NC unit secretary
2020-10-09 10:34:45

This is a discussion channel for “Learning to Learn and the Brain” by Prof. Xiao-Jing Wang (New York University). A link to the talk is the following. Please do not share the link to anyone outside of this slack workspace. Access Passcode can be found at the announcement channel.   URL: https://vimeo.com/471281813 (24 minutes)

Yuichi Iino
2020-10-11 13:45:25

Hi Dr. Wang, thank you for the interesting talk. I am not sure whether I understood the talk fully, but got that the RNN could "learn to learn" the rule, and by repeating the training it could quickly output the right behavior to a novel pair of stimuli. I just wonder whether you have tried to present a pair of images that the network previously saw. Does it remember the pair to some extent, which human and monkeys probably do. Do you have layers in the network that could do this job ?

Raunak Basu
2020-10-11 21:02:53

Thank you for a wonderful talk Dr. Wang. I had two questions about the talk. 1) Can the network do reversal learning where the saccade contingencies of the same pair of pictures reverses in future? 2) How does the dynamics of decision making (in the decision making subspace) compare with the dynamics of stimulus identification (in the stimulus subspace). In other words, Does the decision 'emerge' simultaneously with stimulus identification or does occur in a sequential manner with stimulus recognition followed by decision making?

Hadiseh Hajimohammadi
2020-10-16 01:41:29

*Thread Reply:* the second question is mine as well :) I appreciate if you share the answer

Mingbo Cai
2020-10-11 21:19:07

Thanks for the great talk! I guess for the meta-learning to work, there must still be something inherently shared across the tasks for the RNN to generalize. What do you think is the commonality among the tasks you gave the network? One step further, if we consider all the possible tasks that humans can face daily, do you think one generic RNN is sufficient, or do we need some device to select different RNNs in real time for different classes of tasks? If the latter, what do you think is a good way to organize (or quantify) the space of all tasks?

👍 Seaurchinjinni, Longfei Wang