In this blog post I explore possible theoretical implications of the finding that trained neural networks tend to show similar internal representations. I riff off of two recent papers; Chen & Bonner (2024), and Huh et al. (2024). In this blog post, I focus mostly on the former.
Testing network convergence and brain likeness
Consider the following idealized approach. Let’s consider all neural networks that possibly capture some aspects of how the brain realizes a target cognitive function F. Let’s call this set of networks set A. Next, after training each network using a task objective that operationalizes F, let’s consider every latent dimension that is causally relevant in task performance. To what extent are such latent dimensions shared across networks in our set, and to what extent are these dimensions like the ones the brain employs when it realizes F? Chen & Bonner (2024) refer to the former question as one of universality, and the latter as one of brain similarity.
Maybe neuroscience is sort of like an engineering project where we build bridges from brain to mind. On our side of the river lie the nuts and bolts of the brain, and on the other lies our psychology. Building the bridge is about explaining how the former realizes the latter.
In the case of language comprehension, over at the other riverbank we see rich hierarchical representations of meaning and syntax. The exciting project for us is to figure out what it is about the brain that can transform squiggly air into such linguistic structures. At our pier we have neurons, networks, and anatomical regions, but also dynamical phenomena.
We’ve just published the paper accompanying the Brain Time Toolbox in Nature Human Behaviour (page; PDF). This post serves to give a brief introduction to the toolbox, explaining its raison d’être and how it works. In a sentence, it transforms electrophysiology data in accordance with internal brain dynamics to facilitate our analysis of the cognitive brain.
Let’s start with three observations. First, electrophysiology data is formatted and analyzed in clock time—that is to say, the data unfolds millisecond by millisecond. Second, the brain is mostly indifferent to clock time and instead uses neural oscillations to coordinate the intricate patterns underlying cognition. Third, such neural oscillations are highly dynamic—they speed up, slow down, reset themselves spontaneously, and start at different phases across trials.
If you’re in a room with neuroscientists and somebody brings up philosophy, you can expect at least a few scoffs and eye rolls. I can see where this sentiment comes from. We’re strained from trying to make sense of our data as it stands, so to hear that we’re using the word “representation” wrong is frustrating – especially if it’s unclear whether it changes anything at all about our theorizing. Add philosophical jargon which floods the mind with syllables, and you’ve lost most.