Prefrontal plasticity during learning: transfer of neural changes between active and passive tasks

The study investigates how training on a visuospatial working memory task reshapes prefrontal cortex (PFC) activity and whether such training-induced neural changes transfer to a passive fixation/visual stimulation task that remains unchanged. The central hypotheses are that training would increase neuronal responsiveness and firing rates, alter correlations between neurons (noise correlations), and modify local field potential (LFP) power (notably in beta/gamma bands), and that some of these changes would generalize to the passive task, whereas task-specific decision information (match vs nonmatch) would remain largely absent during passive viewing.

Prior work shows working memory (WM) is malleable and WM training can improve performance with some transfer to untrained tasks (e.g., clinical and behavioral studies). Neurophysiology in trained versus naïve animals has documented increased prefrontal activity, recruitment of more neurons, and changes in LFP bands during WM tasks. Models and EEG/MEG literature link WM maintenance to beta/gamma dynamics. Artificial neural networks provide a framework for transfer learning via changes to hidden representations that generalize across tasks. The study builds on these by longitudinally tracking neural changes through sequential training phases and probing transfer in a passive task.

Subjects: Monkeys were trained over four sequential phases (I–IV) of an active visuospatial match/nonmatch working memory task, interleaved with a passive fixation task with identical stimulus presentations each day. Chronic microelectrode array recordings were obtained from prefrontal cortex across days. Tasks and training phases: Pre-training baseline passive task; Phase I introduced choice targets and reward associations; Phase II introduced match vs nonmatch contingencies (initially block-wise, then randomized); Phase III generalized cues to multiple locations (requiring remembering cue location and comparing to second stimulus); Phase IV increased delay durations (0.25 s up to 1.25 s), amplifying WM demand. Neural data: Single units and multi-unit activity (MUA), plus LFPs. Responsive neurons were identified by comparing firing during stimulus/delay epochs vs fixation baseline (paired t-test, p<0.05), with an additional ≥10% increase criterion for MUA responsiveness. Best (preferred) stimulus location per neuron was determined independently for cue and first delay periods. Analyses: Population PSTHs; baseline and evoked firing rate comparisons across phases using t-tests and ANOVA; day-by-day tracking of MUA per channel. Spike-count (noise) correlation was computed as the Pearson correlation of spike counts during the identical 1 s fixation baseline interval, between simultaneously recorded neuron pairs, analyzed as a function of inter-electrode distance; ANCOVA and ANOVAs tested effects of phase and task. Decoding: Neural Decoding Toolbox used to decode stimulus location and decision (match/nonmatch) from MUA. Only the eight peripheral locations were included; chance levels were 12.5% (location) and 50% (decision). 500 ms bins with 50 ms steps; pseudo-populations of 200 MUAs; 10-fold cross-validation repeated 5 times, over 100 resamples (total 500 estimates). Z-tests compared actual vs shuffled-label performance; additional analyses decoded from the same neurons across tasks (n=72) without resampling. Demixed PCA (dPCA) decomposed population activity into stimulus and decision components; sliding-window 3-way ANOVA (200 ms bins, 50 ms steps) tested encoding of task variables. LFPs: Power spectra were normalized to the pre-cue fixation baseline and compared during cue periods across phases, focusing on alpha (8–14 Hz), beta/low-gamma (20–45 Hz; referred to as beta), and high gamma (46–70 Hz). Grouped daily sessions within phases were compared via 1-way ANOVAs. Data and code: Data available at Mendeley (doi:10.17632/fhx5s7zxg7.2). Acquisition system code on GitHub (ChristosLab/Wave); analysis code upon request.

- Firing rate changes and transfer to passive task: • Passive task: Evoked firing increased across training phases despite unchanged stimuli (1-way ANOVA: cue period F4,285=4.1, p=0.003; first delay period F4,285=5.41, p=0.0003). In Phase IV, delay activity relative to baseline increased in the passive task (two-sided t-test, t(67)=7.17, p=7.49×10^-10), accompanied by a baseline rate decrease. • Active task: During Phase I, activity ramped within trials, from 9.3 spikes/s in the cue period to 12.7 spikes/s at target onset; in Phase II reached 15.4 spikes/s at target onset. Phase II showed an initial rate increase then decrease over training (t(298)=2.12, p=0.035), whereas the passive task did not parallel this (t(65)=0.49, p=0.624). Phase III showed no baseline change (active: t(234)=0.73, p=0.468; passive: t(95)=0.02, p=0.987). Randomized/longer delays in Phase IV abolished ramping and boosted first-delay activity. • Topography-specific passive changes: Responses at a constant passive location changed across phases (e.g., lower-right cue location: cue F3,223=5.89, p=6.97×10^-4; first delay F3,223=7.49, p=8.43×10^-5), reflecting broad network alterations not tethered to a single trained stimulus. - Decoding: • Stimulus location was robustly decoded in both passive and active tasks across phases. • Match/nonmatch decision decoding: Near chance in passive task across phases (mean sample-period accuracy ~0.50–0.53); clearly above chance in active task (sample-period accuracy 0.58, 0.67, 0.58 in Phases II–IV), including when analyzed per location pair. Same-neuron comparisons showed stronger decision decoding in active than passive. • dPCA/ANOVA: Decision-related components were virtually absent in passive activity across all phases; stimulus representations remained stable. - Noise correlations: • Noise correlation decreased with inter-electrode distance and was consistently higher in passive than active tasks. • Training decreased noise correlations predominantly in Phases III–IV and this effect transferred across tasks (Active ANCOVA over phases controlling for distance: F3,2598=10.83, p=4.5×10^-7; Passive: F3,2555=9.37, p=3.7×10^-6). Same-electrode pairs showed significant effects of phase and task (2-way ANOVA; p≈10^-13 for phase, p≈10^-17 for task). - LFP power: • Active task: Beta (20–45 Hz) power during cue presentation decreased progressively with training (highly significant ANOVA; p≈2.27×10^-62), while alpha (8–14 Hz) increased (p≈7.47×10^-53); high gamma generally decreased. • Passive task: Despite unchanged stimuli, beta power also decreased across phases (p≈1.28×10^-24) and alpha increased (p≈1.02×10^-13), paralleling active-task trends. Overall, training increased recruitment and firing rates of responsive neurons, reduced noise correlations, and altered LFP power, with many effects transferring to the passive task. However, decision-related (match/nonmatch) information remained largely specific to the active task.

The findings show that training on a working memory task induces widespread reconfiguration of prefrontal network activity that generalizes to contexts without explicit task demands. Increased responsiveness and firing rate, reduced noise correlations, and systematic LFP beta decreases/alpha increases observed during active task performance also manifested during passive fixation, indicating lasting circuit changes consistent with transfer learning principles. Nevertheless, decision-specific information (match vs nonmatch) emerged robustly only in the active task, underscoring context-dependent neural coding and the unique dynamics associated with active cognitive engagement. The results align with persistent-activity models of spatial working memory, suggesting strengthened prefrontal circuits capable of sustaining delay-period activity after training, and provide a neural substrate for behavioral transfer effects reported in WM training literature.

This longitudinal study demonstrates that working memory training reshapes prefrontal cortex activity in ways that transfer to a passive task: more neurons become responsive with higher firing rates, noise correlations decline, and LFP beta power decreases while alpha increases. Conversely, decision-related codes (match/nonmatch) remain largely confined to active task execution. These results provide a neural basis for transfer learning following WM training and a framework for probing training-induced plasticity. Future work could test generalization to other cognitive tasks, dissect causal mechanisms (e.g., via circuit manipulations), and relate neural changes to individual differences in behavioral gains.

A key consideration is whether passive-task effects reflect covert engagement in the active task. The authors argue this is unlikely given the fixed, daily passive blocks without choice targets and the absence of match/nonmatch decoding in passive data. Some neural changes were transient or phase-dependent, and activity evolved within phases as performance improved. The study focuses on prefrontal cortex in monkeys and relies on chronic recordings and MUA/SUA/LFP analyses; broader generalization across brain areas, species, and task domains remains to be established.