Learning to Transfer with Triply Adversarial Nets

Work in progress, for a tentative paper submission to NIPS'16. Deadline: May 20.

Abstract (tentative)

In classification, transfer learning (or its variants known as co-variate shift or domain adaptation) arises whenever target instances are governed by a distribution that may be arbitrarily different from the distribution of the source instances used at training. This problem has traditionally been solved by re-weighting approaches or by learning robust representations over domains. In this work, we propose a new paradigm based on the assumption that the co-variate shift is only due to a different representation of the same underlying objects. Accordingly, we propose to learn how to transform source instances into target instances, possibly across input spaces of distinct dimensions, structures or supports. For this purpose, we extend the generative adversarial networks framework of \cite{goodfellow2014generative} to a triply adversarial process: a transformer network $G$ for generating target instances from source instances, a discriminative network $D$ for separating transformed source instances from actual target instances, and a classifier network $C \circ G$ for classifying source instances in the projected space. This 3-player game results in a network $G$ capable of transforming source into target instances, while preserving separation between classes as enabled by $C$ in the adversarial setup. Preliminary experiments demonstrate the potential of this novel approach, with promising results when the construction of $C$ can be bootstrapped in a semi-supervised way from a few labeled instances from the target space.

Help is welcome! #openscience

The writing of this paper is open source and welcome contributions!

• Why? Despite being myself a machine learning researcher, this paper is my first serious attempt at a contribution in deep learning. While I am convinced the proposed idea to be worth it, I also firmly believe that it could be strengthened, matured and improved with the help of experienced fellow researchers from the field.

  I would also rather see this idea materialize than die in some random page of my notebook :-)

• Don't you fear being scooped? This is indeed a risk I am willing to take. Yet, given the general openness of the ML and AI community, I bet that the overall outcome will be more positive than the opposite.

• How? Anybody willing to actively take part, in terms of writings, code or suggestions is welcome to do so. Everything should happen publicly through Github. Contributors become co-authors.

• La question qui fâche, what about author order? Co-authors will be listed on the paper in chronological order of first contribution. Period.