We are excited to announce we have partnered with GRiD member Gabriel Andrade and his team to present a series of talks regarding their ongoing project “Recurrent Systems for Agent Decision Making and EMG-based Motor Control”. At a high level, this project involves designing a robotic agent capable of processing both biological and environmental stimuli to perform some task at the optimal time, involving computer vision, signal processing, time series analysis, and more.
Project Description: We aim to design an agent which is able to continuously (but possibly episodically) coordinate percepts from a dynamic environment and incoming EMG signals to perform some prespecified task. This agent should passively reside in an environment, updating the environment state as needed, until it receives a signal which signifies that it should perform some action (described in the signal). It is important to note that the environment may change with time via external influences, which in turn may change how the task should be performed, but also performing the task may change the environment. This interplay between the environment and the agent makes careful management of incoming data extremely important, not just as individual datum, but as a time series. Our approach stresses reliable and efficient (possibly online) data pre-processing across multiple domains (e.g signals from EMG, percepts, internally produced data) and learning algorithms for decision making and control that “optimize” supervision vs. performance (i.e minimal supervision for “acceptable” performance).
The preprocessing not only requires that we venture into active domains such as computer vision, signal processing, etc., but it also requires that we understand emg signals at a biological level, to some extent, so as to be mindful of the implications of what we extract from the raw signals. This takes the form of accounting for how EMG data is extracted, what exactly it is recording from a subject, the inverse relation of what biologically may lead to some signal, and how subject specific variables may affect the the recording.
The decision making and control of the agent requires that we use techniques capable of utilizing memory/context in time series. Furthermore, due to the varying nature of each sub-task involved in the agent’s performance, we will need to create a recurrently connected network composed of the models which perform individual sub-tasks and inform other models appropriately. On top of this, since we aim to coordinate the continuously varying nature of our environment and the fact that incoming EMG initiates the task, we must either place an emphasis on paradigms fit for time-varying, multi-modal input (e.g Reservoir Computing, Limit Cycle SOMs) or we will need to develop an extremely fine grained system which carefully manages/organizes time scales. To state it in a convoluted yet intuitive way, we must construct a network of recurrent sub-networks that perceive the environment and act on given EMG signals.
This first talk will provide a high-level overview of the project and focus on the biological signal processing required to detect stimuli, beginning at 3:00PM. Check out the event and RSVP on the Western Mass Data Science and Statistics Meetup page. We hope to see you there!