A mental forward model for error monitoring during motor planning
Harvard-MIT Division of Health Sciences and Technology, Bioastronautics Training Program
Because movements are inherently variable, even the simplest action can deviate from what was originally intended. This poses a fundamental question: to what extent do we/can we evaluate if a self-generated movement unfolded as planned? During the movement, the brain can use predictive strategies (known as internal forward models) combined with ongoing sensory feedback to assess any departure from the target action. However, a large part of the variability in our movements arises from central disturbances (i.e., noise) during planning, that is, before the movement has even been initiated and in the absence of any external feedback. Whether the brain can internally monitor this planning variability is therefore unclear. To address this question, we sought to test the hypothesis that humans possess trial-by-trial motor confidence about their planned movements. We developed a context-dependent motor timing task in which subjects had to implicitly use motor confidence to detect binary changes in the context and switch their behavioral strategy accordingly. Our behavioral results revealed that subjects do estimate their trial-by-trial planning variability and use the associated confidence to update their strategy across trials. By analogy with the predictive processes the brain employs to evaluate ongoing movements, we hypothesize that the ability to monitor planning errors relies on a mental forward model that uses a copy of the motor command to predictively simulate the motor plan. We implemented this idea in a neurally-plausible model for motor planning and were able to recapitulate confidence-based behavior of our human subjects. Moreover, the model made several predictions that could be verified empirically. The idea that forward models are established as early as during motor planning opens an avenue for bridging the fields of motor control and metacognition.
***** ***** *****
Representations of time-dependent behavior in working memory
Department of Brain and Cognitive Sciences
One hypothesis of working memory is that objects or contexts can be represented by persistent activity in the brain. This can be modeled by dynamical systems with fixed points encoding memories. While this model is compelling, it fails to account for how the brain might hold a time-dependent stimulus in working memory. In this talk, I will discuss my work in using recurrent neural networks (RNNs) to generate hypotheses about how representations of time might be stored in working memory. I will present results from an experiment where subjects had to remember a time interval across a delay, and then reproduce it. I used a recently published algorithm to train an RNN on the same task, which provided a hypothesis with predictions for human behavior. I will also point out some limitations of this paradigm and propose some alternative ways for thinking about similar tasks.