Neural manifold under plasticity in a goal driven learning behaviour

Abstract

Neural activity is often low dimensional and dominated by only a few prominent neural covariation patterns. It has been hypothesised that these covariation patterns could form the building blocks used for fast and flexible motor control. Supporting this idea, recent experiments have shown that monkeys can learn to adapt their neural activity in motor cortex on a timescale of minutes, given that the change lies within the original low-dimensional subspace, also called neural manifold. However, the neural mechanism underlying this within-manifold adaptation remains unknown. Here, we show in a computational model that modification of recurrent weights, driven by a learned feedback signal, can account for the observed behavioural difference between within- and outside-manifold learning. Our findings give a new perspective, showing that recurrent weight changes do not necessarily lead to change in the neural manifold. On the contrary, successful learning is naturally constrained to a common subspace.

Thalamic control of cortical dynamics in a model of flexible motor sequencing

“…This is compatible with experimental reports showing substantial changes in motor cortical activity between different movements and/or contexts (Miri et al, 2017;Al Borno et al, 2020;Sun et al, 2020), as well as during prolonged brain machine interface training (Oby et al, 2019). Further analyses will be needed to investigate whether these large changes of cortical activity relate to changes in the effective cortical connectivity (Feulner and Clopath, 2021). Finally, our model also postulates that the thalamic neurons involved in shaping cortical dynamics during motif execution are segregated into motif-specific subpopulations.…”

Modeling of hyper-adaptability: from motor coordination to rehabilitation

“…Neural network modeling has emerged as a powerful tool to emulate the neural functions and to infer their circuit mechanisms, such as sensory processing [42], decision making [43], and motor control [44][45][46]. One of the biggest successes is the application of feedforward neural networks (i.e.…”

“…Their results showed that monkeys could readily learn to control the cursor when the change was within the original subspace, whereas they were less able to learn the new mapping if it was outside of the original subspace. To understand the circuit mechanisms that cause the difference in the motor learning ability, Feulner and Clopath (2020) trained RNNs and compared their adaptability to change of the BCI mapping within and outside of the original subspace [45]. Their results showed that RNN can predict error feedback signal more correctly when the change was within the original subspace than when it was outside of the original subspace and resulted in the better learning performance, suggesting that the low-dimensional subspace provided a constraint to correctly estimate the error feedback for motor learning.…”

“…This result suggests that the network dynamics of a lower system might constrain a control space where the higher system needs to explore to re-optimize the network dynamics during hyperadaptation. This constraint might help the motor system to achieve faster and robust learning similar to the BCI motor learning within the original subspace [45,55].…”

Small, correlated changes in synaptic connectivity may facilitate rapid motor learning

“…RNNs trained on motor, cognitive and BCI tasks exhibit many striking similarities with the activity of neural populations recorded in animal studies [Mante et al, 2013, Sussillo et al, 2015, Rajan et al, 2016, Song et al, 2017, Wang et al, 2018, Michaels et al, 2020, Perich et al, 2021, suggesting a fundamental similarity between the two. Previous work using RNNs to model the BCI experiment described above [Sadtler et al, 2014] showed that network covariance can be highly preserved even when learning is happening through weight changes within the network [Feulner and Clopath, 2021]. Thus, contrary to widespread intuition, functionally relevant synaptic weight changes may not necessarily lead to measurable changes in statistical interactions across neurons [Das and Fiete, 2020].…”