How Our Brains Perceive a Changing World
Post by Anastasia Sares
What's the science?
We gather evidence to make predictions about the world. Was that a bat flying in the darkness? Am I smelling a skunk, or someone smoking nearby? Is that person smiling or are they in pain? Classical models of evidence accumulation assume that the world stays the same as we gather information about it. According to this, we would accumulate evidence in a linear way: all information has the same value, and the longer we examine something the more certain we become. But the real world is constantly in motion, and our information can quickly become out of date. To better explain how brains accumulate evidence, we need a non-linear model that takes into account a changing world. This week in Nature Neuroscience, Murphy and colleagues showed how people update their predictions in this non-linear way.
How did they do it?
The authors created a simple task, in which they asked participants to look at small shapes appearing and disappearing on a screen. These shapes appeared in different positions, but with enough observation, it was evident that they were coming from a single source (like a spray of drops from a sprinkler). On each trial, participants were supposed to watch the shapes and indicate whether this source was on the right or left side of the screen at the end. However, there was a twist—the source would sometimes change positions partway through a trial. Participants’ neural activity was recorded with magnetoencephalography (MEG) while they performed the task, and the size of their pupils was monitored to track changes in mental arousal.
What did they find?
The authors compared the participants’ performance to different artificial models to see how well the models fit. There were two very good models, one which was more mathematical, and one that was based on a circuit of interconnected neurons.
The mathematical model (a Bayesian ‘ideal observer’ model) updates its predictions after each piece of evidence based on what information was encountered recently, the current uncertainty, and also the expectation that the source would be changing. It performed much better than simpler mathematical models that gave equal weight to the evidence from all previous observations. The circuit-based model consisted of just three groups of neurons, which interact and compete with each other through excitatory and inhibitory connections. This circuit made very similar predictions to the mathematical (Bayesian) model, without being told how to do so.
Over the course of a trial, participants showed widespread brain activity that closely tracked their accumulated evidence predicted by the models. This was seen in parts of the brain linked to action preparation, and also—surprisingly— in early sensory brain areas. This could indicate that the sensory areas were receiving feedback from areas responsible for decision-making and action preparation (this is known as a “top-down” influence). Participants’ pupils also dilated when there was a likely change in the source location, and this dilation predicted changes in brain activity. Altogether, the findings indicated that non-linear evidence accumulation is implemented in brain circuits that are distributed across the brain, and shaped by changes in physiological arousal state.
What's the impact?
This work demonstrates that we don’t treat all sensory information the same—our expectations and biases matter, and they affect how we interpret the world by exerting “top-down” influences. Being able to represent these factors in a mathematical or circuit-based way is difficult, but it is an important step forward. By showing that the mathematical (Bayesian) model and the circuit model both make similar predictions, and by identifying signatures of the underlying processes in human brain activity, the authors hope that their work can bridge the gap between researchers with different analytical approaches.
Murphy et al. Adaptive circuit dynamics across human cortex during evidence accumulation in changing environments. Nature Neuroscience (2021). Access the original scientific publication here.