#23 · ML / AI methods

Learned neural components tested for mechanistic and causal validity beyond prediction

The dominant thread is whether representations learned from neural and cognitive data correspond to the mechanisms they are meant to capture. The multitask study maps how 100 cognitive tasks transfer or interfere within shared hierarchical almost-linear RNN dynamics, and the neuronal identity proposal organizes population dynamics around molecular, morphological, and electrophysiological profiles, a mechanistically motivated basis for latent structure. Other items remain predictive or correspondence-based: the Parkinson's classification and MtReNIRS studies report predictive and transferable performance, the predictive-coding networks are assessed by ventral-stream alignment, and the latent-prediction fMRI encoders show cross-participant alignment for decoding. The emphasis on mechanistic organization rather than predictive accuracy alone distinguishes the multitask and neuronal identity studies, though the current evidence characterizes representational correspondence rather than causal manipulation.