Neural Network from Scratch

Implemented neural-network layers as custom torch.autograd.Function classes — a fully-connected layer and a parameterized generalized-logistic activation — with hand-derived forward and backward passes.

Derived and coded the gradients with respect to inputs, weights, biases, and the activation's lower/upper/growth parameters, then validated every layer with gradient-checking unit tests.

Composed the layers into a trainable feed-forward network and evaluated it on the Iris classification task.