239 — Transductive Adversarial Networks (TAN)

Sean Rowan (1802.02798)

Read on 16 April 2018
#neural-network  #adversarial  #TAN  #unlabelled-data  #MNIST  #SVHN  #drug-interaction 

Oh man — my dude Sean Rowan is killing it.

In this paper, he demonstrates a novel adversarial network architecture called a Transductive Adversarial Network, or a TAN.

TANs address the common problem of having large amounts of in-domain unlabelled data alongside labelled but out-of-domain data. The canonical example is that of the MNIST digits dataset alongside the Google Street View House Numbers (SVHN) dataset: Using the MNIST image inputs and labels, it is possible to “link” the MNIST vectorization to a SVHN vectorization in order to classify SVHN inputs.

This toy example elucidates more nuanced uses of the TAN architecture: Rowan gives another example of drug interactions. Naturally, arbitrary human drug-rug interactions are difficult (or at least immoral) to test in-vivo. But by measuring drug interactions in a yeast culture and cross-training this on the human drug interactions, it is possible (theoretically) to improve human drug trials without more human pharmaceutical tests.

TAN differs from the preexisting $\Delta$-GAN architecture in one fundamental way: Whereas $\Delta$-GANs require matching inputs and labels in both source and target domains, TANS require only source domain labels alongside a marginalized label/vector distribution in the target domain. This makes TANs a more viable solution for more restrictive tasks (such as human drug trials).

Oh and also Sean rocks and all of the code is available open-source on GitHub.

I am very seriously geeking out about this research.