sekwiatkowski/komputation

Supports flexible layer composition across feed-forward, convolutional, and recurrent architectures with dual CPU and GPU execution paths—both powered by automatic differentiation for backpropagation. Provides multiple initialization schemes (Gaussian, He, uniform) and optimizers (SGD, momentum, NAG) for training, alongside standard loss functions and activation layers. Available via Maven Central and demonstrated across tasks ranging from boolean functions to MNIST and text classification on both JVM and CUDA backends.

293 stars. No commits in the last 6 months.