Evolution of neural activity in circuits bridging sensory and abstract knowledge
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
The findings of the paper are of interest to scientists studying the learning of abstract representations. It provides insights into how feedforward networks evolve during a process of learning to map stimuli onto abstract classes via gradient descent. The results are appealing and the analyses thorough. As well, the paper makes some experimental predictions. It could benefit from a deeper discussion on how the findings may generalize to biologically more realistic networks and tasks.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Neuroscience (eLife)
Abstract
The ability to associate sensory stimuli with abstract classes is critical for survival. How are these associations implemented in brain circuits? And what governs how neural activity evolves during abstract knowledge acquisition? To investigate these questions, we consider a circuit model that learns to map sensory input to abstract classes via gradient-descent synaptic plasticity. We focus on typical neuroscience tasks (simple, and context-dependent, categorization), and study how both synaptic connectivity and neural activity evolve during learning. To make contact with the current generation of experiments, we analyze activity via standard measures such as selectivity, correlations, and tuning symmetry. We find that the model is able to recapitulate experimental observations, including seemingly disparate ones. We determine how, in the model, the behaviour of these measures depends on details of the circuit and the task. These dependencies make experimentally testable predictions about the circuitry supporting abstract knowledge acquisition in the brain.
Article activity feed
-
-
-
Author Response
Reviewer 2 (Public Review):
The paper addresses the question of how brain circuits associate stimuli onto abstract representations, and how both the neuronal activity and the synaptic connectivity change during this process. To do so, the authors make use of a feedforward network model that learns to map stimuli vectors onto two categories by means of gradient descent. They show that the model successfully learns the abstract classes in a simple and context-dependent categorisation task. The authors analyse a number of measures, like category and context selectivity to link their results to experimental findings. Moreover, they analyse the network thoroughly and unravel network and task properties that may underlie previous, seemingly contradictory experimental findings. The paper is very well written, the analyses …
Author Response
Reviewer 2 (Public Review):
The paper addresses the question of how brain circuits associate stimuli onto abstract representations, and how both the neuronal activity and the synaptic connectivity change during this process. To do so, the authors make use of a feedforward network model that learns to map stimuli vectors onto two categories by means of gradient descent. They show that the model successfully learns the abstract classes in a simple and context-dependent categorisation task. The authors analyse a number of measures, like category and context selectivity to link their results to experimental findings. Moreover, they analyse the network thoroughly and unravel network and task properties that may underlie previous, seemingly contradictory experimental findings. The paper is very well written, the analyses and mathematical derivations are very thorough and the results are convincing. However, the work and its presentation would benefit from a few changes:
- The paper may benefit from a more thorough discussion on how the results fit into the current literature (neuroscience and machine learning) and how the findings may generalise to more complex tasks and network structures (Dale’s principle, including recurrent/feedback connections, more than two categories, more than one hidden layer, alternatives to gradient descent).
- While the simulations and detailed analyses in the results and methods section are very convincing, some claims should be also supported by more intuitive explanations so that a broader audience can be reached.
- The introduction to the context-dependent task may need to be revised because as now the difference to the simple task presented first is not immediately clear.
- It would be nice if their findings could be related back to the experimental literature more qualitatively. While the authors mention the contradictory findings in monkey and rat PFC vs. monkey LIP in their introduction, a thorough comparison with those findings is missing.
We thank the reviewer for his detailed assessment and his supportive words. We hope that our revision addresses your suggestions. Concerning point 4: we agree with the reviewer that a thorough comparison with experimental findings would be important, and is currently missing. A thorough comparison would require, however, a number of additional steps that we feel lie beyond the scope of this manuscript (adapt the tasks to each different experimental setup, e.g. by increasing the number of categories and changing the structure of context-dependent associations; re-analyse experimental data).
We have thus decided to leave this major effort for future work.
-
Evaluation Summary:
The findings of the paper are of interest to scientists studying the learning of abstract representations. It provides insights into how feedforward networks evolve during a process of learning to map stimuli onto abstract classes via gradient descent. The results are appealing and the analyses thorough. As well, the paper makes some experimental predictions. It could benefit from a deeper discussion on how the findings may generalize to biologically more realistic networks and tasks.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)
-
Reviewer #1 (Public Review):
This is a beautiful paper, which blends strong theoretical results (very well organised in the supplementary material) with intuitive descriptions of the results. The novelty of the theoretical developments in their own right is perhaps eclipsed by similar recent theoretical work in deep learning around the neural tangent kernel, but it is nevertheless great to see these ideas shed light on neural phenomena -- and this paper does this very well. We found that the study is given just the right scope: two learning tasks of increasing difficulty, both simple enough to enable mathematical analysis yet close enough to the type of tasks used in neuroscience as to enable meaningful comparisons to neural data. It is rare enough to be mentioned: the figure are beautiful and we found them of very high illustratory …
Reviewer #1 (Public Review):
This is a beautiful paper, which blends strong theoretical results (very well organised in the supplementary material) with intuitive descriptions of the results. The novelty of the theoretical developments in their own right is perhaps eclipsed by similar recent theoretical work in deep learning around the neural tangent kernel, but it is nevertheless great to see these ideas shed light on neural phenomena -- and this paper does this very well. We found that the study is given just the right scope: two learning tasks of increasing difficulty, both simple enough to enable mathematical analysis yet close enough to the type of tasks used in neuroscience as to enable meaningful comparisons to neural data. It is rare enough to be mentioned: the figure are beautiful and we found them of very high illustratory value (e.g. Figs 3 and 7, in particular, allowed us to understand the main results in a matter of seconds). We haven't found any issue in the analysis and the paper is in great shape already.
-
Reviewer #2 (Public Review):
The paper addresses the question of how brain circuits associate stimuli onto abstract representations, and how both the neuronal activity and the synaptic connectivity change during this process. To do so, the authors make use of a feedforward network model that learns to map stimuli vectors onto two categories by means of gradient descent. They show that the model successfully learns the abstract classes in a simple and context-dependent categorisation task. The authors analyse a number of measures, like category and context selectivity to link their results to experimental findings. Moreover, they analyse the network thoroughly and unravel network and task properties that may underlie previous, seemingly contradictory experimental findings. The paper is very well written, the analyses and mathematical …
Reviewer #2 (Public Review):
The paper addresses the question of how brain circuits associate stimuli onto abstract representations, and how both the neuronal activity and the synaptic connectivity change during this process. To do so, the authors make use of a feedforward network model that learns to map stimuli vectors onto two categories by means of gradient descent. They show that the model successfully learns the abstract classes in a simple and context-dependent categorisation task. The authors analyse a number of measures, like category and context selectivity to link their results to experimental findings. Moreover, they analyse the network thoroughly and unravel network and task properties that may underlie previous, seemingly contradictory experimental findings. The paper is very well written, the analyses and mathematical derivations are very thorough and the results are convincing. However, the work and its presentation would benefit from a few changes:
The paper may benefit from a more thorough discussion on how the results fit into the current literature (neuroscience and machine learning) and how the findings may generalise to more complex tasks and network structures (Dale's principle, including recurrent/feedback connections, more than two categories, more than one hidden layer, alternatives to gradient descent).
While the simulations and detailed analyses in the results and methods section are very convincing, some claims should be also supported by more intuitive explanations so that a broader audience can be reached.
The introduction to the context-dependent task may need to be revised because as now the difference to the simple task presented first is not immediately clear.
It would be nice if their findings could be related back to the experimental literature more qualitatively. While the authors mention the contradictory findings in monkey and rat PFC vs. monkey LIP in their introduction, a thorough comparison with those findings is missing.
-
-