Figure 7. Deep brain stimulation enhances subspace rotation and increases neural response dimensionality.
(A) Experimental design for nucleus basalis stimulation. Intermittent electrical stimulation was applied to the nucleus basalis while single-unit recordings were simultaneously conducted in the prefrontal cortex during the context-dependent working memory task. Stimulation was applied during the inter-trial interval, every few trials. (B) Mean firing rate during the control and NB stimulation conditions (n=106 neurons). (C) Single-cell selectivity during the first delay epoch for the two spatial locations was quantified using the Selectivity Index, SI = (R1 − R2) / (R1 + R2), in both the Remember 1st and Remember 2nd contexts. NB stimulation alters single-cell selectivity by rendering neuronal responses context dependent. (D) Mean subspace rotation for four consecutive task epochs: first stimulus presentation, first delay period, second stimulus presentation, and second delay period, under control and NB-stimulation conditions. Deep brain stimulation induces greater subspace rotation. (E) Mean dimensionality during the two stimulus and two delay period under control and NB-stimulation conditions. Deep brain stimulation induces higher dimensionality, quantified by the participation ratio of neural responses. (F) Analysis of decoding performance in correct and error trials in the Remember 1st context, where most errors occur. Decoding performance for the location of the first stimulus is shown for both contexts under control and stimulation conditions. In control trials, errors were characterized by reduced context-dependent gating of information after the second stimulus, whereas in stimulation trials, errors were characterized by early differentiation of information. (G) Change in the mean information conveyed by each spike in the NB-stimulation relative to the control condition. Red vertical lines in the insets indicate the mean information change, whereas the black curves represent the null distribution obtained by label shuffling. Under NB stimulation, less context-relevant information was conveyed by each spike. (H) Cosine similarity between empirical decoding results and the predictions of the ordinal, pre-allocate, and overwrite models.