Designed a neural network from scratch with Octave GNU, capable of detecting handwritten numerical digits from 0-9 with 99% accuracy. The network is trained with 5000 examples from the MNIST data set. Cost function and gradient are calculated to minimize error of the network’s output.