Nikunj Bhardwaj

ML | DL | Robotics | NLP | CV | IMUCET AIR-986 | NFSU MERIT RANK-121 | IIST RANK-844 | USAR-GGSIPU

[Nikunj Bhardwaj @knee_kunzz]

After a short nap-when i wake up, it’s a whole new szn. 🙌

Thank-you @X Family for supporting 🫶

@LinkedIn - https://t.co/YzHa8xppUR
@github - https://t.co/i2EvVTB9O0
@kaggle - https://t.co/oJSj4HKuZc
@LeetCode - https://t.co/vdIjIIqIlm
@TensorTonic -https://t.co/gEq77SC58V

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 365 of #ML365 #machinelearning
Computer vision Theory

Complete notes - https://t.co/L7NnG8OgIS

[Nikunj Bhardwaj @knee_kunzz]

6/6

6 days ago I started a "backprop from scratch" project.

Full notebook + theory + interview cheat sheet 👇

GitHub link - https://t.co/3Jadyyqc0w
Kaggle - https://t.co/aptzW1UCSY

[Nikunj Bhardwaj @knee_kunzz]

5/6
My favorite part of this project: looking inside the network.

These are the first-layer weights, reshaped as 28×28 images. Each hidden neuron became a little stroke or edge detector.

Nobody programmed them. They emerged from nothing but gradient descent on 60,000 digits.

[Nikunj Bhardwaj @knee_kunzz]

4/6
SGD vs Adam, both written from scratch. Same seed, same init, same batch order
SGD: 97.65% (10 s)
Adam: 97.55% (32 s)
Adam hit 97% two epochs sooner (epoch 3 vs 5) and reached 3× lower training loss. But the test accuracy is a tie. The 0.1% gap is about 10 images out of 10K

[Nikunj Bhardwaj @knee_kunzz]

3/6
How do I know my hand-written gradients are right?
I don't trust myself.
W1 → 2.5e-11 ✅
b1 → 2.8e-11 ✅
W2 → 2.2e-11 ✅
b2 → 1.8e-11 ✅
broke it on purpose: forgot to divide by batch size. Error jumped to 0.78 ❌. Caught instantly.
ML bugs fail silently -build detectors

[Nikunj Bhardwaj @knee_kunzz]

2/6
The entire backward pass of a neural-network layer:

dW = Xᵀ @ dY
db = dY.sum(axis=0)
dX = dY @ Wᵀ
My favorite debugging trick is the shape check. Xᵀ·dY is (in,N)(N,out) = (in,out), exactly W's shape. Wrong shape means wrong derivation

[Nikunj Bhardwaj @knee_kunzz]

1/6
I trained a neural network on MNIST with no PyTorch and no autograd.

Every gradient: derived by hand, coded in NumPy.
Result: 97.65% test accuracy.

[Nikunj Bhardwaj @knee_kunzz]

Hi guys Back here ! -- will be posting my progress with some new projects ,,, First I will start with some warm-up projects and then moving to a Production-Grade End-to-End Capstone projects , See you soon guys !!!!!!!!

[Nikunj Bhardwaj @knee_kunzz]

Hi @X ppl!!!!

Officially closing my Automation & PM intern at @JBMAuto Group ⚙️📈 Lowkey learned a ton, highkey cooked on some massive projects.

On to the next quest! 🚀

#JBMGroup #Internship #Automation #ProjectManagement

[Nikunj Bhardwaj @knee_kunzz]

Hi @X Family

@Flipkart GRiD 8.0 update: cleared R1 ✅, R2 didn't go my way 💀
1 Q solved clean, rest got wrecked by compilation errors + bad time management
Lesson learnt, prep mode: ON for next time 🔥

#GRiD8 #flipkart

[Nikunj Bhardwaj @knee_kunzz]

loss function went down, but so did my sleep schedule 💀 appreciate the love bro 🤝🙌 - @yur_o_0

[𝐲𝐮𝐫𝐨 @yur_o_0]

respect to Nikunj.

bro was consistent for an entire year.
365/365.

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 365 of #ML365 #machinelearning
Computer vision Theory

Complete notes - https://t.co/L7NnG8OgIS

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 364 of #ML365 #machinelearning
Computer vision Theory - 6/6

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 363 of #ML365 #machinelearning
Computer vision Theory - 5/6

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 362 of #ML365 #machinelearning
Computer vision Theory - 4/6

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 361 of #ML365 #machinelearning
Computer vision Theory - 3/6

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 360 of #ML365 #machinelearning
Computer vision Theory - 2/6

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 359 of #ML365 #machinelearning
Computer vision Theory - 1/6

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 358 of #ML365 #machinelearning
Testing Yolov11 Instance Segmentation

[Nikunj Bhardwaj @knee_kunzz]

🚀Day 357 of #ML365 #machinelearning
Training Yolov11 Instance Segmentation