Neural signatures of motor memories emerge in neural network models

Reviewer #1 (Public review):

Summary:

This work investigates the neural basis of continual motor learning, specifically how brains might accommodate new motor memories without interfering with previously learned behaviours. Mainly drawing inspiration from recent experimental studies in monkeys (Losey et al. and Sun, O'Shea et al.), the authors use recurrent neural networks (RNNs) to model sequential learning and examine the emergence and properties of two proposed neural signatures of motor memory: the "uniform shift" observed in preparatory activity and the "memory trace" observed in execution activity.

Strengths:

The work's main contribution is demonstrating that both uniform shifts and memory traces emerge in RNN models trained on a sequential BCI task, without requiring explicit additional mechanisms. The work explores the relationship between these signatures and behavioural savings, finding that the memory trace correlates with immediate retention savings in networks without context, while the uniform shift does not. The study also investigates how properties of the new task perturbation (within- vs. outside-manifold) and the presence of explicit context cues affect these signatures and their relationship to savings, generally finding that context signals and outside-manifold perturbations reduce savings by decreasing the inherent overlap in the neural strategies used to solve the task.

Weaknesses:

A primary weakness is the lack of clear definitions of the uniform shift and the memory trace, which are quite different metrics. Another primary weakness is that the task modelled is well-matched to the Losey et al. BCI paradigm, but not well-matched to the Sun, O'Shea et al.'s curl field paradigm, which is likely impacting some of the results, primarily the lack of a relationship between the uniform shift and motor memories. While there are improvements that could be made in this work, we think it is a demonstration that modeling learning in neural activity using neural network models continues to be a valuable tool, moving the field forward.

Reviewer #2 (Public review):

Summary:

Chang et al. develop an RNN model of a BCI sequential learning task to examine the emergence of motor memory in the network. They use this system to quantify signatures of memory in continual learning, comparing their model with experimental observations from monkeys in prior publications. They show that the RNN model has signatures of shifts associated with sequential learning without any non-standard learning rules. This convincing study contributes to the knowledge of how motor memories are formed and shaped so that they are flexible in acquiring multiple behaviors.

Strengths:

This paper describes a well-designed numerical experiment that comes to a clear interpretation of a set of neural BCI experiments. The learning signatures the authors describe are interesting and well laid out, and the paper is well written. I find it insightful that the neural signature of motor learning emerges in a trained network without special learning rules.

Weaknesses:

The paper could be stronger if it made a stronger interpretation of how memory traces and uniform shifts are related. These two observations are taken from the BCI sequential learning literature and introduced by two different prior experimental papers on two different tasks, so it seems like there is an opportunity here to use the RNN model to unite these concepts, or define another metric for signatures of learning from a more normative approach.

Reviewer #3 (Public review):

Summary:

The authors build and analyze recurrent neural network (RNN) models of brain-computer interface (BCI) multi-task learning, developing a valuable theoretical understanding of learning-related neural population phenomena ("memory traces" and "uniform shifts") that have been reported in recent experimental studies of BCI and motor learning. The authors find that both phenomena emerge in their RNN models, and both correlate in some manner to learning-related behavioral phenomena ("savings" and "forgetting"). The authors also reveal that RNN training details, in particular, incorporating a task-indicating contextual input, can impact these population-level signatures of learning in RNN activity and their relation to those behavioral phenomena.

Strengths:

The text is well written, and the figures are clearly composed to convey the core concepts and findings. The RNN studies are elegant in their ability to recapitulate the memory trace and uniform shift phenomena, and further allow evaluations of novel scenarios that were not tested in the original corpus of the modeled animal experiments. The authors assess the sensitivity of their results to multiple approaches to RNN training, including training connectivity within a model of motor cortex, training only an upstream model that provides inputs to the motor cortex model, and providing task-indicating contextual inputs.

Weaknesses:

(1) It is unclear to what extent these RNN models operate in regimes relevant to biological neural networks (e.g., motor cortex), even at the neural-population level of abstraction studied here. Can the authors speak to how sensitive their results are to details that might speak to these operating regimes (e.g., signal-to-noise ratios or dimensionality of the RNN activities)?

(2) The work could be further strengthened by analyses demonstrating a more direct link between the neural population phenomena (memory trace and uniform shift) and the behavioral phenomena (savings, forgetting, etc). While in animal experiments, it can be exceedingly difficult to demonstrate links beyond correlative effects, the promise of a model is the relative tractability of implementing manipulations that might establish something closer to a causal link between phenomena. Is it the case that the memory trace is a task-dependent, mean-preserving rotation of the across-target task-relevant activity space? And that the uniform shift is a translation (non-mean-preserving) of that space? If so, could the authors design regularization schemes that specifically target each of these effects, enabling a more direct test of the functional role the effects play in driving behavioral phenomena?

Minor Comments:

The current study is based on BCI learning of center-out tasks, analogous to the Losey et al. task that initially reported the memory trace phenomena. However, a rather different behavioral task - involving arm movements through curl force fields - was employed by the Sun, O'Shea, et al. study that originally reported the uniform shift phenomena. How should readers interpret the current study's findings related to the uniform shift? To what extent might the behavioral implications of the uniform shift depend on the demands of the task, e.g., the biomechanics, day-to-day experiencing of different curl-field perturbations, etc.?

Author response:

We thank the reviewers for their thoughtful comments, and we plan to implement many of their suggestions to improve the paper. We agree that the paper can benefit from clearer links between the two neural signatures (memory traces and uniform shifts) themselves, and between the neural signatures and behavioral phenomena. We will address these limitations in multiple ways. First, as the reviewers noted, RNN models have the potential to probe these relationships, so we plan to perform further analyses and modeling experiments to uncover any causal relationships. Second, we will also establish clearer definitions of the neural signatures and explore how these signatures can be unified using our models. Finally, we will compare the experimental paradigms between Losey et al and Sun, O’Shea et al, and discuss how differences between the paradigms may have impacted our observations, particularly in the context of other experimental and modeling papers.