AI for Neuroscience & Neuroscience for AI

Irina Rish, Researcher, AI Science, IBM T.J. Watson Research Center
Presented at MLconf 2018
Abstract: AI and neuroscience share the same age-old goal: to understand the essence of intelligence. Thus, despite different tools used and different questions explored by those disciplines, both have a lot to learn from each other. In this talk, I will summarize some of our recent projects which explore both directions, AI for neuro and neuro for AI. AI for neuro involves using machine learning to recognize mental states and identify statistical biomarkers of various mental disorders from heterogeneous data (neuroimaging, wearables, speech), as well as applications of our recently proposed hashing-based representation learning to dialog generation in depression therapy. Neuro for AI implies drawing inspirations from neuroscience to develop better machine learning algorithms. In particular, I will focus on the continual (lifelong) learning objective, and discuss several examples of neuro-inspired approaches, including (1) neurogenetic online model adaptation in nonstationary environments, (2) more biologically plausible alternatives to backpropagation, e.g., local optimization for neural net learning via alternating minimization with auxiliary activation variables, and co-activation memory, (3) modeling reward-driven attention and attention-driven reward in contextual bandit setting, as well as (4) modeling and forecasting behavior of coupled nonlinear dynamical systems such as brain (from calcium imaging and fMRI) using a combination of analytical van der Pol model with LSTMs, especially in small-data regimes, where such hybrid approach outperforms both of its components used separately.

[teckyify]

The point of feature engineering for machine learning, like classification, is a quite limited issue compared to the things the brain does. So it does not really make sense to oppose generic learning models to single purpose ML tasks. That's the whole point why we look at brains. Since nobody rejects that data-driven and prior must be combined, the question is *which* priors (or innate knowledge) and not *if*.