Orthogonal representations for robust context-dependent task performance in brains and neural networks
Timo Flesch, Keno Juechems, Tsvetomira Dumbalska, Andrew Saxe, Christopher Summerfield
Abstract
(Neuron 110, 1258–1270.e1–e11; April 6, 2022) We thank the authors of Mante et al. (2013) for alerting us to a labeling error in the presentation of our reanalysis of their data. Unfortunately, we switched the labels for "colour" and "motion" in Figures 4 and 6 as well as Figures S4 and S5, which provoked a discrepancy with the findings presented in the original report. The figures have now been corrected online. We apologize for any confusion this may have caused.Figure 4. Task representations in human fMRI and macaque unit recordings (original)View Large Image Figure ViewerDownload Hi-res image Download (PPT)Figure 6. Neural network and NHP data in support of gating theory (corrected)View Large Image Figure ViewerDownload Hi-res image Download (PPT)Figure 6. Neural network and NHP data in support of gating theory (original)View Large Image Figure ViewerDownload Hi-res image Download (PPT)Figure S4. Control analyses on the human fMRI and NHP datasets (corrected)View Large Image Figure ViewerDownload Hi-res image Download (PPT)Figure S4. Control analyses on the human fMRI and NHP datasets (original)View Large Image Figure ViewerDownload Hi-res image Download (PPT)Figure S5. Gating in MLPs, NHPs and RNN model (corrected)View Large Image Figure ViewerDownload Hi-res image Download (PPT)Figure S5. Gating in MLPs, NHPs and RNN model (original)View Large Image Figure ViewerDownload Hi-res image Download (PPT) Orthogonal representations for robust context-dependent task performance in brains and neural networksFlesch et al.NeuronJanuary 26, 2022In BriefHow do brains represent multiple tasks? The authors use artificial neural networks to derive a computational theory of neural coding and validate its predictions in recordings of human and macaque brains. The findings suggest that task-specific information is represented along orthogonal coding axes, which minimizes interference and maximizes robustness to noise. Full-Text PDF Open Access