Part of my series of notes from ICLR 2019 in New Orleans.
Introduction
- using game theory instead of optimization
- community has (mostly) moved on from “make ML work” to other issues
- a brief look at some connections among them…
Cambrian Explosion
- generative modeling – GANs hooray
- there is a Nash equilibrium where generator perfectly recreates data and discriminator guesses, but in practice we don’t reach this
- great progress on images generally and faces especially
- doing this for ImageNet still needs labels
- needs much less supervision
- challenge in doing this for videos and making them temporally consistent
- security
- adversarial examples – breaking the iid, test-matching-train assumption
- see here basically
- model-based optimization
- learn a model of what you want, and optimize it
- e.g. designing DNA for protein function
- (I definitely don’t hear as much about this as I feel like I should! maybe I’m just not listening in the right places)
- RL
- self-play is a form of adversarial learning (and the original ML! Arthur Samuel’s 1959 checkers agent)
- desired behaviours as side effects of “winning the game” (GANs are like this)
- SPIRAL (Ganin et al. 2018) – use GANs to help learn reward function
- eXtreme reliability
- dealing with the worst case
- modeling this is simpler than the average case (why???)
- Reluplex model (Katz et al. 2017)
- dealing with the worst case
- label efficiency – semi-supervised learning
- discriminator distinguishes classes as well as real vs. fake (Odena 2016)
- learn from fake data, real labeled data, and real unlabelled data
- virtual adversarial training (Miyato et al. 2017)
- discriminator distinguishes classes as well as real vs. fake (Odena 2016)
- domain adaptation
- domain adversarial networks (Ganin et al. 2015) – learn features common to domains and ones that distinguish them
- professor forcing (Lamb et al. 2016) – domain adversarial applied to RNN
- one domain is train mode, other is predict mode
- adapting between synthetic and real images to bootstrap models
- fairness
- adversarially learned fair representations, see other talk
- interpretability
- linear doesn’t mean interpretable
- robust might mean more interpretable
- neuroscience
- “bugs” in both neural nets and humans?