The Thing About Machine Learning — It Teaches You Right Back

By Shrikumar Nilawar B.Tech, Artificial Intelligence | GH Raisoni College of Engineering & Management, Nagpur | Batch 2025

— -

I Chose AI. And Then AI Chose to Humble Me.

When I joined the AI branch at GH Raisoni, I’ll be honest — I thought it sounded cool. Artificial Intelligence. Machine Learning. Neural Networks. These were words that felt like the future, and I wanted to be part of that future.

What I didn’t expect was how much the field would make me question everything I thought I knew about problem-solving, about data, and about what it even means to “teach” something. Machine Learning, I quickly realised, isn’t about making computers smarter. It’s about being precise enough with your questions that the data can answer them.

You don’t program a machine to think. You design the conditions under which it can learn to think. That shift in perspective — from coder to teacher — changed how I approached every problem after that.

This post is my attempt to document that shift — the early confusion, the weeks of incremental progress, the real mistakes I made, and the moments where something finally clicked. If you’re somewhere in that journey yourself, I hope this feels familiar.

What is Machine Learning, really? In plain terms: instead of writing explicit rules for every scenario, you show the algorithm thousands of examples and let it figure out the rules itself. It’s pattern recognition at scale — and once you truly get that, everything else starts to make sense.

— -

Four Ideas That Rewired How I Think

Everyone wants to jump straight to deep learning and neural networks. I was guilty of that too. But the concepts that actually changed my thinking were more fundamental. Here are the four that mattered most during my time at Raisoni.

Supervised Learning You feed the model labeled data — input paired with the correct output — and it learns the mapping between them. Think of it like studying with an answer key. It works surprisingly well when your data is clean and well-structured. Most real-world ML problems fall into this category.

Unsupervised Learning No labels, no guide. The algorithm finds its own structure in the data. The first time I ran K-Means clustering on a customer dataset and it surfaced groups I hadn’t even thought to look for, I was genuinely amazed. The model knew something I didn’t.

Reinforcement Learning Learn by doing — reward good actions, penalise bad ones. The same logic behind how a child learns to walk. The same logic behind how AlphaGo learned to beat world champions at Go. It’s the most intuitive type of ML conceptually, and the hardest to implement well in practice.

The Bias–Variance Tradeoff Too simple a model and it misses real patterns (high bias — underfitting). Too complex and it memorises noise instead of signal (high variance — overfitting). Navigating this tradeoff is, I’d argue, the central practical challenge of machine learning. And you don’t really understand it until you’ve been burned by it.

The Bias–Variance tradeoff took me weeks to feel in practice. I kept building models that hit 98% accuracy on training data and collapsed to 61% on new data. That’s overfitting — and it’s the ML equivalent of memorising last year’s exam paper rather than understanding the subject. Painful lesson. Very effective teacher.

Here’s a quick look at the algorithms I tried and what happened:

Linear Regression — Supervised — Housing price prediction — Worked well Logistic Regression — Supervised — Spam email classification — Worked well Decision Tree — Supervised — Iris flower classification — Clean results K-Means — Unsupervised — Customer segmentation — Partially worked Random Forest — Supervised — Titanic survival prediction — Best result

— -

The Honest Account of How It Actually Went

I want to be real here. The first week I understood maybe 40% of what was happening. By the fourth week, that had become something closer to 75% — and the remaining 25% felt like the interesting part, not the scary part anymore. That’s progress worth marking.

Week 1 — Python, Pandas, and the Rude Awakening

I thought I knew Python. Pandas DataFrames exposed exactly how wrong I was. I spent three hours on a single preprocessing step — just cleaning one column of messy data. But by the end of the week I could wrangle a CSV without looking anything up. That small win kept me going.

Week 2 — First Real Model: 8 Lines of Code, Infinite Questions

Built my first Linear Regression model. The code took five minutes. Understanding gradient descent — why the algorithm walks downhill on a cost function to minimise error — took three days and two full rewatches of a 3Blue1Brown video. Worth every second of it.

Week 3 — The Titanic Dataset and the Power of Feature Engineering

This was the turning point for me. I started creating new columns from existing ones — a “family size” feature, a “title” feature extracted from passenger names. My accuracy jumped from 73% to 82% without changing the algorithm at all. That experience taught me something I still believe: good features beat good algorithms. Almost every time.

Week 4 — The 95% Accuracy Lie

I had a fraud detection model scoring 95% accuracy. I was proud of it. Then I realised that 95% of the transactions in the dataset were non-fraudulent — my model was simply predicting “no fraud” every single time and still hitting 95%. It had learned nothing.

That’s when Precision, Recall, and F1-Score entered my life permanently. Accuracy on imbalanced data is a comfortable lie. The confusion matrix is where the truth lives.

— -

Five Mistakes I Made — Written Down So I Never Forget Them

The blogs I trusted most during college were always honest about failure. So here’s my version of that honesty. These aren’t hypothetical mistakes. They’re ones I personally made, some more than once.

Mistake 1 — Data Leakage

I accidentally included future information in my training data. The model hit 99% accuracy and I genuinely thought I was doing something right. Then I learned what data leakage actually means. In real production systems, this mistake ends careers. Now I split my data first — before touching anything else — every single time.

Mistake 2 — Not Scaling Features

If one feature ranges from 0 to 10,000 and another from 0 to 1, distance-based algorithms like KNN and SVM get completely confused by the scale difference. StandardScaler is literally two lines of code. Not using it cost me an entire week of confusion and bad results. Small step, huge impact.

Mistake 3 — Trusting Accuracy on Imbalanced Data

Already mentioned above, but worth repeating here plainly: on any imbalanced dataset — medical diagnosis, fraud detection, rare events of any kind — accuracy tells you almost nothing useful. Always look at precision, recall, and F1-score. Always pull up the confusion matrix. No exceptions.

Mistake 4 — Reaching for Deep Learning Before Earning It

In my third year I wanted to build neural networks before I properly understood logistic regression. That’s like wanting to write novels before learning grammar. A well-tuned Random Forest kept beating my early neural networks in every experiment. The lesson: start with the simplest model that could possibly work. Add complexity only when simplicity genuinely fails you.

Mistake 5 — Underestimating Data Quality

I once spent three full days tuning a model — reading papers, adjusting hyperparameters, trying different architectures — only to eventually discover that the dataset had duplicate rows inflating my evaluation scores. “Garbage in, garbage out” is a cliché for a reason. Cleaning data is 70–80% of real ML work. No algorithm in the world compensates for a fundamentally bad dataset.

— -

The Moment It Stopped Being Abstract

There was a particular moment in my third year when I realised I could trace every ML concept I’d studied directly into apps I used every single day. That connection — between a lecture topic and a live product — made everything feel real in a way that textbooks alone never had.

Spotify’s Discover Weekly → Collaborative Filtering When Spotify recommends a song you’ve never heard but somehow love — that’s your listening history being compared to thousands of similar users to find patterns. K-Nearest Neighbours logic, running at enormous scale, inside your earphones every Monday morning.

Gmail Spam Filter → Naive Bayes Classifier Classic text classification. Certain words and patterns are statistically associated with spam. Naive Bayes is simple, fast, and surprisingly effective — a reminder that elegant, well-understood algorithms often outperform unnecessarily complex ones.

Bank Fraud Alerts → Anomaly Detection Your bank flagging an unusual transaction is ML noticing statistical deviation from your established behavioural pattern. The imbalanced data problem I struggled with in class? This is exactly where it lives in production — and why it matters so much to get right.

Medical Imaging Diagnosis → CNNs Convolutional Neural Networks trained on hundreds of thousands of X-rays can detect anomalies with remarkable accuracy. This application also made me think seriously about responsibility. A biased training set here doesn’t just produce a bad metric — it affects real patients. That reality sat with me.

Self-Driving Systems → RL + Computer Vision Reinforcement Learning determining when to brake. Computer Vision reading lane markings and obstacles. Two concepts from the same semester, working together in a system navigating real roads. When that connection clicked, it hit differently.

The moment that made it personal for me: I built a sentiment analysis model in college that could tell whether a movie review was positive or negative. Watching a machine read actual human words and correctly understand the emotion behind them — that felt like witnessing something quietly remarkable. That project is still the one I talk about most when people ask what I’ve built.

— -

What Four Years in AI Actually Taught Me

It’s strange to sit here after graduating and look back at the person who walked into GH Raisoni in the AI branch thinking he understood what that meant. I didn’t. But I think the not-knowing was exactly the right place to start.

Machine Learning is one of those fields that keeps revealing new depths the further you go. That used to feel intimidating. Now it feels like the best reason to keep going.

Here’s what I would tell myself at the very beginning:

Break things on purpose. The best learning I did came from deliberately writing bad code and watching it fail in instructive ways. A model that fails loudly and clearly teaches you more than one that quietly works by luck. Break things. Read the error. Understand why.

Understand the math, even loosely. You don’t need to derive gradient descent from first principles to be effective. But knowing why an algorithm moves the way it does — what it’s actually optimising for, what it’s trying to minimise — changes how you use it, debug it, and trust it.

Good data beats good algorithms. I spent months chasing better models when the data itself was the problem. A simple, well-tuned model on clean, well-understood data will almost always outperform a sophisticated model on messy data. This is one of the most important things I know, and I learned it slowly.

This field has real consequences. A biased model in healthcare, lending, or hiring doesn’t just produce a bad accuracy number — it affects real people’s lives and opportunities. Choosing AI as a field means choosing that responsibility too. I take it seriously. I think everyone working in this space should.

The confusion is not a sign you’re failing. Every person who is genuinely good at ML was completely lost at some point. The ones who became good at it were just more patient with themselves than the ones who left. If you’re confused right now, that’s the process working correctly. Keep going.

I’m not done learning — I don’t think I’ll ever be done learning, which is honestly part of what I love about this field. But I’m graduating with something more valuable than a list of algorithms: I know how to think about data, how to ask the right questions of it, and how to be honest about what a model is and isn’t telling me.

That, I think, is the real output of four years in AI.

— -

Resources I’d Genuinely Recommend

These aren’t affiliate links. They’re just the things that moved the needle for me during college — the ones I kept coming back to when a concept wouldn’t click from a textbook alone.

• Machine Learning Mastery (machinelearningmastery.com) — My most-used resource. Practical, clear, no fluff.
• Kaggle Learn — Free, hands-on, structured. The best place to actually build things.
• StatQuest with Josh Starmer on YouTube — Makes statistics feel approachable and even enjoyable.
• 3Blue1Brown — Essence of Linear Algebra and Neural Networks series. Watch these.
• Andrew Ng’s Machine Learning Specialisation on Coursera — The foundational course. Still unmatched.
• fast.ai — Practical Deep Learning for Coders. Top-down approach that works brilliantly.
• Hands-On Machine Learning by Aurélien Géron — The book I’d recommend above all others.
• Towards Data Science on Medium — Excellent community writing across all levels.
• Analytics Vidhya — Great for Indian ML community content and competitions.
• Google ML Crash Course — Free, structured, very well made.

If I had to pick one single starting point for someone brand new to ML: Kaggle’s Titanic competition. It’s a rite of passage for a reason. Small dataset, clear problem, massive community, immediate feedback loop. Everyone starts with Titanic. And there’s no shame in that — I learned more from that one dataset than from three months of reading alone.

— -

If you’ve read this far — thank you. Writing this was unexpectedly useful, a way of processing everything I absorbed over four years at Raisoni and putting it somewhere outside my own head. I hope something in here was useful to you, or at least honest enough to feel real.

ML is hard. It’s also endlessly interesting. I’d choose it all over again.