Cognitive models of behavior are a key part of neuroscience. But discovering them is hard!
Neural networks are powerful models. But interpreting them cognitively is hard!
We explore automatically learning interpretable models using "Disentangled RNNs"
Applications are open for the
@GoogleDeepMind
Student Researcher Program!
There will likely be projects available to work with me and with others in computational neuroscience. If interested, please feel free to get in touch!
Learn more and apply here:
How do new habits form?
Many computational models propose that actions become habits when they consistently lead to rewards.
"Habits without Values" showcases a model in which actions become habits when they are repeated consistently (even if unrewarded).
How do brains make multi-step plans?
Neuroscience does not yet know! But it does have many interesting empirical clues and some solid theoretical ideas.
Sarah Jo Venditto and I describe some of these and propose directions for future work:
The brain represents "expected value" of states and actions in the world, but what do these representations do?
Our preprint argues that (at least in the OFC) they play an important role in learning, but no direct role in choice.
I am excited about applying approaches like this to a broader range of datasets, and about using neural networks for scientific discovery more generally.
Thanks to brilliant and inspiring co-authors
@eckstein_maria
,
@mattbotvinick
, and
@zebkDotCom
!
When trained on the rat datasets, these networks could provide better quality-of-fit than the best previously-known cognitive model. Fit quality was similar to that of an LSTM (a classic black-box RNN).
Two key features encourage these networks to learn interpretable models.
First: each latent variable (element of the recurrent state) is updated by a separate sub-network.
Second: information bottlenecks penalize retaining information that is not being used.
We fit these networks to large behavioral datasets from a classic reward learning task.
We consider both synthetic datasets generated by handcrafted artificial agents and laboratory datasets generated by rats ()
@twitemp1
@IrisVanRooij
But: you shouldn't have to take my word for that, and you shouldn't have had to look up a previous paper to understand this one. I'll address both of these in the next preprint iteration. If you have more thoughts on the paper, I'd love to hear them (either here or by email)!
@twitemp1
@IrisVanRooij
Hi Esther, thanks for pointing this out! I see you've found our previous paper, where we did a lot to characterize and validate the behavior. The rats in this paper are like those rats in all the ways that matter.
Builds on previous work in econometrics, and in the neuroscience of RL.
Yule, 1926. Why do we sometimes get nonsense correlations between timeseries?
Elber-Dorozko & Lowenstein, 2018. Striatal action-value neurons reconsidered.
@rei_akaishi
@brody_lab
Best data I know of here are from Matt Gardner at Schoenbaum lab: Silencing OFC in an economic choice task has no effect on behavior.
I agree that checking this out in a one-step learning task would be super informative!
@rei_akaishi
@twitemp1
The simulations we did here are all fully-observable, but I think the ideas can generalize straightforwardly to the partially-observable case. Simplest would probably be to model habits as direct S-R links from observations onto actions.
@QueenieLB
@KateWassum
Great question! I'd been imagining some kind of learning or update process during the test period (as per your other tweet). But this literature is definitely something I want to think about more deeply. Lots of differences between behaviors, so lots of possibilities!
...in an upcoming issue of COBS on computational cognitive neuroscience, edited by Geoff Schoenbaum and Angela Langdon (neither they nor Sarah Jo seem to be on Twitter).
These articles are all great, and more are on their way!
@MelissaJSharpe
I do! In this paper, we model the S-R associations as value-free ("cached policy"), and argue that this has some advantages vs. modeling them as MF-RL's "cached values".