Abstract:

Cognitive flexibility is a fundamental ability that enables humans and animals to exhibit appropriate behaviors in various contexts. The thalamocortical interactions between the prefrontal cortex (PFC) and the mediodorsal thalamus (MD) have been identified as crucial for inferring temporal context, a critical component of cognitive flexibility. However, the neural mechanism responsible for context inference remains unknown. To address this issue, we propose a PFC-MD neural circuit model that utilizes a Hebbian plasticity rule to support rapid, online context inference. Specifically, the model MD thalamus can infer temporal contexts from prefrontal inputs within a few trials. This is achieved through the use of PFC-to-MD synaptic plasticity with pre-synaptic traces and adaptive thresholding, along with winner-take-all normalization in the MD. Furthermore, our model thalamus gates context-irrelevant neurons in the PFC, thus facilitating continual learning. We evaluate our model performance by having it sequentially learn various cognitive tasks. Incorporating an MD-like component alleviates catastrophic forgetting of previously learned contexts and demonstrates the transfer of knowledge to future contexts. Our work provides insight into how biological properties of thalamocortical circuits can be leveraged to achieve rapid context inference and continual learning.

Fig. 1 a. A cortico-thalamic neural network model with a Hebbian learning rule in the PFC-to-MD connections to infer temporal context and MD gating in the PFC. b. The challenge in continual learning. Standard artificial neural networks modeling one brain region are optimized on single contexts and suffer from catastrophic forgetting. The learned critical model parameters of old contexts are changed in new contexts. The right arrows denote the learning process with context switch. c. We propose a synaptic plasticity with pre-synaptic and post-synaptic traces, adaptive thresholding, and winner-take-all in the MD to make the neural network infer temporal contexts and enable continual learning.

Fig.2 a. An attention-guided behavioral task with two cueing contexts presented in blocks. b. Neural networks are trained in similar alternating blocks. Cues 1 and 2 are used in context 1, while cues 3 and 4 used in context 2. Block 3 returns to context 1 to evaluate the model performance in continual learning. c. The PFC and the MD neuronal activities in one single early trial of context 1 and context 2. d. Decoding context (left) and rule (right) in model PFC (red) and MD (blue). MD contains information about the current context, but not the current rule (cue). Shaded area: cue period. e. The decoding performance of context (brown) and rule (green) in the MD with varying numbers of cycles during training. Each cycle consists of two trials of cue 1 and cue 2 during context 1 and two trials of cue 3 and cue 4 during context 2, in random order within each cycle. The MD selectively encodes temporal contexts within a few trials. Shaded areas denote the standard deviations. f. There are different clusters in the t-SNE maps of the PFC and the MD neuronal activities. Both rule and context can be decoded from the PFC while only context can be decoded from the MD. Red and blue colors denote different rules while green and orange colors represent different contexts, respectively.

Fig. 3 a. The MD effects on the PFC neurons, enhancing current-context PFC activities and inhibiting other-context PFC activities.b. The model performance (mean square error, MSE) of the PFC-MD model and the PFC only model during training in the attention-guided behavioral task. c. The model performance of the third block when the learned context was switched back. In comparison with the PFC only model, the PFC-MD model had low prediction errors after context switch, alleviating catastrophic forgetting. d. The experimental data were collected when mice performed similar three-block switching tasks. The left and middle boxplots show the effect of bilateral MD suppression on behavioral performance. The right boxplot shows the comparison of performance on the consecutive sessions. ***P < 0.001, Bonferroni-corrected rank-sum test. Figure 6D adapted from Fig. 5F of Rikhye, R.V., Gilra, A. & Halassa, M.M. Thalamic regulation of switching between cortical representations enables cognitive flexibility. e. The model performance of the PFC-MD model and the PFC model. MD suppression significantly degraded the performance when the model switched back to the previous context. ***P < 0.001, Bonferroni-corrected rank-sum test. f. The boxplots of the changes in connection weights from the current-context and the other-context PFC neurons to the output neurons during Context 1 and Context 2 presentations. **P < 0.005, ***P < 0.001; statistical test with analysis of variance (ANOVA). Adding an MD component protected synaptic weights in neurons that were not currently context-relevant, which reduces interference in model parameters across different temporal contexts. g. The mean performance of the PFC-MD model learning two cognitive tasks in Neurogym sequentially. Orange and green colors represent Task 1 and Task 2, respectively. Shaded areas denote the standard deviations. h. We compared the PFC-MD model with two other continual learning methods: EWC and SI, and the PFC only model (left: task 1 performance, right: task 2 performance). The PFC-MD model obtained the best model performance among these methods. i. The mean model performance of the PFC-MD model with more cognitive task learned sequentially. The PFC-MD model could flexibly switch between different tasks without forgetting.

DOI：10.1038/s41467-024-52289-3