1/ Hark, a #tweeprint! Our new paper is up on #bioRxiv! It’s the results from our #OpenScope project with the @AllenInstitute about learning from unexpected events in the neocortical microcircuit!
Here's a thread with more details...
biorxiv.org/content/10.110…
Here's a thread with more details...
biorxiv.org/content/10.110…
2/ In this paper, we show that unexpected events drive different changes in the responses of somata and distal apical dendrites in primary visual cortex pyramidal neurons. 
3/ Previous research has shown that neurons in sensory areas respond to unexpected events. It’s hypothesized that these responses guide our brain in learning a predictive, hierarchical model of the world in an unsupervised (or self-supervised) manner.
4/ This hypothesis is part of a broad class of models, including predictive coding, Helmholtz machines, contrastive predictive coding, etc. This larger class of models makes four predictions that we tested here.
5/ (Prediction 1) Neural responses to unexpected and expected should be different. They must be, otherwise there would be no way to learn from unexpected events.
6/ (Prediction 2) Unexpected event responses should change with experience, as the system comes to effectively expect them and they transition away from being unexpected.
7/ (Prediction 3) The responses to unexpected events should change differently in top-down
(i.e., distal apical dendrites) and bottom-up (i.e., somata) neuronal compartments. This is implied by the hierarchy in the models.
(i.e., distal apical dendrites) and bottom-up (i.e., somata) neuronal compartments. This is implied by the hierarchy in the models.
8/ (Prediction 4) The responses of neurons to unexpected events should predict how the responses evolve over time: this is critical for learning from the unexpected events.
9/ To test this, we habituated animals to repeating A-B-C-D-Gray stimuli (with local randomness, but global patterns). Then, we introduced unexpected events where we replaced the “D” frames with “U” frames (A-B-C-U-G) that violated the global patterns.
10/ What did we observe?
(Prediction 1) Not only are somata of L2/3 and L5 neurons in primary visual cortex selective to the unexpected events, so are distal apical dendrite branches.
(Prediction 1) Not only are somata of L2/3 and L5 neurons in primary visual cortex selective to the unexpected events, so are distal apical dendrite branches.
11/ What happens over days, as the animals get more experience with the stimulus?
(Predictions 2-3) With experience, the unexpected event selectivity changes in the somata and distal apical dendrites: specifically, it tends to grow in the dendrites, but weaken in the somata.
(Predictions 2-3) With experience, the unexpected event selectivity changes in the somata and distal apical dendrites: specifically, it tends to grow in the dendrites, but weaken in the somata.
12/ Are the changes in unexpected event selectivity that we observe with experience predictable at the individual soma/dendritic branch level?
(Prediction 4) Yes! And how selectivity changes is different in these two neuronal compartments.
(Prediction 4) Yes! And how selectivity changes is different in these two neuronal compartments.
13/ Specifically, the day 1 strongly selective somata tend to show the biggest drop by day 2, whereas the day 2 most selective dendrites show the biggest increase by day 3.
So, both L2/3 and L5 pyramidal neurons appear to be specifically learning from these unexpected events.
So, both L2/3 and L5 pyramidal neurons appear to be specifically learning from these unexpected events.
14/ To sum up, our data matched predictions 1-4 above, strongly supporting the hypothesis that the brain leverages unexpected events to learn a predictive, hierarchical model of the world, with somata and distal apical dendrites playing different computational roles.
15/ Of course, it also raises lots of interesting questions:
How are bottom-up and top-down information integrated in individual neurons?
Where is the learning occurring (i.e., what is plastic)?
Where are the expectations being formed?
How are bottom-up and top-down information integrated in individual neurons?
Where is the learning occurring (i.e., what is plastic)?
Where are the expectations being formed?
16/ We’re very pleased to say that in addition to our code being shared on @github, the full dataset is openly available in @NeurodataWB format on @DANDIarchive. Looking forward to seeing what you discover in the data!
gui.dandiarchive.org/#/dandiset/000…
gui.dandiarchive.org/#/dandiset/000…
17/ As mentioned above, this highly collaborative project was part of an @AllenInstitute #OpenScope collaboration. Thanks to the @AllenInstitute for this fantastic opportunity and beautiful data!
18/ The project was conceptualized by Joel Zylberberg, @tyrell_turing, Timothy Lillicrap and Yoshua Bengio, with @LecoqJerome leading the exceptional data collection team at the @AllenInstitute Brain Observatory, and Jay Pina and me leading the data analysis.
19/ Many thanks to our excellent co-authors, and collaborators who contributed to this team effort, and to @CIFAR_News, @CRC_CRC, @NSERC_CRSNG, @SloanFoundation, @ONgov, @ComputeCanada, @ComputeOntario, …
20/ As well as @UTSC, @UofT, @yorkuniversity, @AllenInstitute, @TheNeuro_MNI, and @Mila_Quebec for making this work possible.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
