Conclusions about Neural Network to Brain Alignment are Profoundly Impacted by the Similarity Measure
Soni, A.; Srivastava, S.; Khosla, M.; Kording, K. P.
Show abstract
Deep neural networks are popular models of brain activity, and many studies ask which neural networks provide the best fit. To make such comparisons, the papers use similarity measures such as Linear Predictivity or Representational Similarity Analysis (RSA). It is often assumed that these measures yield comparable results, making their choice inconsequential, but is it? Here we ask if and how the choice of measure affects conclusions. We find that the choice of measure influences layer-area correspondence as well as the ranking of models. We explore how these choices impact prior conclusions about which neural networks are most "brain-like". Our results suggest that widely held conclusions regarding the relative alignment of different neural network models with brain activity have fragile foundations.
Matching journals
The top 6 journals account for 50% of the predicted probability mass.
Similar papers in this journal
Similar papers in this journal
- Stacked regressions and structured variance partitioning for interpretable brain maps 96%
- Class imbalance should not throw you off balance: Choosing the right classifiers and performance metrics for brain decoding with imbalanced data 95%
- Comparing Functional Connectivity Matrices: A Geometry-Aware Approach applied to Participant Identification 95%
Similar papers in this journal
Similar papers in this journal
- Family Lexicon: using language models to encode memories of personally familiar and famous people and places in the brain 94%
- Small hand-designed convolutional neural networks outperform transfer learning in automated cell shape detection in confluent tissues 94%
- Analysing linear multivariate pattern transformations in neuroimaging data 94%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.