AI and Machine Learning

My group employs a great deal of AI/ML methods. Sometimes, we have had to develop/extend ML methods to solve novel problems that came up in our research; sometimes, insights from natural intelligence or neuroscience led to novel advances in AI/ML.

For example, when I learned about stochastic control theory, I proved that there was a formal mathematical equivalence between the optimal change-detection problem in stochastic control theory and the neuronal problem of detecting a change in pre-synaptic firing rate (Yu, 2007).

Soon after, I collaborated with Peter Frazier to prove that the optimal decision policy for a sequential hypothesis testing problem (such as faced by the brain when doing 2AFC tasks based on noisy sensory input) consists of a pair of converging thresholds on the cumulative log likelihood ratio (Frazier & Yu, 2008).

I then collaborated with Savas Dayanik to prove that the optimal decision policy for a sequential hypothesis testing problem with the objective of maximizing reward rate is equivalent to a special case with the objective of maximizing a linear combination of decision time and error rate, and can be solved using iterative dynamic programming (Dayanik & Yu,

2013).

Later, my student and I (Ahmad, Huang, Yu, 2013) showed how computational intense procedure of optimizing an active sensing problem can be solved approximately using by applying a Gaussian process model to the action value functions and doing off-line sample-based dynamic programming.

Currently, I have a collaboration with Mikhail Belkin to understand how overparameterized systems can self-regularize in supervised learning, based on insights we obtained from modeling human face processing; I am hopeful that this will also lead to novel theoretical insights into why multiple sensory systems (e.g. vision in primates, olfaction in flies) have overcomplete sensory projections, and how that benefits animals at the behavioral level.

Related Papers

  • Ahmad, S, Huang, H, & Yu, A J (2013). Context-sensitivity in human active sensing. Advances in Neural Information Processing Systems 26. MIT Press, Cambridge, MA.
  • Dayanik, S & Yu, A J (2013). Reward-Rate Maximization in Sequential Identification under a Stochastic Deadline. SIAM Journal on Control and Optimization, 51 (4): 2922-2948.
  • Frazier, P, Yu, A J (2008). Sequential hypothesis testing under stochastic deadlines. Advances in Neural Information Processing Systems, 20: 465-72. MIT Press, Cambridge, MA.
  • Yu, A J (2007). Optimal change-detection and spiking neurons. Advances in Neural Information Processing Systems 19: 1545-52. MIT Press, Cambridge, MA.