a mouse it never saw

Neurons do not carry over between animals, so each session is projected into its own 16-dimensional population space and the spaces are rotated into one frame using responses to the shared movies. The dynamics model is then trained on seven mice and asked to forecast the eighth.

latent var

Thin lines: each held-out mouse. Thick line: the mean. The reference line is the ceiling set by the 16-dimensional bottleneck itself — the decoder fed the true next latent (mean R² 0.216). The target’s alignment uses its training period only, so this is few-shot transfer, never zero-shot.

latent gru

did the alignment hold

Alignment quality is the correlation between a target’s rotated movie response and the reference built from the other mice. The training-period figure is what the rotation was fitted to. The test-period figure is a diagnostic only — nothing is fitted to it — asking whether the same rotation still lines the mouse up an hour later.

Per held-out mouse
Target session	Units	Alignment, train	Alignment, test (diagnostic)	Latent VAR, within	Latent VAR, cross aligned	Latent VAR, cross unaligned	Latent GRU, within	Latent GRU, cross aligned	Latent GRU, cross unaligned
715093703	120	0.725	0.653	0.090	0.083	0.086	0.093	0.082	0.085
719161530	120	0.712	0.503	0.095	0.090	0.091	0.098	0.088	0.087
721123822	120	0.747	0.655	0.105	0.099	0.100	0.108	0.097	0.100
732592105	120	0.752	0.644	0.113	0.107	0.107	0.116	0.104	0.105
737581020	96	0.663	0.461	0.088	0.068	0.068	0.094	0.068	0.079
739448407	120	0.669	−0.024	0.066	0.052	0.055	0.073	0.054	0.054
744228101	120	0.725	0.399	0.073	0.059	0.065	0.082	0.070	0.072
754312389	120	0.694	0.600	0.098	0.091	0.090	0.104	0.092	0.091