<?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 Kasrasadatsharifi on Medium]]></title>
        <description><![CDATA[Stories by Kasrasadatsharifi on Medium]]></description>
        <link>https://medium.com/@kasrasadatsharifi?source=rss-ddba48cfe4bc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*qPBtZz-atFS1if8ZuxzZoQ.jpeg</url>
            <title>Stories by Kasrasadatsharifi on Medium</title>
            <link>https://medium.com/@kasrasadatsharifi?source=rss-ddba48cfe4bc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:16:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kasrasadatsharifi/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[Vision Transformers Part1: Why Replace Convolution with Attention]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://levelup.gitconnected.com/vision-transformers-part1-why-replace-convolution-with-attention-000c641c039d?source=rss-ddba48cfe4bc------2"><img src="https://cdn-images-1.medium.com/max/1672/1*up_5LoGUHxtXrXCf6J20rQ.png" width="1672"></a></p><p class="medium-feed-snippet">Understanding vision transformers foundation, architectural differences and performance comparisons with CNNs. Some of the graphs in this&#x2026;</p><p class="medium-feed-link"><a href="https://levelup.gitconnected.com/vision-transformers-part1-why-replace-convolution-with-attention-000c641c039d?source=rss-ddba48cfe4bc------2">Continue reading on Level Up Coding »</a></p></div>]]></description>
            <link>https://levelup.gitconnected.com/vision-transformers-part1-why-replace-convolution-with-attention-000c641c039d?source=rss-ddba48cfe4bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/000c641c039d</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[model-evaluation]]></category>
            <category><![CDATA[vision-transformer]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Kasrasadatsharifi]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 16:49:46 GMT</pubDate>
            <atom:updated>2026-11-06T16:49:46.401Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ConvNeXt Explained: How CNNs Borrowed From Vision Transformers]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pub.towardsai.net/convnext-explained-how-cnns-borrowed-from-vision-transformers-dbf4428801f2?source=rss-ddba48cfe4bc------2"><img src="https://cdn-images-1.medium.com/max/1672/1*EiVNPokFyhLZAPfahdVDPw.png" width="1672"></a></p><p class="medium-feed-snippet">Understanding the architectural decisions behind ConvNeXt, advantages and disadvantages over Vision Transformers and older CNNs.</p><p class="medium-feed-link"><a href="https://pub.towardsai.net/convnext-explained-how-cnns-borrowed-from-vision-transformers-dbf4428801f2?source=rss-ddba48cfe4bc------2">Continue reading on Towards AI »</a></p></div>]]></description>
            <link>https://pub.towardsai.net/convnext-explained-how-cnns-borrowed-from-vision-transformers-dbf4428801f2?source=rss-ddba48cfe4bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/dbf4428801f2</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[model-explanation]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[model-evaluation]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Kasrasadatsharifi]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 23:01:02 GMT</pubDate>
            <atom:updated>2026-10-28T23:01:02.884Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[When Does a VLM Add Value to Your Project?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pub.towardsai.net/when-does-a-vlm-add-value-to-your-project-212dfcda012d?source=rss-ddba48cfe4bc------2"><img src="https://cdn-images-1.medium.com/max/1672/1*7t9hT29s5P2xrh996rvODA.png" width="1672"></a></p><p class="medium-feed-snippet">This is a personal project that I worked on to understand the advantages and disadvantages of VLMs in computer vision projects. Figures&#x2026;</p><p class="medium-feed-link"><a href="https://pub.towardsai.net/when-does-a-vlm-add-value-to-your-project-212dfcda012d?source=rss-ddba48cfe4bc------2">Continue reading on Towards AI »</a></p></div>]]></description>
            <link>https://pub.towardsai.net/when-does-a-vlm-add-value-to-your-project-212dfcda012d?source=rss-ddba48cfe4bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/212dfcda012d</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[multimodal-ai]]></category>
            <category><![CDATA[vision-language-model]]></category>
            <category><![CDATA[model-evaluation]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Kasrasadatsharifi]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 14:31:01 GMT</pubDate>
            <atom:updated>2026-10-20T14:31:01.997Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[How much should you fine-tune your vision transformer with limited data]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://levelup.gitconnected.com/how-much-should-you-fine-tune-your-vision-transformer-with-limited-data-7a35318cd461?source=rss-ddba48cfe4bc------2"><img src="https://cdn-images-1.medium.com/max/1672/1*y8niA43KwlPWanQJRvw_9Q.png" width="1672"></a></p><p class="medium-feed-snippet">In this experiment I wanted to understand how much of my Vision Transformer (ViT) should I fine tune to my limited dataset for a better&#x2026;</p><p class="medium-feed-link"><a href="https://levelup.gitconnected.com/how-much-should-you-fine-tune-your-vision-transformer-with-limited-data-7a35318cd461?source=rss-ddba48cfe4bc------2">Continue reading on Level Up Coding »</a></p></div>]]></description>
            <link>https://levelup.gitconnected.com/how-much-should-you-fine-tune-your-vision-transformer-with-limited-data-7a35318cd461?source=rss-ddba48cfe4bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/7a35318cd461</guid>
            <category><![CDATA[vision-transformer]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[fine-tuning-transformer]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[transfer-learning]]></category>
            <dc:creator><![CDATA[Kasrasadatsharifi]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 15:28:21 GMT</pubDate>
            <atom:updated>2026-10-16T15:28:21.005Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[From OpenCV experiments to an Interactive Image Processor]]></title>
            <link>https://levelup.gitconnected.com/from-opencv-experiments-to-an-interactive-image-processor-81b2c5d7eb6c?source=rss-ddba48cfe4bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/81b2c5d7eb6c</guid>
            <category><![CDATA[opencv]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[image-processing]]></category>
            <category><![CDATA[interactive-content]]></category>
            <category><![CDATA[educational-technology]]></category>
            <dc:creator><![CDATA[Kasrasadatsharifi]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 12:22:25 GMT</pubDate>
            <atom:updated>2026-10-21T17:01:07.160Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This was a personal experiment to understand and implement some of the traditional image processing techniques.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OxjDil9QG4LhH66-kuQVLg.png" /><figcaption>a view of the interactive image processing app</figcaption></figure><p>I wanted to refresh my memory on some of the commonly used openCV image processing techniques that I had used before or currently am using and get familiar with some of the ones I have not worked with yet. I started this as a experimental jupyter notebook where in each block I tried to run a technique and test it on an image.</p><p>As I made progress and implemented more and more of these image processing methods the process became repetitive and cumbersome. I had to write a whole block of code to test a parameter of a functionality or run a loop with different parameters and display the image to see the difference. It got even worse when some of the parameters had subtle changes in the image that I could not just notice with a few different images.</p><p>Calling an OpenCV function is usually the easy part. Understanding why it produced a particular result is harder. Documentation can explain what each argument means. A notebook can show the output of one experiment. Neither automatically gives you an intuition for how the result changes as a parameter changes.</p><p>That is where the idea for the <strong>Interactive Image Processor</strong> came from.</p><h3>What was the experiment</h3><p>The experiment covered different parts of a classical computer-vision pipeline, including:</p><ul><li>color-space conversion;</li><li>smoothing and noise reduction;</li><li>edge detection;</li><li>binary and automatic thresholding;</li><li>contrast enhancement;</li><li>contour detection;</li><li>geometric transformations; and</li><li>simple image augmentations.</li></ul><p>In each block I used the same image or its converted grayscale to run these operations. some operations like edge detection has one or two parameters for lower and upper bound thresholds. Others such as colour domain operations (LAB and HSV) have more parameters to adjust and compare.</p><p>Comparing several static outputs was not enough to understand how parameter changes affected the image. I needed a way to observe those changes continuously.</p><p>This needed a more systematic approach.</p><h3>Turning trial and error into a visible feedback loop</h3><p>The idea behind the app is simple:</p><ol><li>Start with the built-in test image or upload your own.</li><li>Optionally switch the working image to grayscale.</li><li>Choose an image-processing operation.</li><li>Change one or more parameters.</li><li>Compare the result with the original and with nearby parameter values.</li><li>Inspect the luminance histogram when it is useful.</li><li>Copy the matching Python code or download the processed image.</li></ol><p>The app is built with Python and Streamlit and uses openCV for image processing operations. The interactive interface allows you to change parameters for each operation and see the result as you change it.</p><p>As an example, by choosing the Canny edge detection, you can modify the lower and upper thresholds and watch the edges disappear or reappear. In Gaussian blur, you can adjust the kernel size and see the details suppressed live. With Fourier transform, you can map the image to its frequency domain and see why we need to shift and log the frequencies or watch the phase spectrum.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1nt83mTclHsNBc4524B0dw.png" /></figure><h3>What the app does</h3><p>The application starts up as you click the link (if inactive, you just need to wake it up and wait a few seconds for it to load up). For each technique, the image processor brings several pieces of the learning process into one place:</p><ul><li><strong>A plain-language explanation</strong> of what the operation does and when it may be useful.</li><li><strong>Interactive controls</strong> for its important parameters.</li><li><strong>A side-by-side comparison</strong> of the source image and the processed result.</li><li><strong>Parameter sweeps</strong> that show several settings together instead of one isolated output.</li><li><strong>Luminance histograms</strong> for understanding how operations change an image’s intensity distribution.</li><li><strong>Copy-ready Python code</strong> that reflects the settings currently selected in the interface.</li><li><strong>Result downloads</strong> for saving an output and using it in another experiment.</li></ul><h3>What can it be used for</h3><p>I built this app for three main use cases:</p><ul><li><strong>Learning:</strong> The tool helps users explore image-processing operations and understand how different parameters affect their results. I learned a great deal about these techniques myself while developing it.</li><li><strong>Teaching:</strong> The interactive interface makes it easier to demonstrate commonly used image-processing techniques and their effects without writing additional code.</li><li><strong>Prototyping: </strong>which is one of the most important parts of production. There have been many times that I had to run a lot of code just to test a single image. I use this tool now to prototype images and it has saved me a lot of time.</li></ul><h3>Where the project goes next</h3><p>As the application evolves, I will also use this series to document new releases and versions. These updates will explain which operations, controls, and supporting features were added, why they were introduced, and how users can benefit from them.</p><p>Feedback will play an important role in deciding what comes next. Whether you’re a student, educator, developer, or image-processing practitioner, I would love to hear about your experience. Try the app with the built-in image or one of your own, and let me know what you found useful, what could be improved, and which features you would like to see added.</p><h3>Resources</h3><ul><li><strong>Try the Interactive Image Processor:</strong> <a href="https://interactiveimageprocessing.streamlit.app/">The app</a></li><li><strong>View the source and experiments:</strong> <a href="https://github.com/kasrasa/image_processing_experiments">GitHub repository</a></li><li><strong>OpenCV documentation:</strong> <a href="https://docs.opencv.org/4.x/">docs.opencv.org</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=81b2c5d7eb6c" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/from-opencv-experiments-to-an-interactive-image-processor-81b2c5d7eb6c">From OpenCV experiments to an Interactive Image Processor</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bootstrap in statistics]]></title>
            <link>https://medium.com/@kasrasadatsharifi/bootstrap-in-statistics-b6b3718e7022?source=rss-ddba48cfe4bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6b3718e7022</guid>
            <category><![CDATA[bootstrapping]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[bootstrap]]></category>
            <dc:creator><![CDATA[Kasrasadatsharifi]]></dc:creator>
            <pubDate>Thu, 11 Sep 2025 19:50:31 GMT</pubDate>
            <atom:updated>2025-09-11T19:50:31.772Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Definition</strong></h3><p>Bootstrapping is a method to measure the distribution of an estimator, a statistic. The estimator is any functional of the source data. For instance, a defined statistic on a sequence of random variables, X1,…,Xn, Tn(X)=G(X1,…,Xn) is an estimator.</p><p>With bootstrapping one can find the spread of the estimator in question. For instance, if we want to find the average age of a population, bootstrapping help us find the standard error, variation, of this average when we take different samples from the dataset.</p><h3><strong>General procedure</strong></h3><p>Bootstrapping can be done in 4 steps:</p><ol><li>Extract the bootstrapped distribution: First step is to create a bootstrapped distribution so we can then resample from. This distribution is achieved from the original data. For example, we create a limited population of the people on earth as our bootstrapped distribution. This is a smaller population than the people on earth but with a big enough sample set, bootstrap will give us a good enough results.</li></ol><ul><li>Note: to explain the good enough result we can use the Law of Large Numbers (LLN) to show that if our bootstrapped data is large enough the expected value of the estimator will converge to the true estimator from the original dataset. In the case below, the average of the sampled data will converge to the true expected value using the LLN because the distribution will in probability converge to the true distribution of the data.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/702/1*XqyXDTyYgsSeU6xCALxA_A.png" /></figure><ol><li>Create sample population with resampling: This is an important item in bootstrapping. The sampled data is the same size as the bootstrapped data but duplicate samples are allowed, hence the term, sample with resampling.</li><li>Calculate the estimator: With our distribution created we can calculate our estimator parameter.</li><li>Repeat steps 2 and 3 many times.</li></ol><p>At the end of this process we will end up with a distribution representing the calculated estimator parameter of our resampled data. We can use this distribution to measure different things such as standard error, confidence interval, etc.</p><p><strong>Standard Error: </strong>standard error is a measurement that indicates how our estimator varies with our sampled dataset. For instance, if you measure the average age of a sample group of people and it ranges within 30–32 years, the standard error is low. It means most of the samples you have picked to represent the whole population are from 30–32. It is worth mentioning that this does not mean that the parameter we are estimating is good or bad. Maybe age is a wrong parameter to estimate when we are dealing with how high should people be to ride a roller coaster but we have accurately guessed the average age of the population. Therefore standard error, as the name suggests, just shows how much of a spread this estimate has.</p><p><strong>Confidence interval: </strong>confidence interval is a range withins which 95% of our measurement fall. In other words, we can be 95% sure that our next sampled dataset will have the calculated average within the specified range.</p><p>There are different methods of calculating confidence intervals but the simplest method is the normal confidence interval.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/244/1*KEQxKAloT06lU_uRAnjZHw.png" /></figure><p>Where <em>se </em>is the standard error of the estimator distribution. It is essentially the standard deviation calculated from each sampled dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/590/1*JY3bubH_XNd0hOBvvH-GIg.png" /></figure><p><strong>Advantages and disadvantages:</strong></p><p>Advantages:</p><ol><li>Bootstrapping has the freedom to estimate any parameter or estimator that is related to the dataset.</li><li>It can be applied to nonparametric and parametric estimation.</li><li>It has not assumption about the dataset in the non parametric estimation about the dataset</li></ol><p>Disadvantages:</p><ol><li>The obvious disadvantage is the time consuming process of iterating the same steps for thousands of times.</li><li>It assumes i.i.d in the data. Therefore, any dependence in the data will break this method.</li><li>Its results is dependent on the bootstrapped dataset. If the bootstrapped dataset is not representative of the original data then the results of the bootstrapping are not reflective of the true estimator.</li></ol><h3><strong>Conclusion:</strong></h3><p>Bootstrapping can be powerful tool to estimate a parameter from a dataset. It is applicable for any estimator related to the dataset and can be applied in parametric or non parametric estimation. Given its iterative nature, it can be time consuming and have some general assumptions about the dataset and sampled data such as i.i.d but this assumption holds for many applications.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6b3718e7022" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Residual Network Experimentation on CIFAR10]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://levelup.gitconnected.com/residual-network-experimentation-on-cifar10-a52dc5b9d23f?source=rss-ddba48cfe4bc------2"><img src="https://cdn-images-1.medium.com/max/1200/0*SzVpp6mIspg7jF8i.jpg" width="1200"></a></p><p class="medium-feed-snippet">Residual networks, introduced in Deep Residual Learning for Image Recognition (He et al., 2015; CVPR 2016), are among the most widely used&#x2026;</p><p class="medium-feed-link"><a href="https://levelup.gitconnected.com/residual-network-experimentation-on-cifar10-a52dc5b9d23f?source=rss-ddba48cfe4bc------2">Continue reading on Level Up Coding »</a></p></div>]]></description>
            <link>https://levelup.gitconnected.com/residual-network-experimentation-on-cifar10-a52dc5b9d23f?source=rss-ddba48cfe4bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/a52dc5b9d23f</guid>
            <category><![CDATA[latent-space]]></category>
            <category><![CDATA[resnet]]></category>
            <category><![CDATA[supervised-learning]]></category>
            <category><![CDATA[disentanglement]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Kasrasadatsharifi]]></dc:creator>
            <pubDate>Mon, 18 Aug 2025 15:28:36 GMT</pubDate>
            <atom:updated>2026-10-16T14:12:33.991Z</atom:updated>
        </item>
    </channel>
</rss>