Avoiding Catastrophe: Active Dendrites Enable Multi-Task Learning in Dynamic Environments.

A key challenge for AI is to build embodied systems that operate in dynamically changing environments. Such systems must adapt to changing task contexts and learn continuously. Although standard deep learning systems achieve state of the art results on static benchmarks, they often struggle in dynamic scenarios. In these settings, error signals from multiple contexts can interfere with one another, ultimately leading to a phenomenon known as catastrophic forgetting. In this article we investigate biologically inspired architectures as solutions to these problems. Specifically, we show that the biophysical properties of dendrites and local inhibitory systems enable networks to dynamically restrict and route information in a context-specific manner. Our key contributions are as follows: first, we propose a novel artificial neural network architecture that incorporates active dendrites and sparse representations into the standard deep learning framework. Next, we study the performance of this architecture on two separate benchmarks requiring task-based adaptation: Meta-World, a multi-task reinforcement learning environment where a robotic agent must learn to solve a variety of manipulation tasks simultaneously; and a continual learning benchmark in which the model's prediction task changes throughout training. Analysis on both benchmarks demonstrates the emergence of overlapping but distinct and sparse subnetworks, allowing the system to fluidly learn multiple tasks with minimal forgetting. Our neural implementation marks the first time a single architecture has achieved competitive results in both multi-task and continual learning settings. Our research sheds light on how biological properties of neurons can inform deep learning systems to address dynamic scenarios that are typically impossible for traditional ANNs to solve.

Toward Real-Time Muscle Force Inference and Device Control via Optical-Flow-Tracked Muscle Deformation.

Despite the utility of musculoskeletal dynamics modeling, there exists no safe, noninvasive method of measuring in vivo muscle output force in real time - limiting both biomechanical insight into dexterous motion and intuitive control of assistive devices. In this paper, we demonstrate that muscle deformation constitutes a promising, yet unexplored signal from which to 1) infer such forces and 2) build novel device control schemes. Through a case study of the elbow joint on a preliminary cohort of 10 subjects, we show that muscle deformation (specifically, thickness change of the brachioradialis, as measured via ultrasound and tracked via optical flow) correlates well with elbow output force to an extent comparable with standard surface electromyography (sEMG) activation during varied isometric elbow contraction. We then show that, given real-time visual feedback, subjects can readily perform a trajectory tracking task using this deformation signal, and that they largely prefer this method to a comparable sEMG-based control scheme and perform the tracking task with similar accuracy. Together, these contributions illustrate muscle deformation's potential utility for both biomechanical study of individual muscle dynamics and device control, in a manner that - thanks to, unlike sEMG, the localized nature of the signal and its tight mechanistic coupling to output force - is readily extensible to multiple muscles and device degrees of freedom. To enable such future extensions, all modeling, tracking, and visualization software described in this paper, as well as all raw and processed data, have been made available on SimTK as part of the Open-Arm project (https://simtk.org/projects/openarm) for general research use.