gradient-vanishing

High-performance, branchless numerical stability kernels and compiler-optimized core infrastructure for advanced JAX/XLA deep learning architectures.

PyTorch experiments showing what happens when you remove ReLU from a deep network — loss curves, gradient collapse, depth sweep, decision boundaries, and activation comparison on MNIST.

A comparative experiment between RNN and LSTM models to evaluate their ability to perform noise-robust sequence prediction. The project tests short-term vs long-term memory by reconstructing clean input sequences from noisy data, showing how LSTM outperforms RNN under long-dependency conditions.