What is it about?
We study mathematically the optimal way to make decisions between a finite number of options, when evidence in favor of each option is ambiguous. Importantly, the correct option changes in time, a situation commonly faced by humans and other organisms in the natural world. Most work in neuroscience has focused on perceptual decisions, in which organisms accumulate evidence in order to choose the best behavioral response. We build upon previous results, originally established in the case of a fixed environment, for which the correct choice is constant during an observational period. When the environment is in flux, the observer must learn the environmental change rate in order to correctly discount older irrelevant information. We describe the optimal algorithm to achieve this, and a plausible way in which this computation could be approximated by the brain.
Featured Image
Why is it important?
Optimal decision making algorithms provide neuroscientists with a benchmark, against which, animal and human performance may be measured. By extending existing results to the changing environment, we get one step closer to understanding how our brain copes efficiently with its ever-changing surroundings. Humans and other organisms are capable of complex, adaptive evidence accumulation. Understanding how the brain implements these computations requires models that can link behavior to the architecture and activity of neuronal networks. In turn, such understanding will allow us to build better artificial systems, capable of performing real-time decisions better than humans in specific tasks.
Perspectives
I believe that this publication is an example of a new trend in decision science, to tackle more realistic scenarios. The problem addressed in this article presents challenges, both at the mathematical, and the neurophysiological levels. Making progress in this field will unravel many mysteries about the computations implemented by our brain. One very general open question is: How are we able to respond in such an adapted way to so many complex situations?
Adrian Ernesto Radillo
University of Houston Charter School
Decision-making is central to organismal behavior, and the natural world is in constant flux. Therefore it is imperative to develop new models that account for the remarkable ability organisms have to adaptively accumulate evidence to make decisions. Our work fundamentally contributes to the development of inference models of decision-making, and builds a bridge between these models and neuronal network models. Not only did we develop an inference model for evidence accumulation in a changing environment, we were able to reduce this to a low-dimensional model that captured the essential features of the full optimal algorithm. This low-dimensional model could then be mapped to a neuronal network with long term plasticity, where the weights between populations represent the rates of change of the environment. These low-dimensional models will be essential for interpreting neural and behavioral recordings from subjects performing decision-making tasks in dynamic environments.
Zachary Kilpatrick
University of Colorado
Read the Original
This page is a summary of: Evidence Accumulation and Change Rate Inference in Dynamic Environments, Neural Computation, June 2017, The MIT Press,
DOI: 10.1162/neco_a_00957.
You can read the full text:
Resources
Learning the volatility of a dynamic environment
Presentation at BIRS Workshop on Connecting Network Architecture and Network Computation
Code for simulations
The folder NeuralCompCode of this GitHub repository contains the code used to produce the first four figures of the paper. More explanation is provided in the README.md file.
Cosyne 2017 Poster
Scientific poster presented at the Cosyne conference in 2017.
Contributors
The following have contributed to this page
