<?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 Ojasmore on Medium]]></title>
        <description><![CDATA[Stories by Ojasmore on Medium]]></description>
        <link>https://medium.com/@ojasmore143?source=rss-05f56e0fa241------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*b76ZHkOUy19NNj2ho8Md5g.jpeg</url>
            <title>Stories by Ojasmore on Medium</title>
            <link>https://medium.com/@ojasmore143?source=rss-05f56e0fa241------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 17:18:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ojasmore143/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[FastAPI: A modern Python framework for high-performance APIs]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@ojasmore143/fastapi-a-modern-python-framework-for-high-performance-apis-f89f8dbc20b2?source=rss-05f56e0fa241------2"><img src="https://cdn-images-1.medium.com/max/1024/1*3qnS3AT8IotbIBuQNyMlOw.png" width="1024"></a></p><p class="medium-feed-snippet">In the last few years, FastAPI has become the most loved framework in the Python ecosystem.</p><p class="medium-feed-link"><a href="https://medium.com/@ojasmore143/fastapi-a-modern-python-framework-for-high-performance-apis-f89f8dbc20b2?source=rss-05f56e0fa241------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@ojasmore143/fastapi-a-modern-python-framework-for-high-performance-apis-f89f8dbc20b2?source=rss-05f56e0fa241------2</link>
            <guid isPermaLink="false">https://medium.com/p/f89f8dbc20b2</guid>
            <category><![CDATA[rest-api]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[backend-development]]></category>
            <category><![CDATA[fastapi]]></category>
            <dc:creator><![CDATA[Ojasmore]]></dc:creator>
            <pubDate>Tue, 30 Dec 2025 17:54:43 GMT</pubDate>
            <atom:updated>2025-12-30T17:54:43.403Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Architecture of Transformer]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@ojasmore143/architecture-of-transformer-cf28584a0d0f?source=rss-05f56e0fa241------2"><img src="https://cdn-images-1.medium.com/max/1024/1*SZE9r3vaGL2PUztfXgN76Q.png" width="1024"></a></p><p class="medium-feed-snippet">If you are not a member on Medium, then click here to read full story</p><p class="medium-feed-link"><a href="https://medium.com/@ojasmore143/architecture-of-transformer-cf28584a0d0f?source=rss-05f56e0fa241------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@ojasmore143/architecture-of-transformer-cf28584a0d0f?source=rss-05f56e0fa241------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf28584a0d0f</guid>
            <category><![CDATA[architecture]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[gpt]]></category>
            <category><![CDATA[attention]]></category>
            <category><![CDATA[transformers]]></category>
            <dc:creator><![CDATA[Ojasmore]]></dc:creator>
            <pubDate>Sun, 21 Sep 2025 15:10:34 GMT</pubDate>
            <atom:updated>2025-09-21T15:38:11.749Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Vanishing and Exploding Gradients Problems in Deep Learning]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@ojasmore143/vanishing-and-exploding-gradients-problems-in-deep-learning-fa24983b24c0?source=rss-05f56e0fa241------2"><img src="https://cdn-images-1.medium.com/max/1024/1*QwuvyCn9eqGYxPapc8hdyw.png" width="1024"></a></p><p class="medium-feed-snippet">Training a deep neural network sometimes feels like passing a message through a long chain of people. If the message keeps getting weaker&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@ojasmore143/vanishing-and-exploding-gradients-problems-in-deep-learning-fa24983b24c0?source=rss-05f56e0fa241------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@ojasmore143/vanishing-and-exploding-gradients-problems-in-deep-learning-fa24983b24c0?source=rss-05f56e0fa241------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa24983b24c0</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[gradient-descent]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[relu]]></category>
            <dc:creator><![CDATA[Ojasmore]]></dc:creator>
            <pubDate>Sat, 13 Sep 2025 16:47:36 GMT</pubDate>
            <atom:updated>2025-09-13T16:48:57.555Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ Exploring the AI Universe: Your Guide to Picking the Perfect AI Tool For your Uscase]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@ojasmore143/exploring-the-ai-universe-your-guide-to-picking-the-perfect-ai-tool-for-your-uscase-956c772f5349?source=rss-05f56e0fa241------2"><img src="https://cdn-images-1.medium.com/max/1024/1*LghUVelMmEGYeBv3duR8ZQ.png" width="1024"></a></p><p class="medium-feed-snippet">If you&#x2019;ve been scrolling online lately, you&#x2019;ve probably seen mentions of ChatGPT, Gemini, or maybe some jaw-dropping AI art on Instagram&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@ojasmore143/exploring-the-ai-universe-your-guide-to-picking-the-perfect-ai-tool-for-your-uscase-956c772f5349?source=rss-05f56e0fa241------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@ojasmore143/exploring-the-ai-universe-your-guide-to-picking-the-perfect-ai-tool-for-your-uscase-956c772f5349?source=rss-05f56e0fa241------2</link>
            <guid isPermaLink="false">https://medium.com/p/956c772f5349</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai-video-generator]]></category>
            <category><![CDATA[ai-music-generator]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ojasmore]]></dc:creator>
            <pubDate>Wed, 10 Sep 2025 16:59:08 GMT</pubDate>
            <atom:updated>2025-09-10T17:21:29.176Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[What is the XGBOOST MODEL? What techniques it use to prevent the OVERFITTING]]></title>
            <link>https://medium.com/@ojasmore143/what-is-the-xgboost-model-what-techniques-it-use-to-prevent-the-overfitting-d1ccfe2d2290?source=rss-05f56e0fa241------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1ccfe2d2290</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[gradient-boosting]]></category>
            <category><![CDATA[xgboost]]></category>
            <category><![CDATA[overfitting]]></category>
            <dc:creator><![CDATA[Ojasmore]]></dc:creator>
            <pubDate>Fri, 05 Sep 2025 17:23:28 GMT</pubDate>
            <atom:updated>2025-09-09T04:31:03.051Z</atom:updated>
            <content:encoded><![CDATA[<h3>What is the XGBOOST MODEL? What techniques does it uses to prevent OVERFITTING</h3><p>We will be covering maths behind it in a separate blog.</p><p>Before going through the answer, we should know what boosting models are (if you know this part, you can just skip it )</p><p>What are boosting models?</p><p>Ever thought about how it would go if we made a chain of robots in which each robot is correcting the error produced by its previous robot? We will have much less error in the output, or you can say very little residue(observed value — expected value)</p><p>The same thing is happening in the boosting model.</p><p>We combine the multiple weak learners to form a strong model. It is one of the ensemble methods (another one is bagging, eg: Random Forest)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*WhOsSNQb3mpcaAVmKCSyyw.png" /></figure><p><strong>(As the model gets more complex, variance grows, especially if we add too many learners without regularization.)</strong></p><p>If a person says some model is a boosting model, we should keep in mind that it has:</p><h4><strong>Sequential training</strong></h4><p>Models are trained one after another, and each model is learning from the mistakes of the previous one.</p><h4><strong>Weak learners</strong></h4><p>Each model utilizes the weak learners in most cases, and they will be the decision trees with limited depth.</p><h4><strong>Error Corrections</strong></h4><p>So the algorithm will identify the wrongly classified samples and will give more weights to them, so that the further model will give more attention to that complex instance.</p><h4><strong>Bias Reduction</strong></h4><p>Since we are iteratively reducing the error, the overall bias of the model is reduced.</p><h3>Now lets dive into the XGBOOST (Xtream gradient boost)</h3><p>Let’s learn what makes XGBoost less prone to overfitting than traditional gradient boosting models, and a very good accuracy model in complex datasets.</p><h4><strong>1. Regularization function</strong></h4><p>XGBOOST adds a regularization function in the traditional gradient boosting, L1(Lasso regularization), L2(Ridge regularization)</p><p>(A regularization function adds a penalty to the loss function so that we can get a more generalizable solution.)</p><p>So our objective function is</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/447/1*DXYuf2K-f8j9vMmZmvQUFw.png" /></figure><p>l = loss function ,</p><p>Ω = added penalty ,</p><p>γT = penalizes the number of leaves T (encouraging fewer, broader decisions)</p><p>λ (L2, Ridge)= It shrinks the w weight of the leaves towards zero so we can smooth out the extremes, and this will reduce the variance.</p><p>α (L1, Lasso) = It shrinks some weights exactly to zero this promotes the sparsity, automatic feature selections , and It prunes irrelevant features.</p><h4><strong>2. Shrinkage(learning rate)</strong></h4><p>We know that in boosting, each added new tree contributes to the answer. Each model overcorrects the previous model’s error without a learning rate; aggressive updates will lead to rapid overfitting.</p><p>learning rate η(typically 0.01–0.3)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*AC9FF9Cyva8QlwoUG1mE2A.png" /></figure><p>But why are we taking small steps? It’s simple if we take a large step, and that large step may perfectly fit the training residuals(overfitting).</p><p>While small ones allow the ensemble model to converge smoothly, this reduces the variance.</p><p>I’ll ask you a question. If η=1, how will it mimic traditional GBM?</p><h4>3.Column &amp; Row Subsampling</h4><p>Imagine each tree sees the same data and features; don’t you think they would learn the same pattern? Trees will become too similar, which can lead to overfitting.</p><p>So, XGBoost borrowed the idea of subsampling from random forest (bagging method)</p><p>ROW SUBSAMPLING</p><p>The idea is to take a part of the training data for each tree. So no single data point can dominate the learning (especially the noisy or majority class point ). This can help us with imbalanced data.</p><p>COLUMN SUBSAMPLING</p><p>Instead of giving all the features of the tree, we’ll give some features.</p><p><strong>colsample_bytree</strong> : random subset of features for the whole tree.</p><p><strong>colsample_bylevel</strong> : random subset per tree level.</p><p><strong>colsample_bynode</strong> : random subset for every split.(It adds fine grained randomness at splits, so it can be very helpful in enhancing exploration in deep trees)</p><h4>4.Tree Pruning</h4><p>When you look at the trees in boosting, they keep growing greedily, and splitting takes place as long as there is a chance of improvement. But think if this always necessary? The answer is, simply NO. Many times, the improvement is tiny, and the meaningless just fits the noise in it.</p><p>What XGBOOST does. It grows the tree level by level(depth-first). Build a balance tree. It prunes using a gain threshold (γ). In simpler words, it checks each split whether it is a useful split or not; if above the threshold, keep it, else, cut it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/469/1*3915KVs7W1-Q11_0G5ITxA.png" /></figure><p>G and H are gradient and Hessian sums.</p><p>Do you feel max_depth is better than this? Why?</p><h4>5.Early stopping</h4><p>Can we add trees forever? Obviously not, right? If we do so, the model will start fitting extra noise, which will eventually lead to overfitting.</p><p>If the score does not improve after certain rounds, we have to stop. early_stopping_rounds = 10–50, depending upon how big our dataset is.</p><h4>SO DOES IT MEAN XGBOOST CAN NEVER OVERFIT?</h4><p>❌</p><p>LET’S CLEAR IT IN NEXT BLOG ……….</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1ccfe2d2290" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>