How I built a deep learning library in pure Java (and compared it with PyTorch)

1. Why: learning Java and wanting to understand AI under the hood
2. Tensors and matmul first, then gradient checks for every layer
3. Building Conv2D, LSTM and a mini transformer
4. Making it small: int8 and int4 quantization
5. Testing: 283 tests, including autograd checks
6. The PyTorch comparison: same CNNs, same epochs, accuracy within about 1 point
7. What I learned and what's next Links: https://github.com/manishn1877/minidl-java/tree/main and https://manishn1877.github.io/minidl-java/