Deep dive into activations, gradients, and BatchNorm (makemore pt 3). Visualized activation distributions, gradient flow, and dead neurons. Learned why BatchNorm stabilizes training and helps deep nets.
#DL#AI#NNDay
5/6: https://t.co/xbjV3CW76n
Upgraded bigram to MLP (makemore pt 2) in PyTorch. Added embeddings, hidden layers, activations, and softmax. Learned: NLL loss, dev splits, overfitting, hyperparameter tuning,better initialization. Names look way more realistic now! #DL#AI#NN
Day 4/6: https://t.co/xbjV3CW76n
Built a simple bigram-level language model (makemore pt 1). Learned the basics: NLL loss, sampling, and the language modeling framework. Hilarious outputs!
Next: MLP upgrade!
#DL#AI#NN
Day 3/6: https://t.co/lzuPyYdnLU
Rebuilt micrograd concepts in PyTorch – tensors, vectorized ops, efficient gradients. Same math, but scalable and fast. No more scalar loops!
Bridge from toy engine to real frameworks (makemore, PyTorch)
Tomorrow: Bigram model! #DL#AI#NN
Day 2/6: https://t.co/qurhHGWPTn
"NN: Zero to Hero" 🧠
Built micrograd a tiny autograd engine.
Understood forward pass, backprop, and gradient flow.
Thanks Andrej!
Video: https://t.co/QbzCALw2h7
#DeepLearning#AI#NeuralNetworks
Day 1/6:
https://t.co/vrmEjYIWyt
Over the next 6 days, I'm sharing my 2-month journey as a beginner diving into neural networks through Andrej Karpathy's excellent "Neural Networks: Zero to Hero" series. No prior deep learning knowledge—just pure curiosity!
Link to the lectures: https://t.co/QNiF66S0pi
Finally solved the Nvidia driver nightmare on EndeavourOS with dwm! Now my laptop connects to the external monitor seamlessly—no more black screens.
Next up: customizing my own desktop environment for max productivity. #Nvidia#EndeavourOS#dwm#ArchLinux#LinuxCustomization
🧠 Just wrapped up an incredibly insightful Vara
Thanks @AdnanDalal96069
and @0xZeeast Web3 Workshop! Learned so much about building on the network, from the Actor Model to autonomous programs. Excited to put this knowledge into action. Thanks, @VaraNetwork!
#VaraNetwork#Web3