Sleep-like Replay Consolidation

• Sleep-like Replay Consolidation (SRC) is a biologically inspired process that reactivates neural patterns during sleep-like phases to stabilize memory and reorganize synaptic connections.
• SRC employs synthetic replay, local Hebbian/STDP-like updates, and phase-specific dynamics (e.g., NREM/REM) to refine representations under limited-data and sequential learning conditions.
• SRC methods have demonstrated significant accuracy gains in class-imbalanced and continual learning settings by mitigating catastrophic forgetting and enhancing generalization.

Searching arXiv for recent and core papers on Sleep-like Replay Consolidation.

Sleep-like Replay Consolidation (SRC) denotes a family of biologically inspired offline learning procedures in which previously learned activity patterns are reactivated during a sleep-like phase and synapses are reorganized without ordinary supervised updates, typically to stabilize memory traces, reduce interference, and improve learning when data are limited, imbalanced, or presented sequentially (Bazhenov et al., 2024, Hayes et al., 2021). In the ANN literature, SRC most commonly refers to a wake–sleep pipeline that trains a network conventionally, converts it to a spiking or replay-capable form, drives it with spontaneous or stochastic activity, applies local Hebbian or STDP-like plasticity, and then resumes ordinary inference or fine-tuning; in the broader replay literature, the same idea is connected to hippocampal replay, complementary learning systems, offline reinforcement learning, and stage-specific NREM/REM memory reorganization (Aubin et al., 2018, Yoshida et al., 2023).

1. Biological and theoretical foundations

Replay is defined as the reactivation of one or more neural patterns similar to activation patterns experienced during past waking experiences, and in mammals it is described as critical for memory formation, retrieval, and consolidation (Hayes et al., 2021). Within the complementary learning systems view, hippocampus learns quickly and stores specific episodes, cortex learns slowly and integrates across experiences, and sleep replay serves as the bridge between rapid episodic encoding and gradual cortical consolidation (Hayes et al., 2021). The same review emphasizes that biological replay is selective rather than exhaustive, can involve only fragments of memories, can occur in reverse order, and differs across sleep states; NREM is associated more strongly with transfer and storage of recent memories, whereas REM is linked to organizing internal representations, abstraction, schema formation, and interference reduction (Hayes et al., 2021).

Hippocampal replay has also been framed as an offline computational process rather than only a passive consolidation byproduct. During sleep and awake rest, place-cell sequences can replay in forward order, backward order, or as novel shortcuts that concatenate experienced fragments, and these phenomena have been connected to Dyna-style reinforcement learning and prioritized sweeping under limited replay budgets (Aubin et al., 2018). Model-based analyses of rat hippocampal spike trains further support the consolidation interpretation: replay events can be detected by likelihood ratios over template trajectories, their compression rate can be estimated by re-binning REST spike trains, and their timing shows a significant peak near zero lag with sharp-wave ripples (Box et al., 2014).

Recent theoretical work broadens this picture by assigning complementary roles to sleep stages. NREM slow waves, up/down states, and sequential firing patterns are described as mechanisms that can select and stabilize memories, whereas REM dreaming is treated as a mechanism for creative recombination, latent-structure discovery, and representation shaping (Yoshida et al., 2023). This broader framing implies that SRC is not merely a rehearsal heuristic; it is a proposal about how offline activity can both preserve and transform learned structure.

2. Canonical computational pipeline

A canonical SRC implementation trains an ANN in the ordinary wake phase, converts it to a spiking neural network for sleep, drives the SNN with stochastic input reflecting training statistics, applies local synaptic plasticity, remaps the network back to an ANN, and optionally fine-tunes on the original data (Bazhenov et al., 2024). In the specific MNIST and Fashion MNIST implementation, the network is a fully connected ANN with two hidden layers trained with backpropagation on a limited subset; the trained ANN is then transformed into an SNN with the same architecture, and the SNN’s activity is driven by randomly distributed Poisson spiking input that reflects average inputs observed in the training dataset (Bazhenov et al., 2024). During this sleep phase there are no labels and no additional external training data, and synaptic strength is increased if presynaptic activation is followed by postsynaptic activation and reduced if postsynaptic activation occurs without presynaptic activation (Bazhenov et al., 2024).

An explicit pseudocode-style formulation uses membrane-voltage dynamics and a local Hebbian/STDP-like rule during sleep (Bazhenov et al., 2024). In that formulation, for layer $l$,

$$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$

and a neuron fires when

$S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$

The associated weight update is

$$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$

This form makes clear that SRC is neither ordinary gradient descent nor simple noise injection. The replay input is synthetic, the update is local, and the objective is implicit in the induced co-activity statistics rather than in a supervised loss.

Across implementations, the dominant computational motifs are consistent: spontaneous or replay-like stochastic activity, local synaptic updates, and selective strengthening and weakening of synapses (Bazhenov et al., 2024). A plausible implication is that SRC acts as a consolidation operator over an already trained representation, rather than as a substitute for the initial learning phase.

3. Limited-data and class-imbalanced regimes

The strongest direct evidence for SRC comes from limited-data and imbalanced-data settings. When MNIST or Fashion MNIST training used only roughly $0.5\%$ to $10\%$ of the total data, sleep-like unsupervised replay improved accuracy by about $20$–$30$ percentage points on both datasets (Bazhenov et al., 2024). When more than about $10\%$ of the dataset was used, sleep alone had a slight negative effect, with a drop around $10$–$$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$0 in performance, but post-sleep fine-tuning on the original data largely recovered that loss (Bazhenov et al., 2024). The same study reports that even when the number of supervised epochs was increased and the pre-sleep model improved, the sleep phase still produced a substantial gain (Bazhenov et al., 2024).

A later sequential low-data study gives concrete single-task examples of the same pattern. For a $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$1 MNIST subset, overall accuracy rose from about $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$2 before sleep to about $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$3 after SRC, and for a $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$4 Fashion-MNIST subset it rose from about $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$5 to about $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$6 (Bazhenov et al., 2024). In imbalanced settings, the underrepresented class usually improved after sleep, and confusion matrices showed that sleep replay redistributed performance more evenly across classes rather than merely increasing a global score (Bazhenov et al., 2024). The earlier MNIST analysis provides an especially clear example of class collapse and recovery: with $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$7 of MNIST data, classes $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$8, $$v(l) \gets \lambda v(l) + \alpha \, W(l,l-1)\, S(l-1),$$9, $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$0, and $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$1 were all classified as $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$2 before sleep, whereas after sleep classes $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$3, $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$4, and $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$5 were classified correctly (Bazhenov et al., 2024).

The synaptic interpretation offered in these studies is selective rather than uniform. SRC increased the strength of a small fraction of critical synapses while weakening many others (Bazhenov et al., 2024). This suggests a sparsifying consolidation process in which brittle, overfit representations are reorganized into more selective and less interfering structures.

4. Continual learning and catastrophic forgetting

SRC has also been developed as a continual-learning mechanism for catastrophic forgetting. In one line of work, a biologically inspired sleep algorithm converts a feedforward ANN with ReLU units into a leaky integrate-and-fire SNN, drives the input layer with Poisson-distributed spike trains derived from average activation values of all tasks seen so far, applies a modified sigmoidal, weight-dependent STDP rule, and then converts the SNN back into an ANN (Krishnan et al., 2019). That study reports recovery of older tasks that had been forgotten in incremental learning, forward transfer learning of unseen tasks, and improved generalization to noisy or blurred inputs on MNIST, CUB200, and a toy patches dataset (Krishnan et al., 2019). It also explicitly distinguishes its method from strict replay-based SRC: it uses average inputs, noisy inputs, or randomized portions of average images rather than replaying stored episodic samples (Krishnan et al., 2019).

Explicit SRC in sequential low-data training shows stronger replay-style behavior. In a two-task setup with Task 1 consisting of classes $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$6–$S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$7 and Task 2 of classes $S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$8–$S(l)_i = 1 \quad \text{if } v(l)_i > \beta.$9, sleep inserted after Task 2 rescued Task 1 in many conditions and in some cases improved Task 2 as well; the same paper argues that this indicates representation refinement rather than memory protection alone (Bazhenov et al., 2024). It also notes that all layers except the output layer are active during sleep, which supports the interpretation that hidden-layer feature refinement is central to the observed gains (Bazhenov et al., 2024).

The most systematic continual-learning extension in the provided corpus is the adaptation of SRC to Equilibrium Propagation-trained recurrent networks. There, SRC is inserted after each task in sequences such as $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$0, with Poisson-distributed spiking input and an STDP-like local rule during sleep (Kubo et al., 12 Aug 2025). On class-incremental learning benchmarks, the reported improvements are large: on MNIST, MRNN-EP rises from $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$1 sequential accuracy to $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$2 with SRC; on Fashion-MNIST, from $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$3 to $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$4; on Kuzushiji-MNIST, from $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$5 to $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$6; on CIFAR-10, from $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$7 to $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$8; and on ImageNet, from $$W_{ij} \gets \begin{cases} W_{ij} + inc, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 1 \ W_{ij} - dec, & \text{if } S(l)_j = 1 \text{ and } S(l-1)_i = 0 \ W_{ij}, & \text{otherwise.} \end{cases}$$9 to $0.5\%$0 (Kubo et al., 12 Aug 2025). SRC plus rehearsal further improves performance, reaching $0.5\%$1 on MNIST and $0.5\%$2 on CIFAR-10, while rehearsal alone using only $0.5\%$3 of past data gives only modest gains; the paper reports that at least about $0.5\%$4 old data is needed to match the gains SRC achieves without any old data (Kubo et al., 12 Aug 2025).

5. Reinforcement-learning and sequence-based interpretations

A distinct interpretation treats SRC as offline reinforcement learning built on replay. In the Dyna framework, an agent learns online from actual transitions and offline from model-generated replays, with value updates of the form

$0.5\%$5

where

$0.5\%$6

Under limited replay budget, prioritized sweeping replays the most informative transitions first and requires a predecessor model so that value changes can be propagated backward to states that lead to the updated state (Aubin et al., 2018).

The neural version of prioritized sweeping described in this literature includes a Q-network $0.5\%$7 for each action, a reward model $0.5\%$8, and a predecessor model $0.5\%$9, with replay priorities proportional to absolute TD error (Aubin et al., 2018). Because a state can have multiple predecessors for the same action, the paper introduces GALMO, a Growing Algorithm to Learn Multiple Outputs, using multiple output networks and associated gating networks to cope with one-to-many predecessor mappings (Aubin et al., 2018). The empirical claim is that Dyna-Q with prioritized sweeping learns faster than plain Q-learning, while the generated replays are mostly non-sequential but include meaningful backward and forward sequences (Aubin et al., 2018).

This RL interpretation does not replace the consolidation view; it reframes part of it. Replay may contribute to memory consolidation in a broader biological sense while also serving as offline value learning, especially around reward sites and decision points (Aubin et al., 2018). A plausible implication is that SRC can be understood as a more general family of offline reactivation algorithms, spanning supervised consolidation, continual learning, and model-based credit assignment.

6. Multi-stage sleep, boundaries of the term, and open questions

Recent work increasingly treats sleep-like consolidation as a staged process rather than a single replay operation. A review of the computational role of sleep argues that NREM slow waves constrain replay into a low-dimensional dynamical regime that can select and stabilize memories, while REM dreaming promotes creative combinations of fragments of memory and may support representation shaping through GAN-like or contrastive mechanisms (Yoshida et al., 2023). The same review states that “The infomax synaptic plasticity is formulated by increasing mutual information $10\%$0 under the constraint of synaptic weight cost +,” and reports predictions that high baseline firing biases plasticity toward weakening and that global up/down states induce stronger synaptic strengthening than local up/down states (Yoshida et al., 2023).

Wake-Sleep Consolidated Learning is an explicit stage-structured continual-learning framework derived from complementary learning systems theory. During wake, it stores current-task samples in a temporary memory $10\%$1 and selects a dynamic freezing mask; during NREM, it replays samples from both $10\%$2 and the long-term memory $10\%$3 to consolidate synapses; during REM, it trains on previously unseen realistic visual sensory experience to explore the feature space and prepare synapses for future knowledge (Sorrenti et al., 2023). Reported results show substantial gains in final average accuracy and unusually positive forward transfer: for example, on CIFAR-10 with ER-ACE and a 200-sample buffer, FAA rises from $10\%$4 to $10\%$5, and several baselines switch from negative to positive FWT under WSCL (Sorrenti et al., 2023). This supports the view that SRC-like systems can include both replay-driven stabilization and dream-like prospective organization.

At the same time, the literature places clear limits on the term. Deep-learning replay methods often capture only the “mix old with new” aspect of biological replay and usually ignore sleep-state distinctions, temporally structured replay, spontaneity, coordinated multi-region activity, and REM-like reorganization (Hayes et al., 2021). Some biologically inspired sleep algorithms are therefore SRC-adjacent rather than strict SRC implementations, because they rely on average or noisy inputs instead of replaying stored exemplars or internally reconstructed episodes (Krishnan et al., 2019). There is also no claim that SRC is universally beneficial in isolation: in the limited-data MNIST/Fashion-MNIST setting, sleep alone slightly harms performance once more than about $10\%$6 of the dataset is available, and fine-tuning is needed to avoid degradation in better-trained regimes (Bazhenov et al., 2024).

The central open question concerns what part of biological replay is computationally essential. Existing artificial implementations already show that offline stochastic reactivation plus local plasticity can improve low-data learning, reduce class imbalance effects, rescue forgotten tasks, and complement rehearsal (Bazhenov et al., 2024, Bazhenov et al., 2024, Kubo et al., 12 Aug 2025). The broader review literature suggests, however, that full SRC may require explicit awake and sleep phases, selective and temporally structured replay, REM-like representational restructuring, and tighter coupling between replay, normalization, and synaptic rebalancing than most current systems implement (Hayes et al., 2021, Yoshida et al., 2023).