Analyzing 50k fonts using deep neural networks

A personal project scraping 50k fonts, converting them to 64x64 bitmaps, and training a deep neural network (using Lasagne/Theano) to learn a 40-dimensional latent 'font vector' space. The model can reconstruct unseen font characters, interpolate smoothly between fonts, generate new synthetic fonts by sampling the latent space, and visualize font similarity via t-SNE. Key architectural details include 4 fully connected layers of width 1024, L1 loss, leaky ReLU activations, and Gaussian noise regularization on font vectors.