<?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 Shalini Kumari on Medium]]></title>
        <description><![CDATA[Stories by Shalini Kumari on Medium]]></description>
        <link>https://medium.com/@shalini.k_66195?source=rss-a80eaa968c6a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*EBevvXrIxQlgjIvV7Fd4CQ.jpeg</url>
            <title>Stories by Shalini Kumari on Medium</title>
            <link>https://medium.com/@shalini.k_66195?source=rss-a80eaa968c6a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 16:28:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shalini.k_66195/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[Why AI Doesn’t Have to Start from Zero]]></title>
            <link>https://medium.com/@shalini.k_66195/why-ai-doesnt-have-to-start-from-zero-fc4b7717d93a?source=rss-a80eaa968c6a------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc4b7717d93a</guid>
            <dc:creator><![CDATA[Shalini Kumari]]></dc:creator>
            <pubDate>Tue, 13 Oct 2026 09:58:07 GMT</pubDate>
            <atom:updated>2026-10-13T09:58:07.874Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jAmR2AvJ1r2-H2c8t9uKNw.jpeg" /></figure><p>A beginner’s guide to transfer learning, explained through handwriting research.</p><p>Early in my research, I hit a problem. To teach an AI system to recognise a writer’s style, you need handwriting samples, and lots of them. But handwriting datasets are small. Collecting pages from hundreds of people, with their consent, takes months.</p><p>Meanwhile, the most powerful AI models learn from <strong>millions</strong> of images. So how do you get good results from a few thousand pages of handwriting?</p><p>The answer is a simple idea with an intimidating name: <strong>transfer learning.</strong> In this article I explain it without a single equation, using everyday comparisons and one handwriting example.</p><p><strong>THE SHORT VERSION</strong></p><blockquote>Transfer learning means starting from a model that has already learned from a huge amount of data, then adapting it to your smaller task. It saves time and data, and it often beats training from scratch, though it needs some care</blockquote><h3>The idea you already know: skills carry over</h3><p>Picture two people stepping onto a tennis court for the first time. One is a complete beginner. The other is a cricketer.</p><p>The beginner has to learn everything at once: how to hold the racket, how to time the ball, how to move their feet. The cricketer already has hand-eye coordination, timing and footwork. They aren’t a tennis player yet, but they start far ahead.</p><blockquote><em>Transfer learning is the cricketer walking onto the tennis court.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*7f8faO05lMJ9_kWCjF9faA.png" /></figure><p>AI works the same way. Instead of teaching a model everything from zero, we start with one that already knows useful things.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/658/1*OTtZgDD_ZjkNKa4kQjRAZg.png" /></figure><p><strong>What has the AI already learned?</strong></p><p>Many popular image models are first trained on millions of ordinary photos: animals, vehicles, food, buildings. Handwriting isn’t in there at all. Yet something useful happens inside them.</p><p>These models learn in layers, from simple to complex. The first layers pick up basic ingredients like edges, lines and curves. The middle layers combine those into shapes and textures. Only the last layers specialise in the old task, such as telling a cat from a car.</p><p>Many popular image models are first trained on millions of ordinary photos: animals, vehicles, food, buildings. Handwriting isn’t in there at all. Yet something useful happens inside them.</p><p>These models learn in layers, from simple to complex. The first layers pick up basic ingredients like edges, lines and curves. The middle layers combine those into shapes and textures. Only the last layers specialise in the old task, such as telling a cat from a car.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*prRU2Z5fRy_yZUP55ZJCfA.png" /></figure><p>Here’s the interesting part. <strong>Edges, curves and textures are useful for almost any image, including a page of handwriting.</strong> A pen stroke is, after all, a curved line. So the early and middle layers can be reused, and the last layer is swapped for one built for the new task.</p><h3>Two ways to reuse a model</h3><p>There are two common ways to put a pretrained model to work.</p><h3>Option 1: Freeze it</h3><p>Keep the pretrained model exactly as it is and use it as a <strong>feature extractor</strong>. It looks at a handwriting image and produces a list of numbers describing it. You then train only a small, simple classifier on top of those numbers. This works well when you have very little data.</p><h3>Option 2: Fine-tune it</h3><p>Start from the pretrained model and keep training it on your handwriting images, gently and with small adjustments, so it doesn’t forget what it already knows. You can fine-tune just the last few layers or the whole model. This usually works better when you have a moderate amount of data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*ucw-0wplsFTBFpn4ttYbwQ.png" /></figure><p>A simple rule of thumb: <strong>less data, freeze more; more data, fine-tune more.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/423/1*Aymy5QPivYw1alzmKkQIiQ.png" /></figure><h3>Where it fits in writer identification</h3><p>Put these ideas together and you get a typical pipeline for handwriting research. A pretrained network supplies the starting knowledge. You fine-tune it on handwriting images so it pays attention to style rather than objects. It then produces features, which are numbers that describe how a person writes, and those features are compared across samples to identify or verify a writer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*8GtwmDbCS-S08ADfRl-Diw.png" /></figure><p>Writer identification suits transfer learning well, because datasets usually have a limited number of writers and only a few pages per writer. That is exactly where training from scratch struggles.</p><h3>When can it go wrong?</h3><p>Transfer learning is powerful, but it isn’t magic. Four problems come up often.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*8qDyjzOyTeN3NIPnMwOSWw.png" /></figure><p>• <strong>Domain gap.</strong> Photos and handwriting look very different, so some of what the model learned helps less than we’d hope.</p><p>• <strong>Overfitting.</strong> Fine-tuning a large model on a tiny dataset can make it memorise the training pages instead of learning general patterns.</p><p>• <strong>Data leakage.</strong> If pages from the same writer appear in both training and testing, the results look better than they really are. Good research tests on writers the model has never seen.</p><p>• <strong>Negative transfer.</strong> Occasionally the old knowledge actively hurts the new task.</p><p><em>A pretrained model is a head start, not a guarantee.</em></p><p>This is why honest evaluation matters here as much as anywhere else in AI research.</p><h3>Quick answers to common questions</h3><h3>Is transfer learning cheating?</h3><p>No. It’s how most modern AI is built. Starting from existing knowledge is standard practice, as long as researchers say what they started from and test fairly.</p><h3>Do I need an expensive computer?</h3><p>Far less than training from scratch. Freezing a model and training a small classifier can run on an ordinary laptop. Fine-tuning large models is faster with a GPU.</p><h3>Is it only for images?</h3><p>Not at all. Language models, speech systems and many other kinds of AI use the same idea: learn broadly first, then adapt.</p><h3>Does the old task need to be similar?</h3><p>The closer the two tasks, the better the transfer. But even fairly different tasks can share useful basics, like edges and strokes.</p><h3>So, does AI have to start from zero?</h3><p>Usually, no, and that’s good news for anyone working with limited data.</p><p>What I find most interesting is how human the idea feels. We rarely learn anything from nothing. We bring what we already know and adapt it. Teaching machines to do the same has become one of the most useful ideas in modern AI.</p><p><strong>KEY TAKEAWAYS</strong></p><p>✓ <strong>Transfer learning</strong> means reusing what a model learned on one task to help with another.</p><p>✓ Early layers learn general things like edges and curves, which are useful almost everywhere.</p><p>✓ <strong>Freeze</strong> the model when data is very small; <strong>fine-tune</strong> it when you have more.</p><p>✓ It suits writer identification because handwriting datasets are limited.</p><p>✓ Watch for domain gap, overfitting, data leakage and negative transfer.</p><p>✓ A pretrained model is a head start, not a guarantee.</p><p>• • •</p><p><em>This is the second article in my series on AI and handwriting research. If you missed the first one, “Can a Computer Really Tell Who Wrote This?”, it is a good place to start.</em></p><p><strong>Your turn:</strong> what’s a skill you learned faster because of something you already knew? Tell me in the responses below.</p><p>Tags: Artificial Intelligence · Machine Learning · Transfer Learning · Computer Vision · Handwriting Analysis</p><p><strong>Shalini Kumari</strong> is an Assistant Professor in Computer Science and Engineering at CMR University, Bengaluru, and is pursuing doctoral research in the same field. Her interests include AI, machine learning, deep learning, computer vision and</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc4b7717d93a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Can a Computer Really Tell Who Wrote This?]]></title>
            <link>https://medium.com/@shalini.k_66195/can-a-computer-really-tell-who-wrote-this-932c8fdc1bf0?source=rss-a80eaa968c6a------2</link>
            <guid isPermaLink="false">https://medium.com/p/932c8fdc1bf0</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Shalini Kumari]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 07:32:51 GMT</pubDate>
            <atom:updated>2026-09-29T07:32:51.311Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j9ldVU94GpAv9ixdPprqbA.png" /></figure><p>A beginner’s guide to how AI recognises handwriting — no maths, no jargon, just the big idea.</p><blockquote>When I first picked my PhD topic, one question kept nagging me: <em>can a computer really tell who wrote something just by looking at the handwriting?</em></blockquote><blockquote>Honestly, I had no idea. So I started reading. One paper led to another, and soon I was deep into how computers describe handwriting, how machine-learning systems learn patterns, and how researchers test whether any of it actually works.</blockquote><p>This is the explanation I wish someone had handed me on day one. No equations. No heavy terminology. Just the big picture.</p><p><strong>THE SHORT VERSION</strong></p><blockquote>Everyone writes a little differently. Computers can learn those differences from examples and use them to estimate who wrote a page. It works surprisingly well — but it isn’t perfect, and in serious settings it’s best used to <strong>support human experts, not replace them</strong></blockquote><h3>Two questions that sound almost the same</h3><p>Researchers talk about <strong>writer identification</strong> and <strong>writer verification</strong>. They sound alike, but they solve different problems.</p><h3>1. Identification: “Who wrote this?”</h3><p>Picture an unknown handwritten page and a pool of known writers. The computer compares the page with what it has learned about each writer and picks the closest match. Think of a detective comparing a clue against a list of suspects.</p><h3>2. Verification: “Did this person write this?”</h3><p>Now you already have a claim: <em>“This is Ravi’s handwriting.”</em> The system compares the questioned sample with Ravi’s known writing and estimates whether they’re similar enough. It’s less like a detective and more like a bouncer checking ID — a yes or no, within an error threshold.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LtzNuf-_WgHKGmPxve0tAA.png" /></figure><h3>Your handwriting is a behavioural fingerprint</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*gyVxM36FHamqD1idFNmUIg.png" /></figure><blockquote>Think about how you write. You never decided how much to slant each letter, how wide to space your words, or how hard to press the pen. Those habits formed quietly, over years.</blockquote><p><em>Two people can learn the same alphabet, use the same pen, write the same sentence — and still produce noticeably different writing</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*nkEg-pQBzg_IjVrFQNzsNw.png" /></figure><p>A computer can’t look at handwriting with human intuition. Instead, it turns visual patterns into measurable information: shape, spacing, stroke structure, direction, texture and more.</p><p>The core idea is simple: <strong>handwriting contains repeated personal patterns, and AI tries to learn them from examples</strong></p><h3>How the field has evolved</h3><p>People have compared handwriting for a very long time. What’s changed is <em>who — or what — does the comparing</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*7eQ-ZTn6vAeF1mRe0ZvERA.png" /></figure><h3>How does a computer actually do it?</h3><p>I remember the process in three words: <strong>look, learn, decide.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/423/1*zVjq_mXSJioO5sa7dfw67g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/992/1*v2fAQILoPV89PzvCFDVKbg.png" /></figure><h3>How do researchers know it works?</h3><p>Anyone can claim an AI system works. Research needs a fair way to measure it.</p><p>In verification, a system can fail in two ways: it can <strong>reject a genuine writer</strong>, or it can <strong>accept the wrong one</strong>. These mistakes don’t carry the same consequences. Having your own signature rejected is annoying. Having a forged signature accepted could be far more serious.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*UtuLMID-B_1Gl_txaihWvg.png" /></figure><p><em>One headline accuracy number is never the whole story.</em></p><p>Good research explains what data was used, how the evaluation was run, and what kinds of errors occurred. It also asks whether the test data looks like the real world — clean, carefully prepared samples are much easier than rushed, messy, everyday handwriting.</p><h3>Where could this be useful?</h3><p>These are <em>possible</em> uses — not a claim that AI alone should make high-stakes decisions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*p_G6GkQ3-PSkjF7Q_NbBuA.png" /></figure><h3>What’s still hard?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*UudD3XYu-oFwdu6uBvDnmw.png" /></figure><h3>Humans and computers work best together</h3><p>The more I read, the more this looks like a partnership, not a replacement story.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*V8KWATUoL8w4tDfiQsYJTQ.png" /></figure><h3>Quick answers to common questions</h3><h3>Can AI read what’s written?</h3><p>Not necessarily. Recognising <em>what</em> the words say is a different task from recognising <em>who</em> wrote them. A system can focus on the writer without understanding the sentence.</p><h3>Does it work on signatures?</h3><p>Yes — signature checking is a common form of verification. But signatures are short, so they carry less information than a longer passage.</p><h3>Is it 100% accurate?</h3><p>No. No recognition system is. That’s why honest error measurement, realistic testing and transparent reporting matter.</p><h3>Can AI replace a handwriting expert?</h3><p>Think of it as decision support. AI can offer useful evidence and comparisons; human experts provide context and judgement.</p><h3>So, can a computer tell who wrote it?</h3><p>Back to the question that started my journey. The answer: <strong>often, yes — but it’s not magic.</strong></p><p>Performance depends on the data, the method, the evaluation, and the conditions under which the handwriting was produced. For me, the most interesting part isn’t getting a machine to recognise handwriting. It’s understanding <em>why</em> it works, <em>where</em> it fails, and <em>how</em> it can be used responsibly.</p><p>That’s where handwriting research becomes much more than a classification problem.</p><p><strong>KEY TAKEAWAYS</strong></p><p>✓ <strong>Identification</strong> asks “Who wrote this?”; <strong>verification</strong> asks “Did this person write this?”</p><p>✓ Handwriting holds small behavioural patterns that help tell writers apart.</p><p>✓ AI, in short: <strong>look, learn, decide.</strong></p><p>✓ Good evaluation reports errors honestly and tests on realistic data.</p><p>✓ Limited data, forgeries, changing handwriting, different scripts and explainability remain open challenges.</p><p>✓ Human expertise and AI complement each other.</p><p>• • •</p><blockquote><em>I’m still learning, and this article is part of that journey. I hope it made the subject a little less intimidating — and encourages you to ask your own questions about what AI can really see in handwriting.</em></blockquote><p><strong>Your turn:</strong> what’s the first question that comes to your mind when you think about handwriting and AI? Tell me in the responses below.</p><p>Tags: Artificial Intelligence · Machine Learning · Computer Vision · Handwriting Analysis · Research</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=932c8fdc1bf0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>