Unsupervised sleep-like intra- and inter-layer plasticity categorizes and improves energy efficiency in a multilayer spiking network

TL;DR

This study investigates how sleep-like, multi-layer plasticity in a biologically grounded thalamo-cortical spiking network supports memory abstraction and energy efficiency. By enabling full inter- and intra-layer plasticity during sleep and introducing an ATP-based metabolic estimator, the authors show that sleep-driven reorganization yields higher post-sleep accuracy in few-shot learning and substantial reductions in metabolic power, quantified in ATP units. The model reproduces canonical sleep signatures, fosters hierarchical category representations, and reveals that inter-layer plasticity, especially cx→th, plays a key role in consolidation and energy downscaling. Together, these findings offer design principles for energy-aware neuromorphic AI that leverage brain-state dynamics for efficient learning and abstraction, while highlighting avenues for future enhancements such as multi-area extensions and REM-like phases.

Abstract

Sleep is thought to support memory consolidation and the recovery of optimal energetic regime by reorganizing synaptic connectivity, yet how plasticity across hierarchical brain circuits contributes to abstraction and energy efficiency remains unclear. Here we study a spiking multi-layer network alternating wake-like and deep-sleep-like states, with state-dependent dendritic integration and synaptic plasticity in a biologically inspired thalamo-cortical framework. During wakefulness, the model learns from few perceived examples, while during deep sleep it undergoes spontaneous replay driven by slow oscillations. Plasticity enabled not only within intra-layer connections, but also in inter-layer pathways, is critical for memory consolidation and energetic downshift. Compared to restricted plasticity, full inter-layer plasticity yields higher post-sleep visual classification accuracy and promotes the emergence of sharper class-specific associations. Furthermore, we introduce a biophysically grounded estimator of metabolic power expressing network energy consumption in ATP units, partitioned into baseline, synaptic maintenance, action potential, and transmission costs. We find that inter-layer plasticity in sleep leads to a larger reduction in firing rates, synaptic strength and synaptic activity, corresponding to a substantially larger decrease in power consumption. This work suggests promising elements to be integrated in neuromorphic/energy-efficient AI learning systems, supported by brain state-specific apical mechanisms.

Figures (9)

• Figure 1: A biologically grounded multi-layer model with inter- and intra-layer recurrence for sleep and learning. A) Layer-5 cortical pyramidal cell integrating contextual signals (red) with perceptual (blue) and/or local (green) signals. B) Apical amplification: during wakefulness, this mechanism enables the amplification of the spiking activity if apical signal is received in temporal coincidence with (peri-)somatic inputs. C) Apical isolation: during deep sleep, this mechanism prevents from apical signals to reach the soma. Panels A-C adapted from pastorelli2025simplified. D) Two-layer excitatory-inhibitory spiking neural network sustaining learning and sleep through apical mechanisms. Circles: neuronal populations, excitatory and inhibitory; hexagons: inputs to the network, encoded by the rates of Poisson processes. Either HOG-filtered MNIST digits, or CNN-processed CIFAR-10 images, are injected into the thalamus as perceptual (P) signals. Inter-layer (th$\to$cx and cx$\to$th) and recurrent (cx$\to$cx) excitatory connections are plastic (red), while others remain static (grey). Contextual (C) stimuli, administered during training, are necessary for engrams creation, while aspecific (A) random signals are necessary for inducing NREM-like slow oscillations during sleep.
• Figure 2: Network Activity. Evolution of firing rates and spectral content of the thalamo-cortical activity during an awake-sleep learning cycle (training, blue; awake classification, red; NREM sleep, green). The endogenous homeostatic depression happening during sleep affects post-sleep firing rates and spectral content. For legibility, a restricted subset of 180.0 neurons is shown in panels A, B, and C, corresponding to the cortical assemblies encoding the first 9.0 learned training examples (representing only the first three classes). A) Rastergram of cortical activity. B) Average instantaneous firing rate $\overline{\nu}$, obtained from the cortical rastergram in panel A) via a $\sigma=\qty{1.2}{\ms}$ Gaussian convolution. C) Corresponding spectrogram, obtained via a narrower ($\sigma=\qty{0.3}{\ms}$) Gaussian convolution to highlight the high-frequency spectral content. Notably, during NREM sleep, delta-band activity is prominent with high-frequency gamma in coincidence with up states. Early sleep manifests a larger frequency of up states than late sleep. D) Firing rate change from early to late NREM for cortex (left) and thalamus (right). Main panels are scatter plots, stacking together all the neurons from the 100.0 independent trials (each dot representing a neuron). x-axis is the early spiking rate, while y-axis is its relative change. Medians along both axes are represented with a white cross. Smaller panels at the top and on the right represent trial-averaged marginal distributions for early firing rates and relative changes (shaded area: one standard deviation). Medians from scatter plots are also reported here as black tips, for comparison. E) Same as in D), here comparing changes in the firing rate activity during awake post-sleep classification, with respect to pre-sleep classification.
• Figure 3: Sleep effects on cognitive and energetic performance on MNIST classification: comparison with prior ThaCo baseline. In all panels, blue denotes the previously published model capone2019sleeplike with sleep plasticity restricted to cx$\to$cx synapses only, whereas red denotes the present full-plasticity model (also th$\to$cx and cx$\to$th plasticity are enabled during sleep). Observables are measured post-sleep for incremental sleep duration (20.0 rounds of 100-long NREM-sleep). Network is trained on three MNIST digit examples per class (10.0 classes in total). Classification is performed over a balanced [$\to$ continues on next page]
• Figure 4: Network synaptic changes during sleep. During sleep, apical isolation mechanism drives an endogenous dynamics inducing associative and homeostatic effects in synaptic weights. A) Sleep effects on the cx$\to$cx synaptic matrix for a representative trial, zoomed on the first three digit classes, three examples per class. Synaptic weight values are color-coded in log scale (reference weight is 1). Left: before sleep, stronger synapses (red) emerge within example-specific assemblies, sculpted by apical amplification during the initial training, while cross-example synapses (blue) show no sign of association, staying around the randomly initialized values some orders of magnitude below. Right: after 2000 of sleep, the strength of example-specific synapses reduces (homeostatic depression; from darker to lighter red), while intermediate-strength synapses spontaneously emerge among cell assemblies encoding for the same digit class (memory association; light blue). B) Trial-average (100.0 independent trials) of the entire post-sleep/pre-sleep cx$\to$cx synaptic matrix (ten digit classes), coherently showing the homeostatic and associative effects previously reported for a single trial. C) Homeostatic effects for example-specific synapses and associative effects for class-specific synapses shown as a function of sleep time for for the connectivity matrices: cx$\to$cx, cx$\to$th and th$\to$cx; data are computed as the ratio between post-sleep/pre-sleep averages over the three functional classes of connectivity for every given trial; medians and inter-quartile ranges over the means from the 100.0 independent trials are shown. See the analogous for CIFAR-10 in Suppl. Fig. \ref{['supp_fig:net_synapses_CIFAR']}.
• Figure 5: MNIST and CIFAR-10 input preprocessings and resulting structures. Raw images drawn from MNIST and CIFAR-10 datasets are processed into 81.0 float values using HOG filters and CNN algorithms, respectively. A) Each input feature (81.0 floating-point values in range [range-phrase=--]01) is one-hot encoded using a four-level discretization scheme ([range-phrase=--]00.25, [range-phrase=--]0.250.50, [range-phrase=--]0.500.75, [range-phrase=--]0.751) (adapted from golosio2021thalamocortical). B) Histogram of oriented gradients computed from 28.0$\times$28.0 MNIST images using 9.0 overlapping 7.0$\times$7.0 kernels. 81.0-floats resulting features are expanded by the one-hot encoding digitization to produce the 324.0-bit input to ThaCo. C) Residual convolutional neural network (ResNet) [$\to$ continues on next page]