Decoding Alignment without Encoding Alignment: A critique of similarity analysis in neuroscience

A new paper from Bertram, Dyballa, Keller, Kinger, and Zucker (arXiv, May 2026) takes aim at a foundational assumption in computational neuroscience: that similarity in decoding representations (as measured by RSA, DSA, and other alignment metrics) implies similarity in underlying neural computation. The researchers argue this is a critical flaw because such metrics can be driven by just a small subset of neurons that happen to decode information the same way, while the majority of the population may operate very differently. This means conclusions about brain region homology, cross-species comparisons, or brain-to-AI alignment may be fundamentally misleading.

To prove their point, the team ran controlled experiments on both biological neural recordings and deep learning models. In a clear MNIST example, they causally altered the encoding manifold topology (how function is distributed across neurons) by manipulating the training loss, but the decoding alignment metrics remained unchanged. This demonstrates that high representational similarity can exist without shared computation. The authors advocate for a complementary "encoding manifold" approach that characterizes how neurons are organized globally, offering a more complete picture of neural computation and function.