Written by
Caroline Jahn
March 2, 2021

In February 2021, the Pillow lab published a new study “Extracting the dynamics of behavior in sensory decision-making experiments” in Neuron in collaboration with Ji Hyun Bak (UCSF), Athena Akrami (PNI, now UCL) and Carlos Brody (PNI). Led by Nick Roy, who now works at DeepMind, this study presents a publicly available tool called PsyTrack for tracking the trial-by-trial behavioral strategy during decision-making. PsyTrack is packaged in a Google collaboration Notebook for easy download and use.

This study is part of a collaboration of twenty-one neuroscience groups from the US, UK, France, Switzerland and Portugal called the International Brain Laboratory. Funded by the Simons Foundation and the Wellcome Trust, this initiative aims to understand how the brain controls learning and decision making. The different groups work together to understand a single behavior in mice and will record neural activity in many brain regions using NeuroPixels, an electrode that can record hundreds of neurons simultaneously. They use a single open platform for conducting experiments and organizing data across labs. The tool developed by Roy and colleagues enables the standardization of the behavioral analysis for better reproducibility.

The International Brain Laboratory task consists of a series of striped disks that are displayed on either the right or the left side of the screen. Mice are trained to turn a steering wheel that controls the position of the stimulus. If they bring the stimulus into the center of the display, they are given a reward (see Fig. 1A). The difficulty of the task depends on the contrast of the stripes in the stimulus. When the contrast is high, it is very easy to see where the stimulus is, but not when it is low. After the mice have learned the task, the probability that they turn the wheel to the right is well-modelled by a psychometric curve, showing that on average mice make fewer errors when the contrast in high (see Fig. 1B).

The Neuro Resource Figure 1

Roy and colleagues saw an opportunity to better understand the learning process by examining the dynamics of the mice behavior throughout training. Most animal experiments in decision-making involve over-trained animals. Mice perform this task thousands of times until their strategy is fixed. Of course, this leaves out interesting information about how the mice learn. Early in training, the mice use a simplistic strategy: always steer in the same direction regardless of the stimulus. As learning progresses, the mice begin to use the stimuli to inform their choice, eventually converging on the psychometric curve of a fully trained mouse. Using conventional tools, neuroscientists were unable to track this dynamic change of strategy. By modelling the influence of the strategy on choices as a smooth drift over time, PsyTrack tracks the learning of the correct strategy (blue and red curves in Fig. 2) and the forgetting of the simplistic one (yellow curve in Fig. 2).

The Neuro Resource Figure 2

PsyTrack is a versatile tool and can be used to model learning in a variety of tasks. In this study, the researchers were also able to examine the strategies used by rats and humans in an auditory discrimination task originally published by Arkami and colleagues (Nature, 2018). Subjects heard one tone after another and had to decide which was louder. If the first was louder, they had to respond right, if the second was louder, they had to respond left (see Fig. 3). Roy and colleagues showed that rats tend to use various erroneous strategies at the start of learning (such as always choosing the same side or repeating a choice that has been rewarded), while humans do not. This can be explained by the fact that humans are told what the correct strategy is, whereas rats had to discover it through trial and error.

The Neuro Resource Figure 3

Being able to track the strategies used during learning opens many exciting possibilities in the study of behavior. Researchers are now better able to investigate the neural correlates of dynamic learning processes. PsyTrack could also be used to optimize training, intelligently choosing the conditions that will quickly guide the subject towards the desired strategy.  As neuroscientists refine our understanding of learning and its neural correlates, new tasks will need to be developed. The authors hope that this tool can give researchers reliable and easy access to their subject’s underlying strategies in real-time.