Fashion is a visual medium so it makes sense for our models of fashion to include visual features. In this presentation, I'll describe how we've build a general purpose visual fashion representation using CNNs. The network is multi-task (multiple labels per image), multi-image (multiple images per label) and it runs on multiple GPUs. We used the python library Chainer to fit the network.

I'll visually explore what is going on inside the black box of a neural network and discover how a fashion specific model sees the world differently from generic visual models. Lastly, I'll demonstrate some applications of the representation learned by the model.

The initial part of this presentation will be technical but the remainder will be accessible.