A Virtuous Circle Between Neuroscience and Deep Learning: Short-Term Memory as a Case Study

A Virtuous Circle Between Neuroscience and Deep Learning: Short-Term Memory as a Case StudyDr. Emin Orhan
New York University
The second candidate for the joint position with neurobiology is interviewing next week.
Emin Orhan is currently a postdoc with the Center for Data Science at New York University. Emin’s talk will take place Tuesday Feb 4th at 10am – Room 4500 Pacific Hall (see attached flyer and abstract below). Also attached is the candidates CV. Please sign up to meet with Emin on Wednesday Feb 5th at the following link:
https://docs.google.com/spreadsheets/d/1pllicnp4_A6lE2_k94vif7CNOlfpnLIaerEo31CfE9c/edit?usp=sharing
Please sign up only on Wednesday Feb 5th (the blue shaded section on the right). The peach-shaded section on the left will display Emin’s schedule for their day visiting with neurobiology.
Abstract:
Dr. Emin Orhan’s research is aimed at understanding the high level computational principles and lower level circuit mechanisms underlying biological intelligence. His recent work uses deep learning to address these fundamental questions. In the first part of his talk, Dr. Orhan will describe his recent efforts to understand the nature of the neural representations underlying short-term memory maintenance in the brain, using task-trained recurrent neural networks (Orhan & Ma, Nature Neuroscience, 20191). Consistent with experimental findings, his results show that these models display a spectrum of maintenance mechanisms ranging from sequential delay activity to more persistent delay activity depending on a variety of task- and circuit-related factors. The complete experimental tractability of these models also aids a detailed understanding of the circuit level mechanisms underlying this spectrum of solutions. In the second part of his talk, Dr. Orhan will discuss how he can use the insights gained from the first part to improve artificial neural networks (Orhan & Pitkow, ICLR, 20202). Specifically, he will show that sequential neural activity can be used as a novel, useful inductive bias to improve memory in recurrent neural networks solving practical machine learning tasks. Together these two projects illustrate how deep learning and neuroscience can mutually inform each other, to the benefit of both fields.
Emin_Orhan_CV (PDF Document)