Fast Learners Have High-Dimensional Representations of Neural Activity

Post by Shireen Parimoo

What's the science?

People naturally group new pieces of information into categories to differentiate between them, eliciting distinct patterns of brain activity. One measure of quantifying how information is coded in the brain is by estimating the geometric representation or dimensionality of neural responses. For example, stimuli can be coded in a psychological dimension and represented based on their size and color. Patterns of brain activity can be represented in a high dimension, meaning that the information is coded along many dimensions and that neural responses to each stimulus are likely more distinct from one another. Alternatively, low-dimensionality of neural representations would suggest that there are fewer dimensions along which new information can be coded in the brain, resulting in less distinction between stimuli. Is there a relationship between the dimensionality of neural responses and how fast people can learn new information? This week in Nature Neuroscience, Tang and colleagues used functional magnetic resonance imaging (fMRI) to examine how the dimensionality of neural representations is related to the speed with which people learn new information.

How did they do it?

Nineteen young adults completed various tasks for four days while undergoing fMRI scanning. On each day, they first completed a value-learning task, in which they learned the monetary value of twelve stimuli that ranged from $1 and $12. They were then tested on the stimulus-value associations in a value-judgment task and their response accuracy was used as a measure of learning effectiveness. In the value-judgment task, participants were shown the stimuli and had to indicate if their monetary value was in the top 50% ($7-$12) or the bottom 50% ($1-$6) of the range. In the fMRI data, dimensionality was estimated from label assortativity and separability values of the neural responses. Label assortativity provides a measure of how easily neural responses can be distinguished from each other. In general, higher assortativity is associated with greater dimensionality, but it is also possible for high dimensional representations to have low assortativity. Assortativity was calculated using a linear support vector machine, which is a machine learning technique used to classify information into categories. On the other hand, separability values provide a more direct measure of dimensionality; this was calculated by assigning binary labels to neural activity for each stimulus and using a support vector machine to classify the neural responses according to these labels. High separability values indicate that neural responses are organized in a larger number of dimensions, whereas low values indicate lower dimensionality of representations in neural responses.

The authors examined the relationship between dimensionality and how effectively participants learned the stimulus-value associations. They also conducted a virtual lesioning analysis, in which neural activity in specific brain regions was systematically excluded from the analysis to see which regions contributed to the association between separability and learning effectiveness. Finally, as high dimensionality of neural responses is associated with greater reliance on neural resources, the authors investigated whether resources are allocated efficiently through a balance between high- and low-dimensionality of representations during learning. Participants’ response accuracy was correlated with the dimensionality of representations to specific stimuli (stimulus dimension) during learning and to neural responses independent of the stimulus (embedding dimension), which represents the overall dimensionality of participants’ neural representations.

What did they find?

Response accuracy on the value-judgment task improved over the course of four days, and fast learners – participants who performed above average on the first day – had high accuracy on subsequent days. Assortativity of brain activity was higher among participants who learned fast. That is, stimulus-specific neural responses of fast learners could be distinguished from each other more easily than that of slower learners. Fast learners also had higher dimensionality of neural representations, which is another indication that their representations of the different stimuli were more distinct from one another. In fact, high performance on the first day of the experiment was correlated with high dimensionality on the fourth day, suggesting that effective learning facilitates more precise coding of information in the brain. Interestingly, there was a positive correlation between response accuracy and dimensionality in the stimulus dimension, but a negative correlation in the embedding dimension. This means that fast learners allocate neural resources efficiently by representing all stimuli in an overall low dimensional space, but within that low-dimensional embedding space, specific stimulus-value associations are represented by higher dimensional, and thus more distinct patterns of activity.


What's the impact?

This is the first study to demonstrate how the efficiency of neural responses is optimized during learning, depending on whether it is associated with learning stimulus associations or independently of specific stimuli. The approach of quantifying the dimensionality of representations in neural responses and the finding that dimensionality is correlated with behavioral performance both have important implications for our understanding of how information is represented in the brain and how that affects behavior.

Tang et al. Effective learning is accompanied by high-dimensional and efficient representations of neural activity. Nature Neuroscience (2019). Access the original scientific publication here.