to be within red cluster boundaries. Is there a problem with our classification result such that samples near decision boundary (or discriminant surface) are not correctly calculated? 2D plot implies that something is wrong. Since the current figure has 3D rotation enabled, let’s start rotating the figure and see if we can better understand the classification results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/604/1*ErtiUDU4xd96s6U7IikQOQ.png" /><figcaption>k-means clustering of random data points in 3D</figcaption></figure><p>2D plot displayed the cluster members as inseparable — we couldn’t see a discriminant surface in any form, let alone a linear function. However, 3D plot shows that there is a clear linear discrimination surface. We just had to think outside box.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*ubICcMfVZPQzzfj57a6oNw.png" /><figcaption>Linear discriminant function in 3D plot</figcaption></figure><p>The dimensional expansion explained above can even be applied after feature extraction is complete and dimensionality reduction techniques are used to find the more representative features in the original input space. By changing the context (or our viewing angle of the problem), we can start to work with the subspace and develop our techniques for better results.</p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] <a href="https://www.imdb.com/title/tt0892700/">https://www.imdb.com/title/tt0892700/</a></p><p>[2] Many parts of this article refers to Prof. Z. Dokur Olmez’s “Pattern Recognition by Neural Networks” course notes from 20+ years ago</p><p>[3] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[4] Duda, R.O., Hart, P.E., Stork, D.G., Pattern Classification, 2nd Edition, Wiley-Interscience, November 9, 2000</p><p>[4] Drew Wilimitis, “The Kernel Trick in Support Vector Classification”, <a href="https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f">https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f</a></p><p>[5] <a href="https://www.mathsisfun.com/algebra/eigenvalue.html">https://www.mathsisfun.com/algebra/eigenvalue.html</a></p><p>[6] Test images are taken from University of Southern California Signal and Image Processing Institute USC-SIPI Image Database, volume 3 miscellaneous, <a href="https://sipi.usc.edu/database/database.php?volume=misc">https://sipi.usc.edu/database/database.php?volume=misc</a></p><p>[7] <a href="https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583">https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583</a></p><p>[8] <a href="https://www.section.io/engineering-education/image-compression-using-pca/">https://www.section.io/engineering-education/image-compression-using-pca/</a></p><p>[9] <a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p>[10] <a href="https://www.youtube.com/watch?v=UnURElCzGc0">https://www.youtube.com/watch?v=UnURElCzGc0</a></p><p>[11] The code snippets partially make use of the following resources:</p><p><a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle">https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors">https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac432cb82e25" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/nonlinear-discriminant-functions-ac432cb82e25">Nonlinear Discriminant Functions</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Linear Discriminant Functions]]></title>
            <link>https://medium.com/orion-innovation-techclub/linear-discriminant-functions-5b0d91e8e577?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b0d91e8e577</guid>
            <category><![CDATA[linear-discriminant]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[classification]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Thu, 16 Mar 2023 14:53:43 GMT</pubDate>
            <atom:updated>2023-04-12T07:06:33.185Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://medium.com/@selinmetin/how-carl-sagans-apple-expanded-the-universe-2e0b981e7027"><strong>Classifiers</strong></a></p><p>In the 2-dimensional input space below there are two classes which can be easily separated by a linear discriminant function:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/390/1*4Oe2cD6nNiN_BKQ22F2GgA.png" /></figure><p>The discriminant function is a line which can be represented by the below equation:</p><p><em>d</em>(<em>x</em>)=<em>w</em>1<em>x</em>1+<em>w</em>2<em>x</em>2+<em>w</em>3=0</p><p>Using this equation, any feature x belonging to class S1 results in a positive value while any feature x belonging to class S2 results in a negative value.</p><p>The equation that represents a linear discriminant function in an n-dimensional space is below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/526/1*ADDD7OIH60Uf3hCMmMM7rw.png" /></figure><p>In the below example, three classes can be separated by three linear discriminant functions:</p><p><em>d</em>1(<em>x</em>)= -<em>x</em>1+<em>x</em>2</p><p><em>d</em>2(<em>x</em>)=<em>x</em>1+<em>x</em>2–5</p><p><em>d</em>3(<em>x</em>)= -<em>x</em>2+1</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/998/1*mHiAk7LxoO4Fh7VIkZInNw.png" /></figure><p><a href="https://medium.com/@selinmetin/nonlinear-discriminant-functions-ac432cb82e25"><strong>Nonlinear Discriminant Functions</strong></a></p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] <a href="https://www.imdb.com/title/tt0892700/">https://www.imdb.com/title/tt0892700/</a></p><p>[2] Many parts of this article refers to Prof. Z. Dokur Olmez’s “Pattern Recognition by Neural Networks” course notes from 20+ years ago</p><p>[3] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[4] Duda, R.O., Hart, P.E., Stork, D.G., Pattern Classification, 2nd Edition, Wiley-Interscience, November 9, 2000</p><p>[4] Drew Wilimitis, “The Kernel Trick in Support Vector Classification”, <a href="https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f">https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f</a></p><p>[5] <a href="https://www.mathsisfun.com/algebra/eigenvalue.html">https://www.mathsisfun.com/algebra/eigenvalue.html</a></p><p>[6] Test images are taken from University of Southern California Signal and Image Processing Institute USC-SIPI Image Database, volume 3 miscellaneous, <a href="https://sipi.usc.edu/database/database.php?volume=misc">https://sipi.usc.edu/database/database.php?volume=misc</a></p><p>[7] <a href="https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583">https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583</a></p><p>[8] <a href="https://www.section.io/engineering-education/image-compression-using-pca/">https://www.section.io/engineering-education/image-compression-using-pca/</a></p><p>[9] <a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p>[10] <a href="https://www.youtube.com/watch?v=UnURElCzGc0">https://www.youtube.com/watch?v=UnURElCzGc0</a></p><p>[11] The code snippets partially make use of the following resources:</p><p><a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle">https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors">https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b0d91e8e577" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/linear-discriminant-functions-5b0d91e8e577">Linear Discriminant Functions</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Carl Sagan’s Apple Expanded the Universe]]></title>
            <link>https://medium.com/orion-innovation-techclub/how-carl-sagans-apple-expanded-the-universe-2e0b981e7027?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/2e0b981e7027</guid>
            <category><![CDATA[techclub-tr]]></category>
            <category><![CDATA[pattern-recognition]]></category>
            <category><![CDATA[classifier]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Thu, 16 Mar 2023 14:50:38 GMT</pubDate>
            <atom:updated>2023-04-12T07:07:25.674Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://medium.com/@selinmetin/on-beloved-tv-shows-and-why-dimensionality-of-data-matters-caf85b7314e7"><strong>On Beloved TV Shows and Why Dimensionality of Data Matters</strong></a></p><p>I remember watching the original Cosmos series hosted by Carl Sagan on TV. I was only in elementary school and was not able to realize all the concepts but one of the key moments that remained with me until this day is the depiction of the 4th dimension. Although I can find the original scene online now [10], I can clearly remember how he explained the extra dimensions with the help of an apple. Many years later after reading Flatland by Edwin A. Abbott, I remembered that bright red apple again and marveled at how adding another dimension made things easier to understand.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/604/1*bk4x96uIIe28e67f58ZYKQ.png" /><figcaption>Carl Sagan explaining the fourth dimension in original Cosmos series [10]</figcaption></figure><p>I think this is a perfect example of viewing things from different angles and how the perspective could make a difference. Although modern trends usually dictate us to downsize and get rid of extras, in some cases upscaling may interestingly help us to find better explanations.</p><p>Since I focused on dimensionality reduction in machine learning so far, it may feel contradictory but we can see the benefits of an increase in dimensions in pattern recognition.</p><h3>Classifiers</h3><p>Classifiers form a decision surface in feature space to depict the distribution of classes. In communication theory, classification problem is handled as discriminating different signals sent from the transmitter at the receiver. Modulation techniques used at the receiver artificially reconstruct transmitted signals with enough distinctive properties and can get rid of the noise in the transmission medium. On the contrary, machine learning practices work with natural signals instead of artificially reconstructed ones. Usually a transform operation is applied to uncover discriminative properties of the signals. The goal of transform operation is to minimize the scattering within the vectors of a single class while maximizing the distance between the centers of different clusters. Such transforms provide linear discrimination surfaces (like straight lines or planes). Otherwise cluster boundaries cross over each other and vectors of the same class form clusters in different parts of the input space. Such distributions can be represented by nonlinear discrimination surfaces (like hyperbola).</p><p><a href="https://medium.com/@selinmetin/linear-discriminant-functions-5b0d91e8e577"><strong>Linear Discriminant Functions</strong></a></p><p><a href="https://medium.com/@selinmetin/nonlinear-discriminant-functions-ac432cb82e25"><strong>Nonlinear Discriminant Functions</strong></a></p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] <a href="https://www.imdb.com/title/tt0892700/">https://www.imdb.com/title/tt0892700/</a></p><p>[2] Many parts of this article refers to Prof. Z. Dokur Olmez’s “Pattern Recognition by Neural Networks” course notes from 20+ years ago</p><p>[3] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[4] Duda, R.O., Hart, P.E., Stork, D.G., Pattern Classification, 2nd Edition, Wiley-Interscience, November 9, 2000</p><p>[4] Drew Wilimitis, “The Kernel Trick in Support Vector Classification”, <a href="https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f">https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f</a></p><p>[5] <a href="https://www.mathsisfun.com/algebra/eigenvalue.html">https://www.mathsisfun.com/algebra/eigenvalue.html</a></p><p>[6] Test images are taken from University of Southern California Signal and Image Processing Institute USC-SIPI Image Database, volume 3 miscellaneous, <a href="https://sipi.usc.edu/database/database.php?volume=misc">https://sipi.usc.edu/database/database.php?volume=misc</a></p><p>[7] <a href="https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583">https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583</a></p><p>[8] <a href="https://www.section.io/engineering-education/image-compression-using-pca/">https://www.section.io/engineering-education/image-compression-using-pca/</a></p><p>[9] <a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p>[10] <a href="https://www.youtube.com/watch?v=UnURElCzGc0">https://www.youtube.com/watch?v=UnURElCzGc0</a></p><p>[11] The code snippets partially make use of the following resources:</p><p><a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle">https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors">https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2e0b981e7027" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/how-carl-sagans-apple-expanded-the-universe-2e0b981e7027">How Carl Sagan’s Apple Expanded the Universe</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Principal Components Analysis — Karhunen-Loéve Expansion]]></title>
            <link>https://medium.com/orion-innovation-techclub/principal-components-analysis-karhunen-lo%C3%A9ve-expansion-7dcd2b27bb53?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/7dcd2b27bb53</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[dimensionality-reduction]]></category>
            <category><![CDATA[principal-component]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Thu, 16 Mar 2023 14:48:37 GMT</pubDate>
            <atom:updated>2023-04-12T07:09:44.599Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Principal Components Analysis — Karhunen-Loéve Expansion</strong></h3><p><a href="https://medium.com/@selinmetin/king-juliens-wisdom-dimensions-and-classification-42ba0cfe9ae1"><strong>Dimension Reduction in Pattern Recognition</strong></a></p><p>PCA-KLE is used to obtain features of a distribution that best represents the sample signals. It identifies patterns in data, expressing this data in a way as to highlight their similarities and differences. PCA-KLE reduces the number of data dimensions by “compressing” the data (this makes it easier to visualize the hidden data pattern). To do the compression, the eigenvectors and eigenvalues are computed and sorted according to decreasing eigenvalue.</p><p>We have some new terms now. Eigen is a word from German origin and in linear algebra eigenvalues and eigenvectors refer to characteristics of a square matrix. For a good example with proofs, you may refer to [5].</p><p>The goal of PCA-KLE is to find the optimum number of characteristics that best describe the data pattern so that instead of using the whole data, this subset can be used without losing much from the information content of the original data.</p><p>Let’s try to visualize this on a 2D space. There are 1000 random points in the image below. On the Cartesian coordinate system, we need 2 values to store each of these data points, i.e. we need both x and y coordinates. The dimension of this space is 2 and we can’t apply any compression because both dimensions are required. However, we can see that all the data points lie around a linear path, as depicted by the blue line. If we rotate the coordinate system such that x-axis lies along the blue line, y-axis becomes the green line which is perpendicular to the blue line. If we use this new axis system to represent our data points, we can see that most of the variance lies along the new x-axis, i.e. the blue line. Therefore, we can use only new x-axis values to represent each data point. This reduces the dimension of our dataset to one and requires half as much the initial storage space to store and reconstruct the dataset. In other words, there is only one principal component now that is sufficient to reconstruct the original data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/604/1*sNmejK8wCJg7w9S25r2xFw.png" /></figure><p>Let’s see PCA results on image processing. Below piece of code written in MATLAB applies PCA to compress a test image. Colored input image is first converted to grayscale for easier operations.</p><pre>% Read graphics file into an array<br>clear all, close all, clc<br>% Get the initial image<br>filename = &#39;4.1.01.tiff&#39;;<br>img=imread(filename);<br>imgs = rgb2gray(img);<br>A = imgs;<br>A=double(A); % Create the douple<br>figure(1);<br>subplot(2,2,1);<br>imshow(img);<br>title(&#39;Original image&#39;);<br>subplot(2,2,2);<br>imshow(imgs);<br>title(&#39;Input image&#39;);<br>[m n]=size(A);<br>mn = mean(A,2);<br>X=A-repmat(mn,1,n);<br>Z=1/sqrt(n-1)*X&#39;;<br>covZ=Z&#39;*Z; % find covariance matrix<br>[eigvec, eigval]=eig(covZ); % find eigen values &amp; vectors<br>[U,S,V]=svd(covZ);<br>variances=diag(S).*diag(S); % find variances<br>PCs = 50; % Select first 50 Principle components<br>subplot(2,2,3); bar(variances(1:PCs)); % plot first n variances<br>title([&#39;First &#39;, num2str(PCs), &#39; variances&#39;]);<br>P = V(:,1:PCs);<br>Y= P&#39;*X;<br>ratio=256/(2*PCs+1);<br>XX = P*Y;<br>XX=XX+repmat(mn,1,n); % compress the image<br>subplot(2,2,4); imshow(XX,gray), axis off<br>title(&#39;Compressed image&#39;);<br>imwrite(imgs,&#39;OriginalBWImage.jpg&#39;);<br>fileinfo = dir(&#39;OriginalBWImage.jpg&#39;);<br>SIZE = fileinfo.bytes;<br>OrigSize = SIZE/1024;<br>imwrite(XX,&#39;PCACompImage.jpg&#39;);<br>fileinfo = dir(&#39;PCACompImage.jpg&#39;);<br>SIZE = fileinfo.bytes;<br>CompSize = SIZE/1024;<br>fprintf(&#39;\n The size of input black&amp;white image is %0.4f KB&#39;, OrigSize);<br>fprintf(&#39;\n The size of PCA compressed image is %0.4f KB\n&#39;, CompSize);</pre><p>The output of this code is given in the below image file:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*L7rVeHd9TU-BAiwT5Btj8Q.png" /></figure><pre>The size of input black&amp;white image is 8.2178 KB<br>The size of PCA compressed image is 2.3379 KB</pre><p>Above example uses the first 50 principal components to compress the image. The variances of these components is shown in the graph. The relevance of the variances decreases exponentially and is not even visible after the first 10. Let’s see what happens if we use less principle components, e.g. first 10:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/604/1*y08vjQ8Pq2DFQ9qWGz4Zjw.png" /></figure><pre>The size of input black&amp;white image is 8.2178 KB<br>The size of PCA compressed image is 2.3379 KB</pre><p>Although the bar graph still shows that first few variance values are greater and should carry most of the information, the compressed image file has considerable amount of quality loss. Actually we are fooled by the scale of the variance graph because the first few variances have larger values which change the graph scale automatically and since we look at the whole picture instead of focusing on areas of the graph, we can’t correctly estimate the number of most useful principal components.</p><p>Let’s continue experimenting and look at what happens when we use the first 100 principal components:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*Ul61vLQxSHMozm62QkTj2Q.png" /></figure><pre>The size of input black&amp;white image is 8.2178 KB<br>The size of PCA compressed image is 1.5576 KB</pre><p>The quality of the compressed image is almost the same as the original one and we can’t notice a remarkable difference between the two compressed images which use 50 and 100 principal components.</p><p>There is one interesting output in the above examples. The size of the compressed image file is not relatively small in most of the examples. In fact, it can become even larger. The compression ratio for one image with PCA is a negative number because you need more information to reconstruct it with PCA than the original image. The reason for this is PCA works better if you want to compress a set of images rather than a single image file. PCA is a good technique to reduce dimensionality of datasets when the elements of that dataset have something in common such as a set of human face photographs. Since all humans have similar facial features like eyes, nose, or mouth, a template can be used to express each one of these features and then these templates can be combined to generate original faces in the dataset. The templates still have the original size of the image dimensions but since the required number of templates will be small in comparison to the whole size of the dataset, the storage size required for the PCA compressed version of the dataset will be much smaller than the original dataset [7]. The image compression use case example in [8] utilizes the first 100 principle components instead of the original 4096 dimensions for each image file.</p><p>The final compressed file size for the single image case is also dependent on the type of the file selected for storage. JPG, PNG, BMP, TIFF formats have different built-in compression algorithms per system. Below code snippet is a simpler implementation that directly uses PCA function from the MATLAB statistics toolbox:</p><pre>I = imread(&#39;4.1.01.tiff&#39;);<br>X = rgb2gray(I);<br>figure, imshow(X);<br>[coeff,score,~] = pca(double(X),&#39;Centered&#39;,false);<br>% reconstruct the image using the first n Principal Components<br>nPC = 10;<br>imgRC=uint8(score(:,1:nPC)*coeff(:,1:nPC)&#39;);<br>figure, imshow(imgRC);<br>imwrite(X,&#39;OriginalBWImage.jpg&#39;);<br>fileinfo = dir(&#39;OriginalBWImage.jpg&#39;);<br>SIZE = fileinfo.bytes;<br>OrigSize = SIZE/1024;<br>imwrite(I,&#39;PCACompImage2ndMethod.jpg&#39;);<br>fileinfo = dir(&#39;PCACompImage2ndMethod.jpg&#39;);<br>SIZE = fileinfo.bytes;<br>CompSize = SIZE/1024;<br>fprintf(&#39;\n File sizes when storage format is JPG&#39;);<br>fprintf(&#39;\n The size of input black&amp;white image is %0.4f KB&#39;, OrigSize);<br>fprintf(&#39;\n The size of PCA compressed image is %0.4f KB\n&#39;, CompSize);<br>% Storage file format changed to PNG<br>imwrite(X,&#39;OriginalBWImagePNG.png&#39;);<br>fileinfo = dir(&#39;OriginalBWImagePNG.png&#39;);<br>SIZE = fileinfo.bytes;<br>OrigSize = SIZE/1024;<br>imwrite(I,&#39;PCACompImage2ndMethodPNG.png&#39;);<br>fileinfo = dir(&#39;PCACompImage2ndMethodPNG.png&#39;);<br>SIZE = fileinfo.bytes;<br>CompSize = SIZE/1024;<br>fprintf(&#39;\n File sizes when storage format is PNG&#39;);<br>fprintf(&#39;\n The size of input black&amp;white image is %0.4f KB&#39;, OrigSize);<br>fprintf(&#39;\n The size of PCA compressed image is %0.4f KB\n&#39;, CompSize);</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/542/1*-gbN-_3AMJkW0gbTrAYwew.png" /></figure><pre>File sizes when storage format is JPG<br>The size of input black&amp;white image is 8.2178 KB<br>The size of PCA compressed image is 9.7139 KB<br>File sizes when storage format is PNG<br>The size of input black&amp;white image is 36.1699 KB<br>The size of PCA compressed image is 116.0273 KB</pre><p>When the first 10 principal components are used for reconstruction, the output is visually as good as the longer code that calculates variances. However, compressed file size is much greater. You can also notice that when the storage format changes, the same algorithm uses quite different amount of disk size to store the image array contents. (Above examples partially use the code and explanations from [9].)</p><p>When the first 50 principal components are used with the same example, output image has quite high visual quality as seen below. Due to the above mentioned storage file format differences, the input and compressed file sizes are equal for different number of principle components used.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/541/1*m43GoSvoZPBz2Bt5-NJtIQ.png" /></figure><pre>File sizes when storage format is JPG<br>The size of input black&amp;white image is 8.2178 KB<br>The size of PCA compressed image is 9.7139 KB<br>File sizes when storage format is PNG<br>The size of input black&amp;white image is 36.1699 KB<br>The size of PCA compressed image is 116.0273 KB</pre><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] <a href="https://www.imdb.com/title/tt0892700/">https://www.imdb.com/title/tt0892700/</a></p><p>[2] Many parts of this article refers to Prof. Z. Dokur Olmez’s “Pattern Recognition by Neural Networks” course notes from 20+ years ago</p><p>[3] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[4] Duda, R.O., Hart, P.E., Stork, D.G., Pattern Classification, 2nd Edition, Wiley-Interscience, November 9, 2000</p><p>[4] Drew Wilimitis, “The Kernel Trick in Support Vector Classification”, <a href="https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f">https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f</a></p><p>[5] <a href="https://www.mathsisfun.com/algebra/eigenvalue.html">https://www.mathsisfun.com/algebra/eigenvalue.html</a></p><p>[6] Test images are taken from University of Southern California Signal and Image Processing Institute USC-SIPI Image Database, volume 3 miscellaneous, <a href="https://sipi.usc.edu/database/database.php?volume=misc">https://sipi.usc.edu/database/database.php?volume=misc</a></p><p>[7] <a href="https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583">https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583</a></p><p>[8] <a href="https://www.section.io/engineering-education/image-compression-using-pca/">https://www.section.io/engineering-education/image-compression-using-pca/</a></p><p>[9] <a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p>[10] <a href="https://www.youtube.com/watch?v=UnURElCzGc0">https://www.youtube.com/watch?v=UnURElCzGc0</a></p><p>[11] The code snippets partially make use of the following resources:</p><p><a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle">https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors">https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7dcd2b27bb53" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/principal-components-analysis-karhunen-lo%C3%A9ve-expansion-7dcd2b27bb53">Principal Components Analysis — Karhunen-Loéve Expansion</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[King Julien’s Wisdom, Dimensions and Classification]]></title>
            <link>https://medium.com/orion-innovation-techclub/king-juliens-wisdom-dimensions-and-classification-42ba0cfe9ae1?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/42ba0cfe9ae1</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <category><![CDATA[classifier]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[dimensionality-reduction]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Thu, 16 Mar 2023 14:40:30 GMT</pubDate>
            <atom:updated>2023-03-31T23:19:21.024Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://medium.com/@selinmetin/on-beloved-tv-shows-and-why-dimensionality-of-data-matters-caf85b7314e7"><strong>On Beloved TV Shows and Why Dimensionality of Data Matters</strong></a></p><p>I love Penguins of Madagascar, especially the TV series. Although the show was intended for kids, it portrays colorful personalities with prominent characteristics and contains many anecdotal moments for the grown up audience as well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/491/1*btX73zxiKlUfIF5o_y2ZmA.png" /><figcaption>Snapshot taken from [1]</figcaption></figure><p>In one of the episodes King Julien solves the eternal question about priority: Which comes first, the chicken or the egg? The conversation between him and his minion Maurice flows somewhat like below:</p><p>- This is a riddle, which is eaten first? Egg or chicken? I think egg because egg is served on breakfast.</p><p>- But it is not like that…</p><p>- So you think chicken is eaten first? A-ha! Well, in this case it will be easier to take the egg.</p><p>Although it sounds like he misunderstood the question, King Julien cleverly confines the universal set of probabilities to a dining menu and answers within this context. He is also open minded enough to accept the opposite answer due to practical reasons about obtaining the egg. He simply decreases the dimensions of the input space to be able to focus on an important feature (in this case the course of eating).</p><p>There is a valuable lesson here: we need to define our area of focus depending on our goals to get finer results. From a mathematical point of view, we need to apply dimension reduction to concentrate our points of interest.</p><h3>Dimension Reduction in Pattern Recognition</h3><p>Pattern recognition is one of the most popular research areas of data analysis methods in machine learning. Since the aim of this article is not to delve on the problem of feature extraction for classification tasks, we will fast forward to the topic of feature space dimensions by summarizing the features as the important properties of the input space that yield maximum information in order to perform classification. The number of features defines the dimension of feature space and is dependent on the nature of the input data as well as the goal of classification (i.e. you can define different features from the same input data if you have different classification tasks like categorizing a class of students by their interests depending on the courses they take or by their heights depending on their physical appearances).</p><p>Based on the classification task and the classifier algorithm, some of the extracted features will have better performance in defining class boundaries or discriminating between the categories in the feature space. Since the feature space dimension is directly related to the computational complexity of the classifier algorithm, reducing the number of features by selecting the most important ones will simplify the overall computational cost. “Most important” is a subjective quality and needs to be defined via experimentation on the dataset — this is maybe the first rule of machine learning: know your data!</p><p>Dimensionality reduction is a well-known and extensively studied topic in statistics and communication theory. The goal of dimensionality reduction is to transform the feature space to a subspace with less dimensions. This will simplify the performance criteria of the classification task and reduce the number of calculations.</p><h3>Information and Probability</h3><p>Entropy is a measure of the disorder in a system. Shannon’s Information Theorem tells us that the less likely an event or symbol si is, the more uncertainty exists, and the more information one gets if this event/symbol occurs. In other words, an event delivers more information if we are surprised by it. If the event is expected, i.e. the frequency of occurrence is high, it does not yield much information. Consider this example: if the dog bites the man, this is expected. If the man bites the dog, this is news worth telling. The entropy η of an information source with alphabet X = {x1, x2, . . . , xn} is given by the following equation in the discrete space[3]:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/946/1*lCZU5LaSJYnBthHu2rnECA.png" /></figure><p>or in the continuous space denoted as [4]:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/946/1*6oif-g8cmqJzAAy-qrXPDw.png" /></figure><p>In most cases there is a good chance that a substantial amount of correlation is inherent among neighboring samples of feature data. The goal of dimension reduction is to efficiently squeeze most of the information of feature space to a subspace which is simpler to operate on. To do this compression, a transform operation is applied on the feature space. Feature space and transform operations are denoted by vectors (or matrices) and linear algebra principles apply. If most information is accurately described by the first few components of a transformed vector, then the remaining components can be set to zero, with little information loss. In technical terms, if Y is the result of a linear transform T of the feature vector X in such a way that the components of Y are much less correlated, then Y can be coded more efficiently than X.</p><p>Features which decrease the amount of disorder contain more information. Thus minimizing the entropy of a system also minimizes the dispersion of classes distributed in the space. Therefore, the applied transform operation should also compact the classes in the space in addition to decreasing the dimensions of the space. The entropy criterion tries to minimize the entropy of the subspace in order to find the optimal transform operation.</p><p>Various methods have been proposed for dimensionality reduction. The classical procedures used in statistics are principal components analysis (PCA) and factor analysis, both of which reduce dimensionality by forming linear combinations of the features. The object of principal components analysis (known in communication theory as the Karhunen-Loéve expansion or PCA-KLE) is to find a lower-dimensional representation that accounts for the variance of the features. The object of factor analysis is to find a lower-dimensional representation that accounts for the correlations among the features.</p><p><a href="https://medium.com/@selinmetin/principal-components-analysis-karhunen-lo%C3%A9ve-expansion-7dcd2b27bb53"><strong>Principal Components Analysis — Karhunen-Loéve Expansion</strong></a></p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] <a href="https://www.imdb.com/title/tt0892700/">https://www.imdb.com/title/tt0892700/</a></p><p>[2] Many parts of this article refers to Prof. Z. Dokur Olmez’s “Pattern Recognition by Neural Networks” course notes from 20+ years ago</p><p>[3] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[4] Duda, R.O., Hart, P.E., Stork, D.G., Pattern Classification, 2nd Edition, Wiley-Interscience, November 9, 2000</p><p>[4] Drew Wilimitis, “The Kernel Trick in Support Vector Classification”, <a href="https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f">https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f</a></p><p>[5] <a href="https://www.mathsisfun.com/algebra/eigenvalue.html">https://www.mathsisfun.com/algebra/eigenvalue.html</a></p><p>[6] Test images are taken from University of Southern California Signal and Image Processing Institute USC-SIPI Image Database, volume 3 miscellaneous, <a href="https://sipi.usc.edu/database/database.php?volume=misc">https://sipi.usc.edu/database/database.php?volume=misc</a></p><p>[7] <a href="https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583">https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583</a></p><p>[8] <a href="https://www.section.io/engineering-education/image-compression-using-pca/">https://www.section.io/engineering-education/image-compression-using-pca/</a></p><p>[9] <a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p>[10] <a href="https://www.youtube.com/watch?v=UnURElCzGc0">https://www.youtube.com/watch?v=UnURElCzGc0</a></p><p>[11] The code snippets partially make use of the following resources:</p><p><a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle">https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors">https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=42ba0cfe9ae1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/king-juliens-wisdom-dimensions-and-classification-42ba0cfe9ae1">King Julien’s Wisdom, Dimensions and Classification</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[On Beloved TV Shows and Why Dimensionality of Data Matters]]></title>
            <link>https://medium.com/orion-innovation-techclub/on-beloved-tv-shows-and-why-dimensionality-of-data-matters-caf85b7314e7?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/caf85b7314e7</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[dimensionality-reduction]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <category><![CDATA[classifier]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Thu, 16 Mar 2023 14:36:41 GMT</pubDate>
            <atom:updated>2023-03-31T23:19:15.544Z</atom:updated>
            <content:encoded><![CDATA[<p>Thinking outside the box helps to view problems from different perspectives and is useful in finding creative solutions. Following puzzle is a common example used to emphasize this concept: can you connect the 9 dots below with only 4 lines?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/209/1*lxg6_MDYJgwtGLXD3Dm-vg.png" /><figcaption>Can you connect the 9 dots with 4 lines?</figcaption></figure><p>These dots form a square for us, or maybe 4 little squares and a big square containing the others. We contemplate the directions of the possible lines but most of us fail to solve the problem initially. We are told that there is a simple solution but it starts to feel frustrating as we struggle to find the answer. Here is the solution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/310/1*wrm_6cMVGWjkYSyiAp-X9Q.png" /><figcaption>Solution</figcaption></figure><p>Our perspective is usually confined with what we see initially — the dots seem to form a square with invisible lines so our perception becomes dependent on this invisible square. Our minds are conditioned to see the invisible lines: all the childhood puzzles we solved had these invisible lines and when we drew them with pencil, we could see the hidden figures. Therefore, it is difficult to extend the line beyond the outer dots or to imagine a triangle instead of a square. Neuroscience and psychology can explain this interesting phenomenon better than me. I prefer to work in the perfect world of technology in which mathematics undoubtedly explains concepts. Hence I want to discuss how thinking outside the box helps to simplify problems related with the dimensions of the input space in machine learning and classification tasks.</p><p><a href="https://medium.com/@selinmetin/king-juliens-wisdom-dimensions-and-classification-42ba0cfe9ae1">King Julien’s Wisdom, Dimensions and Classification</a></p><p><a href="https://medium.com/@selinmetin/how-carl-sagans-apple-expanded-the-universe-2e0b981e7027">How Carl Sagan’s Apple Expanded the Universe</a></p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] <a href="https://www.imdb.com/title/tt0892700/">https://www.imdb.com/title/tt0892700/</a></p><p>[2] Many parts of this article refers to Prof. Z. Dokur Olmez’s “Pattern Recognition by Neural Networks” course notes from 20+ years ago</p><p>[3] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[4] Duda, R.O., Hart, P.E., Stork, D.G., Pattern Classification, 2nd Edition, Wiley-Interscience, November 9, 2000</p><p>[4] Drew Wilimitis, “The Kernel Trick in Support Vector Classification”, <a href="https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f">https://towardsdatascience.com/the-kernel-trick-c98cdbcaeb3f</a></p><p>[5] <a href="https://www.mathsisfun.com/algebra/eigenvalue.html">https://www.mathsisfun.com/algebra/eigenvalue.html</a></p><p>[6] Test images are taken from University of Southern California Signal and Image Processing Institute USC-SIPI Image Database, volume 3 miscellaneous, <a href="https://sipi.usc.edu/database/database.php?volume=misc">https://sipi.usc.edu/database/database.php?volume=misc</a></p><p>[7] <a href="https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583">https://towardsdatascience.com/face-dataset-compression-using-pca-cddf13c63583</a></p><p>[8] <a href="https://www.section.io/engineering-education/image-compression-using-pca/">https://www.section.io/engineering-education/image-compression-using-pca/</a></p><p>[9] <a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p>[10] <a href="https://www.youtube.com/watch?v=UnURElCzGc0">https://www.youtube.com/watch?v=UnURElCzGc0</a></p><p>[11] The code snippets partially make use of the following resources:</p><p><a href="https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress">https://www.mathworks.com/matlabcentral/answers/1761175-with-pca-how-much-of-the-photo-did-i-compress</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle">https://www.mathworks.com/matlabcentral/fileexchange/88748-lrotate?s_tid=srchtitle</a></p><p><a href="https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors">https://www.mathworks.com/matlabcentral/fileexchange/31710-plotclusters-data-idx-centers-colors</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=caf85b7314e7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/on-beloved-tv-shows-and-why-dimensionality-of-data-matters-caf85b7314e7">On Beloved TV Shows and Why Dimensionality of Data Matters</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deep Learning for Audio Signal Processing]]></title>
            <link>https://medium.com/orion-innovation-techclub/deep-learning-for-audio-signal-processing-4fe97665e0aa?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/4fe97665e0aa</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[audio-compression]]></category>
            <category><![CDATA[neural-codecs]]></category>
            <category><![CDATA[multimedia-compression]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Wed, 10 Aug 2022 09:36:43 GMT</pubDate>
            <atom:updated>2022-09-09T06:32:21.438Z</atom:updated>
            <content:encoded><![CDATA[<p>This is part of a series of articles in <a href="https://medium.com/@selinmetin/ai-codecs-a128d2896362">AI Codecs</a>.</p><p>While many deep learning methods have been adopted from image processing, there are important differences between audio and image/video compression.</p><p>· Raw audio samples form a one-dimensional time series signal, which is fundamentally different from two-dimensional images.</p><p>· Audio signals are commonly transformed into two-dimensional time frequency representations for processing, but the two axes, time and frequency, are not homogeneous as horizontal and vertical axes in an image.</p><p>· Images are instantaneous snapshots of a target and often analyzed as a whole or in patches with little order constraints; however audio signals have to be studied sequentially in chronological order.</p><p>Deep learning models have a large modeling capacity so end-to-end trained systems that directly map the input audio signal to the target audio sequences can be developed. These systems are trained to optimize criteria that are related to the final evaluation metric (such as word error rate) [16]. Despite the success of GANs for image synthesis, their use in the audio domain has been limited. GANs have been used for source separation, music instrument transformation and speech enhancement to transform noisy speech input to denoised versions.</p><h3>Applications of Deep Learning in Audio</h3><p>Purwins and his colleagues present a survey of recent deep learning techniques used for audio signal processing [16]. They present these methods under several groups (convolutional NNs, recurrent NNs, sequence-to-sequence models, GANs) some of which are listed below.</p><p><strong>Speech recognition by neural network based hybrid models</strong>: After 2012, we start to see deep neural network based models with millions of parameters trained on thousands of hours of data.</p><p>· Virtual assistants, such as Google Home, Amazon Alexa and Microsoft Cortana, all adopt voice as the main interaction modality. Speech transcriptions also find their way to various applications for retrieving information from multimedia, such as YouTube speech captioning.</p><p>· With increasing adoption of speech based applications, extending speech support for more speakers and languages has become more important.</p><p>· Transfer learning (using large amounts of similar data labeled for another task and adapting the knowledge learned from it to the target domain) has been used to boost the performance of Automatic Speech Recognition (ASR) systems on low resource languages with data from rich resource languages. With the success of deep learning models in ASR, other speech related tasks also embraces deep learning techniques, such as voice activity detection, speaker recognition, language recognition and speech translation. (English language is maybe the most studied language in speech recognition and therefore has the largest resources. For example, Youtube automatically added subtitles to videos in English for quite a long time but since last year, videos in many other languages can have automatic subtitles only hours after they are uploaded to the cloud. This was made possible partially due to transfer learning approach.)</p><p><strong>Music:</strong> Compared to speech, music recordings typically contain a wider variety of sound sources of interest. In many kinds of music, their occurrence follows common constraints in terms of time and frequency, creating complex dependencies within and between sources.</p><p>Tasks in music processing contain low-level analysis (onset and offset detection, fundamental frequency estimation), rhythm analysis (beat tracking, meter identification, downbeat tracking, tempo estimation), harmonic analysis (key detection, melody extraction, chord estimation), high-level analysis (instrument detection, instrument separation, transcription, structural segmentation, artist recognition, genre classification, mood classification) and high-level comparison (discovery of repeated themes, cover song identification, music similarity estimation, score alignment). Each of these has originally been approached with hand-designed algorithms or features combined with shallow classifiers, but are now tackled with deep learning.</p><p>Deep learning has been applied successfully to numerous music processing tasks, and drives industrial applications with automatic descriptions for browsing large catalogues, with content-based music recommendations in the absence of usage data, and also profanely with automatically derived chords for a song to play along with.</p><p><strong>Environmental Sounds:</strong> Signals and other sounds also carry a wealth of relevant information about our environments. Computational analysis of environmental sounds has several applications, for example in context-aware devices, acoustic surveillance, or multimedia indexing and retrieval. It is typically done with three basic approaches: a) acoustic scene classification, b) acoustic event detection, and c) tagging.</p><p>· Acoustic scene classification aims to label a whole audio recording with a single scene label. Possible scene labels include for example “home”, “street”, “in car”, “restaurant”, etc. The set of scene labels is defined in advance.</p><p>· Acoustic event detection aims to estimate the start and end times of individual sound events such as footsteps, traffic light acoustic signaling, dogs barking, and assign them an event label. The set of possible event classes should be defined in advance.</p><p><strong>Localization and Tracking:</strong> Multichannel audio allows for the localization and tracking of sound sources, i.e. determining their spatial locations, and tracking them over time and can, for example, be used as a part of a source separation or speech enhancement system to separate a source from the estimated source direction, or in a diary system to estimate the activity of multiple speakers. By combining information from multiple microphone arrays, directions can be merged to obtain source locations.</p><p><strong>Synthesis and Transformation:</strong> Among other applications, synthesis domain is where we see generative deep learning models being utilized more.</p><p>· Source separation is the process of extracting the signal corresponding to individual sources from a mixture of multiple sources. In realistic environments, often multiple sources are present which sum to a mixture signal, negatively affecting downstream signal processing tasks. Example application areas related to source separation include music editing and remixing, preprocessing for robust classification of speech and other sounds, or preprocessing to improve speech intelligibility.</p><p>· Speech enhancement techniques aim to improve the quality of speech by reducing noise. Recently, GANs have been shown to perform well in speech enhancement in the presence of additive noise, when enhancement is posed as a translation task from noisy signals to clean ones.</p><p>· Generative sound models synthesize sounds according to characteristics learned from a sound database, yielding realistic sound samples. The generated sound should be similar to sounds from which the model is trained, in terms of typical acoustic features (timbre, pitch content, rhythm). A basic requirement is that the sound should be recognizable as stemming from a particular object/process or intelligible, in the case of speech generation. At the same time, the generated sound should be original, i.e. it should be significantly different from sounds in the training set. A further requirement is that the generated sounds should show diversity. In addition, training and generation time should be small; ideally generation should be possible in real-time. Generative models can be evaluated both objectively and subjectively. Generated sounds can be tested objectively through a classifier or subjectively in a forced choice test with humans. A Turing test, asking a human to distinguish between real and synthesized audio examples, is a hard test for a model, since passing the Turing test requires that there is no perceivable difference between an examples being real or being synthesized.</p><h3>Neural Audio Codecs</h3><p>As mentioned above, audio signal processing has different requirements from image or video processing. In communication systems, transmitting sounds that are within the hearing range of humans are good enough (and thus we can reduce sampling bitrates to even as low as 8 kbps) but music data processing requires precision so much higher bitrates (usually 44.1 kHz or 48 kHz) are used. The lower the bitrate for an audio codec, the less intelligible and more robotic the voice signal becomes.</p><p>The drawback of waveform codecs is that they achieve this high quality by compressing and sending over the signal sample-by-sample, which requires a higher bitrate and, in most cases, isn’t necessary to achieve natural sounding speech.</p><p>Currently, the royalty-free open-source codec Opus is the most widely used codec for WebRTC-based VOIP applications and, with audio at 32kbps, typically obtains transparent speech quality, i.e., indistinguishable from the original. However, while Opus can be used in more bandwidth constrained environments down to 6kbps, it starts to demonstrate degraded audio quality at lower bitrates.</p><p>The most competitive results in AI usage in audio codecs domain come from Google Research. These experimental codecs will probably be commercialized soon and start to appear in open source chromium libraries.</p><p><strong>Lyra</strong> is a high-quality, very low-bitrate speech codec of which features are log mel spectrograms (logarithmic scale in companding) [17]. Lyra uses generative model-based packet-loss-concealment system which compensates for sample losses by generating completely new ones. Using these models as a baseline, Lyra is capable of reconstructing speech using minimal amounts of data. Computational complexity is a major concern of generative models but Lyra overcomes this problem by using a cheaper recurrent generative model that works at a lower rate, but generates in parallel multiple signals in different frequency ranges that it later combines into a single output signal at the desired sample rate. This trick enables Lyra to perform well on mid-range mobile phones despite the limited processor capacities and hence makes it quite suitable for the mobile applications. When compared to Opus, Lyra provides more than 60% bandwidth reduction at 3 kbps.</p><p><strong>Soundstream</strong> is another neural audio codec developed by Google Research on top of Lyra [18]. During training, the encoder, quantizer and decoder parameters are optimized using a combination of reconstruction and adversarial losses, computed by a discriminator, which is trained to distinguish between the original input audio and the reconstructed audio. SoundStream is designed in such a way that compression and enhancement can be carried out jointly by the same model, without increasing the overall latency. Soundstream uses vector quantization but enhances this method by splitting the quantization process in several layers so that the codebook size can be reduced drastically. During training, some quantization layers are randomly dropped to simulate a varying bitrate. This pushes the decoder to perform well at any bitrate of the incoming audio stream, and thus helps SoundStream to become “scalable” so that a single trained model can operate at any bitrate.</p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[2] Cisco Annual Internet Report (2018–2023) White Paper, <a href="https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html">https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html</a></p><p>[3] ITU JVET — Joint Video Experts Team, <a href="https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx">https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx</a></p><p>[4] <a href="https://en.wikipedia.org/wiki/Versatile_Video_Coding">https://en.wikipedia.org/wiki/Versatile_Video_Coding</a></p><p>[5] <a href="https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs">https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs</a></p><p>[6] Qualcomm Research Webinar: How AI research is enabling next gen codecs, <a href="https://www.youtube.com/watch?v=2lphrufyP0I">https://www.youtube.com/watch?v=2lphrufyP0I</a></p><p>[7] Generative Compression, Santurkar et al., <a href="https://arxiv.org/abs/1703.01467">https://arxiv.org/abs/1703.01467</a></p><p>[8] What is data redundancy in computer vision? , <a href="https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision">https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision</a></p><p>[9] Semantic Redundancies in Image-Classification Datasets: The 10% You Don’t Need, Birodkar et al., <a href="https://arxiv.org/pdf/1901.11409.pdf">https://arxiv.org/pdf/1901.11409.pdf</a></p><p>[10] Real-Time Adaptive Image Compression, Rippel &amp; Bourdev, <a href="https://doi.org/10.48550/arXiv.1705.05823">https://doi.org/10.48550/arXiv.1705.05823</a></p><p>[11] Probabilistic Programming for Deep Learning, Dustin Tran, PhD Thesis, Columbia University, 2020, <a href="https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download">https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download</a></p><p>[12] What The Heck Are VAE-GANs?, <a href="https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a">https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a</a></p><p>[13]<a href="https://developer.nvidia.com/maxine?ncid=so-yout-26905#cid=dl13_so-yout_en-us">What is NVIDIA Maxine?</a></p><p>[14] Inventing Virtual Meetings of Tomorrow with NVIDIA AI Research, <a href="https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be">https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be</a></p><p>[15] <a href="https://www.usenix.org/system/files/osdi18-yeo.pdf">Neural Adaptive Content-aware Internet Video Delivery</a> , Yeo et al.</p><p>[16] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schlüter, Shuoyiin Chang, Tara Sainath. Deep Learning for Audio Signal Processing. In Journal of Selected Topics of Signal Processing, Vol. 13, №2, May 2019, pages 206–219. <a href="http://doi.org/10.1109/JSTSP.2019.2908700">http://doi.org/10.1109/JSTSP.2019.2908700</a></p><p>[17] <a href="https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html">https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html</a></p><p>[18] <a href="https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html">https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html</a></p><p>[19] Machine learning for media compression: challenges and opportunities, Amir Said, <a href="https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities">https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities</a></p><p>[20] The next video codec will be based on AI, <a href="https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai">https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4fe97665e0aa" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/deep-learning-for-audio-signal-processing-4fe97665e0aa">Deep Learning for Audio Signal Processing</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Discussion on the Results of Selected Research in AI-Based Image and Video Compression]]></title>
            <link>https://medium.com/orion-innovation-techclub/a-discussion-on-the-results-of-selected-research-in-ai-based-image-and-video-compression-803024b956da?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/803024b956da</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[codec]]></category>
            <category><![CDATA[image-compression]]></category>
            <category><![CDATA[video-compression]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Wed, 10 Aug 2022 09:35:45 GMT</pubDate>
            <atom:updated>2022-09-09T06:31:58.081Z</atom:updated>
            <content:encoded><![CDATA[<p>This is part of a series of articles in <a href="https://medium.com/@selinmetin/ai-codecs-a128d2896362">AI Codecs</a>.</p><p>In this section I will mention a couple of selected research in video and image compression and comment on their results. There is no specific order or a reason for selection other than my own interest but so far the selection is capable of covering basic topics for neural compression techniques so a curious reader may start from here and proceed to learn more using the reference list and similar works.</p><p>In [10] Rippel and Bourdev propose a GAN-based approach to image compression with quite promising results. They compare their results to several commercial compression algorithms such as JPEG and their algorithm can perform in real-time. Although real-time requirement limits the capacity of the algorithm, non-real-time (e.g. storage related) applications can present growth area for this algorithm.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*kpre_uUTfVULtdiglNrqlw.png" /><figcaption>Sample output of the algorithm in [10].</figcaption></figure><p>Santurkar and her colleagues present generative compression systems for image and video compression in [7]. Their results are interesting to understand the impact of semantic redundancy in generative compression. In the figure below, rows e, f, and g show the results of generative image compression. In the shoes data set, we can see how generated data changes the semantics of the original data. For example, the red high heel shoes are always reconstructed similarly except for the final row where the color changes (and also highest compression factor is used). Similarly we can easily understand that the last column shows sandals but in fact generated outputs show quite different sandals (consider if you were online shopping and the image on the website is quite different from the product). Or the sports shoes in the middle rows look similar in average but a fashionista would definitely notice the difference. In the final group of landscape images, we understand this “making up facts instead of describing them” impact more. For example the last column starts with a snow capped mountain at the back and another dark looking mountain in the front but the final picture shows a dark but much shorter mountain at the back. The blue sky is there, mountain is there again but this is a completely different picture — as if an author described this landscape vaguely and different artists painted different versions of this description. This is an example of the problem when reconstructed images/videos have quite different outputs although they are semantically correct!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/1*fN5LUCFiq_AgtoEPmt3flA.png" /><figcaption>Taken from Figure 3 in [7].</figcaption></figure><p>For the performance of the video compression algorithm described in the same study [7], we can make similar comments again. In the original video sequences in the figure below, hands are lowered towards the end of the sequence while generational algorithm still shows them stretched to the sides.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*VBsJNb9zwJMIoDz79fs7WA.png" /><figcaption>Taken from Figure 5 in [7].</figcaption></figure><p>Qualcomm Research present many examples of deep generative learning based algorithms which also look like one of the most commercially viable results in AI based codecs domain [5, 6]. They use end-to-end deep learning for video compression. They also apply semantic-aware image compression: More bits are allocated for regions of interest in an image to increase the visual quality. (Region of interest coding is actually one of the enhancements of JPEG2000 standard and origins from the fact that humans pay more attention to faces, other human beings, etc. in an environment so we can increase bitrate for some regions on purpose while keeping the rest of the image at a low bitrate compression level. The perceived visual quality would still be quite acceptable. This technique is useful when you want to focus on specific parts of an image, e.g. in ads.)</p><p>Qualcomm Research proposes the use of overfitting (fine tuning your neural network [NN] to a specific data set so you obtain extremely good results for that data set only) through instance-adaptive video coding. In general, neural compression approaches are adaptive which means you can train your system on a particular kind of data. They exert this idea by training their NN system on a particular video to obtain better results. This method may not be useful for real-time compression tasks (e.g. video conferences) but since NN architectures have the flexibility of changing via training, overfitting can be applied for specific purposes where statistics of the input video are very well known (e.g. you don’t receive random images but it could be same type of video shot in a studio environment or always with similar background).</p><p>NVIDIA recently demonstrated the applications of Maxine which is a suite of GPU-accelerated SDKs that elevate standard microphones and cameras with AI to enhance audio and video communications [13, 14]. Maxine provides state-of-the-art real-time AI audio, video, and augmented reality features that can be built into customizable, end-to-end deep learning pipelines. The demo shows quite interesting usage of the AI capabilities such as making the speaker appear as if looking at the camera in fact he is really looking away from the camera. Since this is an already commercialized product and is part of NVIDIA’s AI platform, it is compatible with other SDKs from the same platform and can be combined to deliver enhanced experiences on multimedia communication applications.</p><p>An interesting application of deep neural network (DNN) based quality enhancement on video is proposed by Yeo and his colleagues [15]. The limitation of existing content distribution networks (CDNs) (e.g. Youtube) is that the quality heavily depends on the bandwidth between servers and clients. Predictions in Cisco’s Internet Report [2] show that streaming media will devour the network bandwidth and an important percentage of this consumption will be caused by CDN usage. Hence improvements proposed in this area will find numerous applications soon. Yeo’s model learns a mapping from a low-quality video to a high-quality version, e.g., superresolution. This enables clients to obtain high-definition (e.g., 1080p) video from lower quality transmissions. Content-aware model uses a different DNN for each video episode.</p><p>The content-aware DNN developed in this study [15] captures redundancies occurring at large time scales and operates over the entire video. In contrast, the conventional codecs deal with redundancy within a frame or between frames within a portion of the video segment. The steps of this scalable DNN model are below and results of a sample video frame are in shown in the following figure:</p><p>· A client’s video player first downloads a manifest file, which contains a list of available DNNs for the video. The client then selects one of the DNNs that fit its computing power.</p><p>· The player then downloads the selected DNN or video chunks.</p><p>· The DNN processor initializes the DNN as DNN chunks arrive.</p><p>· The player opportunistically passes video chunks in the playback buffer to the DNN processor for quality enhancement along with their associated playback time, which indicates the deadline for neural processing.</p><p>· The DNN processor first decodes a video chunk into frames and then applies the super resolution DNN.</p><p>· The resulting frames are then re-encoded to video chunks which replace the original chunks in the playback buffer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*p87oQDBDflvYhM-b3Ni1Vw.png" /><figcaption>Taken from Figure 2 in [15]</figcaption></figure><p>Above studies demonstrate various use cases of image and video compression using AI or machine learning (ML) based approaches. As noted before, GAN-based approaches may have the risk of introducing fake data or altering the original input such that authenticity of the information might be impacted. Despite this anticipated problem which is caused by the nature of GAN architectures, this method seems to be the most promising approach for compression goals.</p><p>Another clever approach is the one proposed in [15] where a manifest file announces the available DNNs so that the decoder can pick up the flavor which is applicable to its local resources. This approach can be elaborated in many other use cases. For example during peer to peer communications (e.g. video calls or conferences) both endpoints negotiate the codecs and decide on the optimal one for that session. The negotiation can include AI based compression &amp; decompression parameters so that bandwidth and local resource consumption can be improved. This use case will be most applicable to mobile communication scenarios where resources are usually limited, bandwidth is not optimal, interruptions are inevitable and end users usually pay per data size instead of minutes.</p><p>Although most of the work in this area is still done by academic researchers, we can see a couple of commercially available or ready for demonstration type of products such as Maxine of NVIDIA. Thus AI based video compression techniques have proved maturity for end user products and technology investment. Whatever the reason for your interest is, this is truly an interesting research area to monitor for the near future.</p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[2] Cisco Annual Internet Report (2018–2023) White Paper, <a href="https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html">https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html</a></p><p>[3] ITU JVET — Joint Video Experts Team, <a href="https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx">https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx</a></p><p>[4] <a href="https://en.wikipedia.org/wiki/Versatile_Video_Coding">https://en.wikipedia.org/wiki/Versatile_Video_Coding</a></p><p>[5] <a href="https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs">https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs</a></p><p>[6] Qualcomm Research Webinar: How AI research is enabling next gen codecs, <a href="https://www.youtube.com/watch?v=2lphrufyP0I">https://www.youtube.com/watch?v=2lphrufyP0I</a></p><p>[7] Generative Compression, Santurkar et al., <a href="https://arxiv.org/abs/1703.01467">https://arxiv.org/abs/1703.01467</a></p><p>[8] What is data redundancy in computer vision? , <a href="https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision">https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision</a></p><p>[9] Semantic Redundancies in Image-Classification Datasets: The 10% You Don’t Need, Birodkar et al., <a href="https://arxiv.org/pdf/1901.11409.pdf">https://arxiv.org/pdf/1901.11409.pdf</a></p><p>[10] Real-Time Adaptive Image Compression, Rippel &amp; Bourdev, <a href="https://doi.org/10.48550/arXiv.1705.05823">https://doi.org/10.48550/arXiv.1705.05823</a></p><p>[11] Probabilistic Programming for Deep Learning, Dustin Tran, PhD Thesis, Columbia University, 2020, <a href="https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download">https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download</a></p><p>[12] What The Heck Are VAE-GANs?, <a href="https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a">https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a</a></p><p>[13]<a href="https://developer.nvidia.com/maxine?ncid=so-yout-26905#cid=dl13_so-yout_en-us">What is NVIDIA Maxine?</a></p><p>[14] Inventing Virtual Meetings of Tomorrow with NVIDIA AI Research, <a href="https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be">https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be</a></p><p>[15] <a href="https://www.usenix.org/system/files/osdi18-yeo.pdf">Neural Adaptive Content-aware Internet Video Delivery</a> , Yeo et al.</p><p>[16] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schlüter, Shuoyiin Chang, Tara Sainath. Deep Learning for Audio Signal Processing. In Journal of Selected Topics of Signal Processing, Vol. 13, №2, May 2019, pages 206–219. <a href="http://doi.org/10.1109/JSTSP.2019.2908700">http://doi.org/10.1109/JSTSP.2019.2908700</a></p><p>[17] <a href="https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html">https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html</a></p><p>[18] <a href="https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html">https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html</a></p><p>[19] Machine learning for media compression: challenges and opportunities, Amir Said, <a href="https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities">https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities</a></p><p>[20] The next video codec will be based on AI, <a href="https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai">https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=803024b956da" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/a-discussion-on-the-results-of-selected-research-in-ai-based-image-and-video-compression-803024b956da">A Discussion on the Results of Selected Research in AI-Based Image and Video Compression</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Generative Compression]]></title>
            <link>https://medium.com/orion-innovation-techclub/generative-compression-8b1e6253b104?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/8b1e6253b104</guid>
            <category><![CDATA[generative-adversarial]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[multimedia-compression]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <category><![CDATA[codec]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Wed, 10 Aug 2022 09:34:49 GMT</pubDate>
            <atom:updated>2022-09-09T06:31:19.643Z</atom:updated>
            <content:encoded><![CDATA[<p>This is part of a series of articles in <a href="https://medium.com/@selinmetin/ai-codecs-a128d2896362">AI Codecs</a>.</p><p>As mentioned before, traditional image and video compression algorithms rely on hand-crafted encoder/decoder pairs (codecs) that lack adaptability and are agnostic to the data being compressed. Hence they do not degrade gracefully, i.e. resulting losses may harm the optimality of perceived quality.</p><p>Traditional codecs apply linear analysis and synthesis transforms (e.g. DCT [JPEG] and multi-scale orthogonal wavelet decomposition [JPEG2000]). Although the ground theory of transform coding is very strong (it roots from Shannon’s Theorem like many other communication technology principles), unfortunately a linear function is not optimal for compressing the full spectrum of natural images (and sequence of images, hence videos). Even presuming that current transform coding algorithms are optimal for a particular class of bitmap images, this performance is unlikely to generalize to emerging media formats.</p><p>A clever approach to solve this problem was introduced in the past decade by replacing hand-crafted linear transforms with artificial neural networks so that both encoder and decoder algorithms are both learnt neural network functions.</p><h3>Generative Adversarial Networks</h3><p>Generative Adversarial Networks (GANs) apply a generator function to approximate high dimensional input space to a transformation of some prior latent. In other words, GANs invent new data which align with the original input data statistics so that the new data looks like the original (both perceptibly and mathematically). To be able to generate realistic looking samples at the decoder output, we would like to train the decoder function to minimize the difference between its output distribution and the unknown true distribution. A popular method to achieve this is to introduce an auxiliary discriminator network which learns to map the input data to the probability that it was sampled from the original input instead of generated samples [7]. Although GANs provide an appealing method for reconstructing quality images, they lack the encoder function necessary for image compression.</p><p>The adversarial training framework is particularly relevant to the compression world. In traditional codecs, distortions often take the form of blurriness, pixelation, and so on. These artifacts are unappealing, but are increasingly noticeable as the bitrate is lowered [8]. Generative structure of GAN-based codecs can fill in for data loss during decoding stage and discriminator structures are used to increase quality of the generated sequences. Since there is a high potential of error concealment during reconstruction stage, compression rates can be increased much more than the traditional algorithms.</p><p>GAN-based codecs can produce much more visually appealing content at very low bits per pixel. Figure below gives the simple architecture of GAN-based codecs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*7APlpyYwQhyNQKsUgdletQ.png" /><figcaption>Taken from [6], around 40:11.</figcaption></figure><p>Encoder and generator (decoder) and prior are used to compress the latents. A third NN called discriminator classifies whether the input is an image from the data distribution or something that is output by the decoder. Discriminator looks for artifacts and other clues to understand that the image is not real but a compressed one. The task of the blue components of the system is to fool the discriminator. The system tries to produce realistic looking outputs.</p><p>GANs produce data that look like the original but is not coming from the original input data set. Although this is quite practical since you can use extreme methods to get rid of redundancy knowing that you will be able to produce the lost input sequence during reconstruction, there is a danger also: Reconstructed images/videos may have quite different outputs although they are semantically correct! We will see examples of this later.</p><h3>Auto Encoder Model</h3><p>An alternative to GANs for generative image modeling are variational autoencoders (VAEs). Similar to GANs, VAEs introduce an auxiliary network to facilitate training. Unlike GANs, this encoder function is trained to learn an approximation of the true posterior and thus can be used as an encoder for image compression.</p><p>Auto encoder based models contain two neural networks, one at each encoder and decoder sides. Encoder architecture is fine tuned for the input type of data. This NN outputs features that capture most important info. This feature representation can be quantized and compressed into a bitstream and transmitted then. The decoder side feeds features into NN to reproduce the input as closely as possible. Variational auto encoders (VAEs) introduce a third component to this system which is called the priors. VAEs try to predict the feature representation. It is possible to losslessly compress the latent representation to squeeze out any redundancy that may be left.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/563/1*qnz0-T8s_N7b522GUeOh-g.png" /><figcaption>Taken from [6], around 10:13.</figcaption></figure><p>In [12] a simple explanation of how GANs, VAEs and proposed VAE-GAN architectures work is given. However, a limitation of VAEs (and autoencoders more generally) is that maximizing a Gaussian likelihood is equivalent to minimizing the L2 loss between pixel intensity vectors. This loss is known to correlate poorly with human perception and leads to blurry reconstructions.</p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[2] Cisco Annual Internet Report (2018–2023) White Paper, <a href="https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html">https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html</a></p><p>[3] ITU JVET — Joint Video Experts Team, <a href="https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx">https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx</a></p><p>[4] <a href="https://en.wikipedia.org/wiki/Versatile_Video_Coding">https://en.wikipedia.org/wiki/Versatile_Video_Coding</a></p><p>[5] <a href="https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs">https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs</a></p><p>[6] Qualcomm Research Webinar: How AI research is enabling next gen codecs, <a href="https://www.youtube.com/watch?v=2lphrufyP0I">https://www.youtube.com/watch?v=2lphrufyP0I</a></p><p>[7] Generative Compression, Santurkar et al., <a href="https://arxiv.org/abs/1703.01467">https://arxiv.org/abs/1703.01467</a></p><p>[8] What is data redundancy in computer vision? , <a href="https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision">https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision</a></p><p>[9] Semantic Redundancies in Image-Classification Datasets: The 10% You Don’t Need, Birodkar et al., <a href="https://arxiv.org/pdf/1901.11409.pdf">https://arxiv.org/pdf/1901.11409.pdf</a></p><p>[10] Real-Time Adaptive Image Compression, Rippel &amp; Bourdev, <a href="https://doi.org/10.48550/arXiv.1705.05823">https://doi.org/10.48550/arXiv.1705.05823</a></p><p>[11] Probabilistic Programming for Deep Learning, Dustin Tran, PhD Thesis, Columbia University, 2020, <a href="https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download">https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download</a></p><p>[12] What The Heck Are VAE-GANs?, <a href="https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a">https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a</a></p><p>[13]<a href="https://developer.nvidia.com/maxine?ncid=so-yout-26905#cid=dl13_so-yout_en-us">What is NVIDIA Maxine?</a></p><p>[14] Inventing Virtual Meetings of Tomorrow with NVIDIA AI Research, <a href="https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be">https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be</a></p><p>[15] <a href="https://www.usenix.org/system/files/osdi18-yeo.pdf">Neural Adaptive Content-aware Internet Video Delivery</a> , Yeo et al.</p><p>[16] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schlüter, Shuoyiin Chang, Tara Sainath. Deep Learning for Audio Signal Processing. In Journal of Selected Topics of Signal Processing, Vol. 13, №2, May 2019, pages 206–219. <a href="http://doi.org/10.1109/JSTSP.2019.2908700">http://doi.org/10.1109/JSTSP.2019.2908700</a></p><p>[17] <a href="https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html">https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html</a></p><p>[18] <a href="https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html">https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html</a></p><p>[19] Machine learning for media compression: challenges and opportunities, Amir Said, <a href="https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities">https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities</a></p><p>[20] The next video codec will be based on AI, <a href="https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai">https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8b1e6253b104" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/generative-compression-8b1e6253b104">Generative Compression</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deep Learning and Compression]]></title>
            <link>https://medium.com/orion-innovation-techclub/deep-learning-and-compression-50731505f6af?source=rss-3bf0699bda15------2</link>
            <guid isPermaLink="false">https://medium.com/p/50731505f6af</guid>
            <category><![CDATA[codec]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[multimedia-compression]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[techclub-tr]]></category>
            <dc:creator><![CDATA[Selin Metin]]></dc:creator>
            <pubDate>Wed, 10 Aug 2022 09:33:35 GMT</pubDate>
            <atom:updated>2022-09-09T13:58:07.409Z</atom:updated>
            <content:encoded><![CDATA[<p>This is part of a series of articles in <a href="https://medium.com/@selinmetin/ai-codecs-a128d2896362">AI Codecs</a>.</p><p>While digital media are transmitted in a wide variety of settings, the available codecs are “one-size-fits-all”: they are hard-coded, and cannot be customized to particular use cases beyond high-level hyperparameter tuning [10]. In the last few years, deep learning has revolutionized many tasks such as machine translation, speech recognition, face recognition, natural language processing and photo-realistic image generation. The world of compression has not yet benefited from these advancements, for two main reasons:</p><p>· Deep learning primitives (extracted characteristic features of data), in their raw forms, are not well-suited to construct representations sufficiently compact.</p><p>· It is difficult to develop a deep learning compression approach sufficiently efficient for deployment in environments constrained by computation power, memory footprint and battery life.</p><p>Given unlabeled training data, deep learning based models generate new samples from the input data distribution [6]. These are called deep generative models and have powerful capabilities such as extracting features by learning a low-dimension feature representation of the input space and sampling to generate, restore, predict or compress data. Despite the above mentioned difficulties about applying deep learning algorithms, the research and outcomes in this emerging technological area promise serious improvements for codecs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/580/1*cKaT5QCDiUZvWpJ1KQ-m2A.png" /><figcaption>Comic taken from <a href="https://xkcd.com/1838/">https://xkcd.com/1838/</a></figcaption></figure><h3>Common Probability Terms Used for Deep Learning</h3><p>Before we dive into the depths of neural compression methods, we’d rather remember some of the terminology and mathematical concepts which are extensively used in AI related tasks [11].</p><p>A <strong>probabilistic model</strong> asserts how observations from a natural phenomenon arise. The model is a joint distribution of observed variables corresponding to data and latent variables that provide the hidden structure to generate from the observed variables.</p><p><strong>Latent variables</strong> are variables that are not directly observed but rather calculated via mathematical models from other variables that are observed or directly measured. When the distribution of these latents follows some other distribution, this other distribution is called the prior.</p><p>The <strong>prior</strong> is a probability distribution that describes the latent variables present in the data.</p><p>The <strong>posterior</strong> is the distribution of the latent variables conditioned on the observed data. It is a probabilistic description of the data’s hidden representation.</p><p>The <strong>likelihood</strong> is a probability distribution that describes how any data depend on the latent variables.</p><p><strong>Loss functions</strong> are used to minimize errors in machine learning. L1 Loss (Least Absolute Deviations — LAD) is used to minimize the error which is the sum of the all the absolute differences between the true value and the predicted value. L2 Loss (Least Square Errors — LS) is used to minimize the error which is the sum of the all the squared differences between the true value and the predicted value. Generally L2 loss is preferred but when the dataset has many outliers, taking squares of differences quickly leads to much larger error values.</p><h3>References</h3><p>The reference list is common for all articles in this series.</p><p>[1] Ze-Nian Li, Mark S Drew, “Fundamentals of Multimedia”, Second Edition 2014</p><p>[2] Cisco Annual Internet Report (2018–2023) White Paper, <a href="https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html">https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.html</a></p><p>[3] ITU JVET — Joint Video Experts Team, <a href="https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx">https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx</a></p><p>[4] <a href="https://en.wikipedia.org/wiki/Versatile_Video_Coding">https://en.wikipedia.org/wiki/Versatile_Video_Coding</a></p><p>[5] <a href="https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs">https://www.qualcomm.com/news/onq/2021/07/how-ai-research-enabling-next-gen-codecs</a></p><p>[6] Qualcomm Research Webinar: How AI research is enabling next gen codecs, <a href="https://www.youtube.com/watch?v=2lphrufyP0I">https://www.youtube.com/watch?v=2lphrufyP0I</a></p><p>[7] Generative Compression, Santurkar et al., <a href="https://arxiv.org/abs/1703.01467">https://arxiv.org/abs/1703.01467</a></p><p>[8] What is data redundancy in computer vision? , <a href="https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision">https://www.lightly.ai/post/what-is-data-redundancy-in-computer-vision</a></p><p>[9] Semantic Redundancies in Image-Classification Datasets: The 10% You Don’t Need, Birodkar et al., <a href="https://arxiv.org/pdf/1901.11409.pdf">https://arxiv.org/pdf/1901.11409.pdf</a></p><p>[10] Real-Time Adaptive Image Compression, Rippel &amp; Bourdev, <a href="https://doi.org/10.48550/arXiv.1705.05823">https://doi.org/10.48550/arXiv.1705.05823</a></p><p>[11] Probabilistic Programming for Deep Learning, Dustin Tran, PhD Thesis, Columbia University, 2020, <a href="https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download">https://academiccommons.columbia.edu/doi/10.7916/d8-0yf2-wf27/download</a></p><p>[12] What The Heck Are VAE-GANs?, <a href="https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a">https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a</a></p><p>[13]<a href="https://developer.nvidia.com/maxine?ncid=so-yout-26905#cid=dl13_so-yout_en-us">What is NVIDIA Maxine?</a></p><p>[14] Inventing Virtual Meetings of Tomorrow with NVIDIA AI Research, <a href="https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be">https://www.youtube.com/watch?app=desktop&amp;v=NqmMnjJ6GEg&amp;feature=youtu.be</a></p><p>[15] <a href="https://www.usenix.org/system/files/osdi18-yeo.pdf">Neural Adaptive Content-aware Internet Video Delivery</a> , Yeo et al.</p><p>[16] Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schlüter, Shuoyiin Chang, Tara Sainath. Deep Learning for Audio Signal Processing. In Journal of Selected Topics of Signal Processing, Vol. 13, №2, May 2019, pages 206–219. <a href="http://doi.org/10.1109/JSTSP.2019.2908700">http://doi.org/10.1109/JSTSP.2019.2908700</a></p><p>[17] <a href="https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html">https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.html</a></p><p>[18] <a href="https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html">https://ai.googleblog.com/2021/08/soundstream-end-to-end-neural-audio.html</a></p><p>[19] Machine learning for media compression: challenges and opportunities, Amir Said, <a href="https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities">https://www.researchgate.net/publication/327576518_Machine_learning_for_media_compression_Challenges_and_opportunities</a></p><p>[20] The next video codec will be based on AI, <a href="https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai">https://www.redsharknews.com/the-next-video-codec-will-be-based-on-ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50731505f6af" width="1" height="1" alt=""><hr><p><a href="https://medium.com/orion-innovation-techclub/deep-learning-and-compression-50731505f6af">Deep Learning and Compression</a> was originally published in <a href="https://medium.com/orion-innovation-techclub">Orion Innovation techClub</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>