A theory of brain-computer interface learning via low-dimensional control

A remarkable demonstration of the flexibility of mammalian motor systems is primates' ability to learn to control brain-computer interfaces (BCIs). This constitutes a completely novel motor behavior, yet primates are capable of learning to control BCIs under a wide range of conditions. BCIs with carefully calibrated decoders, for example, can be learned with only minutes to hours of practice. With a few weeks of practice, even BCIs with randomly constructed decoders can be learned. What are the biological substrates of this learning process? Here, we develop a theory based on a re-aiming strategy, whereby learning operates within a low-dimensional subspace of task-relevant inputs driving the local population of recorded neurons. Through comprehensive numerical and formal analysis, we demonstrate that this theory can provide a unifying explanation for disparate phenomena previously reported in three different BCI learning tasks, and we derive a novel experimental prediction that we verify with previously published data. By explicitly modeling the underlying neural circuitry, the theory reveals an interpretation of these phenomena in terms of biological constraints on neural activity.

Learning leaves a memory trace in motor cortex

How are we able to learn new behaviors without disrupting previously learned ones? To understand how the brain achieves this, we used a brain-computer interface (BCI) learning paradigm, which enables us to detect the presence of a memory of one behavior while performing another. We found that learning to use a new BCI map altered the neural activity that monkeys produced when they returned to using a familiar BCI map in a way that was specific to the learning experience. That is, learning left a "memory trace" in the primary motor cortex. This memory trace coexisted with proficient performance under the familiar map, primarily by altering neural activity in dimensions that did not impact behavior. Forming memory traces might be how the brain is able to provide for the joint learning of multiple behaviors without interference.

The role of prospective contingency in the control of behavior and dopamine signals during associative learning

Associative learning depends on contingency, the degree to which a stimulus predicts an outcome. Despite its importance, the neural mechanisms linking contingency to behavior remain elusive. Here we examined the dopamine activity in the ventral striatum - a signal implicated in associative learning - in a Pavlovian contingency degradation task in mice. We show that both anticipatory licking and dopamine responses to a conditioned stimulus decreased when additional rewards were delivered uncued, but remained unchanged if additional rewards were cued. These results conflict with contingency-based accounts using a traditional definition of contingency or a novel causal learning model (ANCCR), but can be explained by temporal difference (TD) learning models equipped with an appropriate inter-trial-interval (ITI) state representation. Recurrent neural networks trained within a TD framework develop state representations like our best 'handcrafted' model. Our findings suggest that the TD error can be a measure that describes both contingency and dopaminergic activity.

Emergence of belief-like representations through reinforcement learning

To behave adaptively, animals must learn to predict future reward, or value. To do this, animals are thought to learn reward predictions using reinforcement learning. However, in contrast to classical models, animals must learn to estimate value using only incomplete state information. Previous work suggests that animals estimate value in partially observable tasks by first forming "beliefs"-optimal Bayesian estimates of the hidden states in the task. Although this is one way to solve the problem of partial observability, it is not the only way, nor is it the most computationally scalable solution in complex, real-world environments. Here we show that a recurrent neural network (RNN) can learn to estimate value directly from observations, generating reward prediction errors that resemble those observed experimentally, without any explicit objective of estimating beliefs. We integrate statistical, functional, and dynamical systems perspectives on beliefs to show that the RNN's learned representation encodes belief information, but only when the RNN's capacity is sufficiently large. These results illustrate how animals can estimate value in tasks without explicitly estimating beliefs, yielding a representation useful for systems with limited capacity.

Emergence of belief-like representations through reinforcement learning

To behave adaptively, animals must learn to predict future reward, or value. To do this, animals are thought to learn reward predictions using reinforcement learning. However, in contrast to classical models, animals must learn to estimate value using only incomplete state information. Previous work suggests that animals estimate value in partially observable tasks by first forming "beliefs"-optimal Bayesian estimates of the hidden states in the task. Although this is one way to solve the problem of partial observability, it is not the only way, nor is it the most computationally scalable solution in complex, real-world environments. Here we show that a recurrent neural network (RNN) can learn to estimate value directly from observations, generating reward prediction errors that resemble those observed experimentally, without any explicit objective of estimating beliefs. We integrate statistical, functional, and dynamical systems perspectives on beliefs to show that the RNN's learned representation encodes belief information, but only when the RNN's capacity is sufficiently large. These results illustrate how animals can estimate value in tasks without explicitly estimating beliefs, yielding a representation useful for systems with limited capacity.

Author Summary

Natural environments are full of uncertainty. For example, just because my fridge had food in it yesterday does not mean it will have food today. Despite such uncertainty, animals can estimate which states and actions are the most valuable. Previous work suggests that animals estimate value using a brain area called the basal ganglia, using a process resembling a reinforcement learning algorithm called TD learning. However, traditional reinforcement learning algorithms cannot accurately estimate value in environments with state uncertainty (e.g., when my fridge's contents are unknown). One way around this problem is if agents form "beliefs," a probabilistic estimate of how likely each state is, given any observations so far. However, estimating beliefs is a demanding process that may not be possible for animals in more complex environments. Here we show that an artificial recurrent neural network (RNN) trained with TD learning can estimate value from observations, without explicitly estimating beliefs. The trained RNN's error signals resembled the neural activity of dopamine neurons measured during the same task. Importantly, the RNN's activity resembled beliefs, but only when the RNN had enough capacity. This work illustrates how animals could estimate value in uncertain environments without needing to first form beliefs, which may be useful in environments where computing the true beliefs is too costly.

How learning unfolds in the brain: toward an optimization view

How do changes in the brain lead to learning? To answer this question, consider an artificial neural network (ANN), where learning proceeds by optimizing a given objective or cost function. This "optimization framework" may provide new insights into how the brain learns, as many idiosyncratic features of neural activity can be recapitulated by an ANN trained to perform the same task. Nevertheless, there are key features of how neural population activity changes throughout learning that cannot be readily explained in terms of optimization and are not typically features of ANNs. Here we detail three of these features: (1) the inflexibility of neural variability throughout learning, (2) the use of multiple learning processes even during simple tasks, and (3) the presence of large task-nonspecific activity changes. We propose that understanding the role of these features in the brain will be key to describing biological learning using an optimization framework.

Learning is shaped by abrupt changes in neural engagement

Internal states such as arousal, attention and motivation modulate brain-wide neural activity, but how these processes interact with learning is not well understood. During learning, the brain modifies its neural activity to improve behavior. How do internal states affect this process? Using a brain-computer interface learning paradigm in monkeys, we identified large, abrupt fluctuations in neural population activity in motor cortex indicative of arousal-like internal state changes, which we term 'neural engagement.' In a brain-computer interface, the causal relationship between neural activity and behavior is known, allowing us to understand how neural engagement impacted behavioral performance for different task goals. We observed stereotyped changes in neural engagement that occurred regardless of how they impacted performance. This allowed us to predict how quickly different task goals were learned. These results suggest that changes in internal states, even those seemingly unrelated to goal-seeking behavior, can systematically influence how behavior improves with learning.

New neural activity patterns emerge with long-term learning

Learning has been associated with changes in the brain at every level of organization. However, it remains difficult to establish a causal link between specific changes in the brain and new behavioral abilities. We establish that new neural activity patterns emerge with learning. We demonstrate that these new neural activity patterns cause the new behavior. Thus, the formation of new patterns of neural population activity can underlie the learning of new skills.

Constraints on neural redundancy

Millions of neurons drive the activity of hundreds of muscles, meaning many different neural population activity patterns could generate the same movement. Studies have suggested that these redundant (i.e. behaviorally equivalent) activity patterns may be beneficial for neural computation. However, it is unknown what constraints may limit the selection of different redundant activity patterns. We leveraged a brain-computer interface, allowing us to define precisely which neural activity patterns were redundant. Rhesus monkeys made cursor movements by modulating neural activity in primary motor cortex. We attempted to predict the observed distribution of redundant neural activity. Principles inspired by work on muscular redundancy did not accurately predict these distributions. Surprisingly, the distributions of redundant neural activity and task-relevant activity were coupled, which enabled accurate predictions of the distributions of redundant activity. This suggests limits on the extent to which redundancy may be exploited by the brain for computation.

When you swing a tennis racket, muscles in your arm contract in a specific sequence. For this to happen, millions of neurons in your brain and spinal cord must fire to make those muscles contract. If you swing the racket a second time, the same muscles in your arm will contract again. But the firing pattern of the underlying neurons will probably be different. This phenomenon, in which different patterns of neural activity generate the same outcome, is called neural redundancy.

Neural redundancy allows a set of neurons to perform multiple tasks at once. For example, the same neurons may drive an arm movement while simultaneously planning the next activity. But does performing a given task constrain how often different patterns of neural activity can be produced? If so, this would limit whether other tasks could be carried out at the same time. To address this, Hennig et al. trained macaque monkeys to use a brain-computer interface (BCI). This is a device that reads out electrical brain activity and converts it into signals that can be used to control another device. The key advantage of a BCI is that the redundant activity patterns are precisely known. The monkeys learned to use their brain activity, via the BCI, to move a cursor on a computer screen in different directions.

The results revealed that monkeys could only produce a limited number of different patterns of brain activity for a given BCI cursor movement. This suggests that the ability of a group of neurons to multitask is restricted. For example, if the same set of neurons is involved in both planning and performing movements, then an animal’s ability to plan a future movement will depend on the one it is currently performing.

BCIs can help patients who have suffered stroke or paralysis. They enable patients to use their brain activity to control a computer or even robotic limbs. Understanding how the brain controls BCIs will help us improve their performance and deepen our knowledge of how the brain plans and performs movements. This might include designing BCIs that allow users to multitask more effectively.