The ability of biological brains to continuously learn and integrate new knowledge contrasts sharply with the catastrophic forgetting exhibited by artificial neural networks (ANNs). ANNs often achieve optimal performance on newer tasks while losing proficiency in previously learned ones. This discrepancy highlights the stability-plasticity dilemma: networks must be plastic enough to adapt to new tasks but stable enough to retain existing knowledge. While ANNs demonstrate impressive capabilities in various domains, from complex games to image recognition, they remain suboptimal in navigating the stability-plasticity spectrum. Sleep is posited as crucial for memory consolidation and knowledge generalization in biological systems. During sleep, spontaneous neuronal activity generates complex patterns of synchronized activity across brain regions. Two essential components believed to underlie memory consolidation during sleep are spontaneous replay of memory traces and local unsupervised synaptic plasticity. Replay of recent and relevant older memories allows the network to create orthogonal memory representations, promoting the coexistence of competing memories. Local plasticity ensures that synaptic changes impact only the relevant memories. This research focuses on hippocampus-independent memory consolidation, mirroring REM sleep-like activity. The study proposes that integrating a sleep-like phase after an ANN learns a new task facilitates replay and enables continual learning without catastrophic forgetting. This is formalized as the Sleep Replay Consolidation (SRC) algorithm, interleaving standard backpropagation training ('awake' training) with a simulated sleep phase. The sleep phase involves a one-time step of forward propagation of spontaneous activity, followed by a backward pass applying local Hebbian plasticity rules. This approach can be used independently or combined with existing rehearsal methods.

Existing research extensively documents the phenomenon of catastrophic forgetting in artificial neural networks. The stability-plasticity dilemma, a central challenge in continual learning, has been explored through various approaches including rehearsal methods that reintroduce previous data during training and regularization methods that constrain weight updates to preserve previously learned knowledge. The work of McClelland et al. (1995) on complementary learning systems in the hippocampus and neocortex highlights the biological mechanisms of memory consolidation, particularly the roles of replay and synaptic plasticity. Studies by Ji and Wilson (2007) and Walker and Stickgold (2004) further support the importance of sleep in memory consolidation and generalization in biological systems. These biological observations motivate the development of artificial sleep-like mechanisms to improve the continual learning capabilities of ANNs.

The Sleep Replay Consolidation (SRC) algorithm was implemented and tested using several datasets. The algorithm interleaves standard backpropagation training with a simulated sleep phase. In the sleep phase, the network's activation function is replaced with a Heaviside function to mimic neuronal spiking, and weights are scaled to enhance activity. An unsupervised Hebbian-type learning rule modifies synaptic weights based on pre- and post-synaptic neuron activations. Noisy binary inputs, drawn from a Poisson distribution reflecting the mean input intensity across past training sessions, drive network activity during sleep. Crucially, no specific memory traces are presented during sleep; the network's state implicitly determines the replay patterns. The SRC algorithm was evaluated in class-incremental learning and cross-modal learning scenarios using datasets like MNIST, Fashion MNIST, CUB-200, and CIFAR-10. In class-incremental learning, the network learns classes sequentially without access to previously learned classes. Cross-modal tasks involved learning two distinct datasets sequentially (e.g., MNIST and Fashion MNIST). A toy example using binary patterns was also employed to analyze synaptic weight changes during sleep and to visualize the impact of SRC on alleviating catastrophic interference. Comparisons were made with baseline ANNs trained sequentially without sleep, and with state-of-the-art continual learning methods including Elastic Weight Consolidation (EWC), Synaptic Intelligence (SI), and Orthogonal Weight Modification (OWM). In addition, the complementary effects of SRC with rehearsal methods (incorporating a small percentage of old data during new task training) were investigated using a modified loss function promoting balanced learning of old and new tasks. Specifically, the study also analyzed the combination of SRC with iCaRL, a state-of-the-art rehearsal method, examining its impact on performance with different memory capacities (K, number of stored examples from previous classes) and evaluating training time savings.

The SRC algorithm demonstrated a significant improvement in overall and task-specific performance compared to baseline sequential training without sleep across various datasets. On the CUB-200 dataset, where the baseline ANN showed catastrophic forgetting (5% accuracy on the first task, 95% on the second), SRC yielded significantly higher and more balanced accuracies (63.2% and 45.4% respectively). Similar improvements were observed for CIFAR-10, where SRC boosted accuracy from 19% to 44%. In the cross-modal task (MNIST and Fashion MNIST), SRC improved overall accuracy from 47% to 61%. Analysis of synaptic weight changes revealed that SRC downscales task-irrelevant synaptic weights, reducing cross-talk between tasks. In the toy model, SRC prevented catastrophic forgetting by inhibiting connections from task-irrelevant neurons to output neurons of the old task, resulting in correct classification. Further analysis demonstrated that while catastrophic forgetting was observed from a classification perspective, network weights retained information about previous tasks. SRC was found to be complementary to state-of-the-art rehearsal methods. Incorporating a small percentage of old data during training along with SRC significantly improved performance compared to using only old data or SRC alone. Combining SRC with iCaRL resulted in higher accuracy than iCaRL alone for most tested memory capacities, sometimes even with lower memory capacity, demonstrating a reduction in data storage requirements and training time.

The SRC algorithm effectively alleviates catastrophic forgetting by simulating key aspects of biological sleep replay. The spontaneous replay of neuronal activity during the sleep phase, combined with unsupervised Hebbian plasticity, leads to the redistribution of network resources and the formation of unique representations for each task, thereby reducing representational overlap and interference. This contrasts with rehearsal methods, which require storing or generating old data, and regularization methods, which constrain weight updates. While rehearsal methods show high performance, they lack the biological plausibility of SRC's spontaneous replay mechanism. SRC's ability to improve upon regularization methods like EWC and SI in an incremental learning setting, and its compatibility with rehearsal methods (reducing data storage and training time), highlights its potential as a valuable tool for continual learning in ANNs.

This study introduces the Sleep Replay Consolidation (SRC) algorithm, a biologically inspired approach to mitigate catastrophic forgetting in ANNs. SRC leverages unsupervised replay and Hebbian plasticity to recover old task performance without storing task-specific data. Results demonstrate its effectiveness across different datasets and its complementarity with state-of-the-art rehearsal methods. Future research could explore more complex sleep dynamics and plasticity rules to further enhance SRC's capabilities. The findings offer valuable insights into the mechanisms underlying continual learning and provide a promising avenue for developing more robust and adaptable ANNs.

The current implementation of SRC is a simplified model of biological sleep, employing a basic Hebbian-type learning rule and noisy input. More complex sleep patterns and learning rules might lead to further performance improvements. Furthermore, convolutional layers were kept frozen during the study; future research should investigate the impact of SRC on these layers. The study primarily focused on fully connected networks; the generalizability of SRC to different network architectures requires further exploration. Finally, while SRC showed promising results, its performance didn't always surpass all state-of-the-art rehearsal methods in every scenario.