How much neuroscience does a neuroscientist need to know?

TL;DR

A lot, but surprisingly few details matter; this perspective aligns computational neuroscience with mechanistic interpretability in AI, suggesting a more unified approach to studying the mechanisms of intelligence, both natural and artificial.

Abstract

How much of the brain's learned algorithms depend on the fact it is a brain? We argue: a lot, but surprisingly few details matter. We point to simple biological details -- e.g. nonnegative firing and energetic/space budgets in connectionist architectures -- which, when mixed with the requirements of solving a task, produce models that predict brain responses down to single-neuron tuning. We understand this as details constraining the set of plausible algorithms, and their implementations, such that only `brain-like' algorithms are learned. In particular, each biological detail breaks a symmetry in connectionist models (scale, rotation, permutation) leading to interpretable single-neuron responses that are meaningfully characteristic of particular algorithms. This view helps us not only understand the brain's choice of algorithm but also infer algorithm from measured neural responses. Further, this perspective aligns computational neuroscience with mechanistic interpretability in AI, suggesting a more unified approach to studying the mechanisms of intelligence, both natural and artificial.

Figures (3)

• Figure 1: Hypothesis that few biological details constrain the space of effective algorithms.Left: The space of potential algorithms taking place in the brain is constrained by computation and biological details. Right: We contend that, conditioned on being a neural network, relatively few further biological details constrain the optimal algorithm.
• Figure 2: Intuition for how biological details can constrain neural response. Two uniformly distributed independent factors represented with two entangled neurons (left). The neural population can be made nonnegative at the expense of activity energy (middle). Activity energy is minimised under a nonnegativity (and variance) constraint when the neurons are axis aligned to task factors (i.e. disentangled, right).
• Figure 3: Twisted XORA: The four 3D datapoints in the twisted XOR task. In the (x,y) plane this is the classic XOR problem, whereas the z direction simply encodes the label. $\Delta$ measures the size of the $z$ direction. There are two viable solutions, B: using two neurons to map the z direction directly to labels, or C: using one neuron per datapoint to solve the task as an XOR. Which is learnt depends on the size of $\Delta$. Figure from the original work of Jarvis et al.jarvis_make_2025.