Forcing low-dimensional representations boosts AI generalization, study finds

A new arXiv paper from Hardik Rajpal and Dan Goodman provides strong evidence that low-dimensional representations in neural networks are not just a reflection of neuron-level activity but confer a functional advantage for generalization. By applying an explicit information bottleneck to a recurrent neural network (RNN), they forced the network to learn compact representations. This was shown to be necessary for the model to achieve rotational and out-of-distribution generalization on a time-series prediction task—abilities that elude networks without such constraints.

Using information-theoretic measures of causal emergence, the researchers characterized the dynamics of these representations across the memorization-to-generalization transition. They found a non-monotonic trajectory: the emergent structure initially decreased, hit a minimum, and then rose to a maximum, even as prediction loss fell monotonically. This trajectory scaled with task complexity, and the magnitude of emergent structure reliably predicted generalization performance. The non-monotonic pattern suggests that the network first compresses information before building generalizable representations.

Crucially, the team validated their findings in biological neural networks. They analyzed CA1 hippocampal activity in mice learning an alternating maze task and discovered analogous non-monotonic emergence dynamics that tracked behavioral performance. This cross-domain alignment—from artificial RNNs to mouse hippocampus—supports a causal role for distributed, compact representations in learning and generalization. The work bridges AI and neuroscience, offering a principled explanation for why explicit representation learning is critical in both engineered and biological systems.