Hierarchical predictive coding proposes a compelling hypothesis of brain computation, suggesting that the cortex builds layered predictions to minimize surprise. Yet most models rely on error-coding neurons or generative modeling of unclear biological plausibility. Here, we examine a biologically plausible framework in which the functional goals of predictive coding emerge from local contrastive learning and simple activity cancellation. Building on recent machine learning advances, we present a recurrent variant of the Forward-Forward (FF) algorithm with an inverted objective that increases activity for negative data. This setup yields predictive representations across layers, capturing hallmark features of cortical computation such as top-down modulation and surprise signaling. Our results suggest that key principles of predictive coding can emerge from simple, local learning rules, offering a new bridge between neuroscience and machine learning.

Abstract Imaging flow cytometry (IFC) provides spatial resolution and multi-parametric measurements to characterise cells of interest. It enables profound insights into cell signalling, co-localisation, cell-to-cell interaction, and DNA studies. However, the use of traditional frame-based sensors (FBS) introduces the classic trade-off between speed, resolution, and sensitivity, where improving one often degrades the others. This limitation has hindered both technological progress and widespread adoption. The large data volumes generated by IFC further complicate real-time applications. The spatial richness of IFC data, however, makes it highly compatible with machine learning (ML), enabling automated analysis, gating, and the discovery of rare events. To overcome FBS constraints, neuromorphic vision sensors (NVS) have recently been introduced. Their event-driven design, characterised by data sparsity, provides greater data-processing efficiency, temporal resolution, and fluorescence sensitivity compared to conventional imaging methods. This review examines advances in flow cytometry (FC) and IFC, evaluating their respective strengths and drawbacks. It highlights innovations in NVS that address long-standing limitations, explores recent advances in neuromorphic imaging cytometry (NIC) across different modalities and ML approaches, and discusses current challenges and future directions shaping this emerging field. NIC represents a pivotal step toward real-time, data-efficient cytometry platforms that integrate event-driven sensing with spiking neural computation.

Continual learning (CL) enables artificial neural networks to adapt to new information while ideally retaining prior knowledge. Despite progress, most CL methods are still evaluated under idealized, unrealistic conditions that fail to capture the complexity of real-world learning. While previous reviews have surveyed continual learning algorithms and categorized strategies, they have not systematically addressed the disconnect between experimental protocols and real-world deployment requirements. In this work, we critically assess the limitations of current CL benchmarks and propose a structured framework consisting of three interconnected components: (1) benchmark datasets that reflect real-world complexity, (2) training protocols that accommodate non-stationary, task-agnostic data streams, and (3) comprehensive evaluation metrics beyond accuracy. We demonstrate critical flaws in widely-used benchmarks, which we reframe as instances of structural information leakage. Drawing on neuroscience principles-including synaptic plasticity mechanisms, complementary learning systems, and memory consolidation processes-we discuss how biological insights can inform more robust algorithm design. Through this review, we aim to inspire more robust, flexible CL systems and evaluation platforms capable of tackling the challenges of practical learning environments.

Event cameras provide superior temporal resolution, dynamic range, power efficiency, and pixel bandwidth. Spiking Neural Networks (SNNs) naturally complement event data through discrete spike signals, making them ideal for event-based tracking. However, current approaches that combine Artificial Neural Networks (ANNs) and SNNs, along with suboptimal architectures, compromise energy efficiency and limit tracking performance. To address these limitations, we propose the first Transformer-based spike-driven tracking pipeline. Our Global Trajectory Prompt (GTP) method effectively captures global trajectory information and aggregates it with event streams into event images to enhance spatiotemporal representation. We then introduce SDTrack, a Transformer-based spike-driven tracker comprising a Spiking MetaFormer backbone and a tracking head that directly predicts normalized coordinates using spike signals. The framework is end-to-end, does not require data augmentation or post-processing. Extensive experiments demonstrate that SDTrack achieves state-of-the-art performance while maintaining the lowest parameter count and energy consumption across multiple event-based tracking benchmarks, establishing a solid baseline for future research in the field of neuromorphic vision.

Looped language models improve reasoning and knowledge manipulation by applying shared computation repeatedly. Existing systems usually repeat an entire layer stack, although a mixer and a dense feed-forward network (FFN) perform different operations and have different costs. We ask a narrower question: what should loop? We view recurrence as repeated composition of a state update and argue that an application is valuable when it exposes a new cross-position influence direction that remains observable at the task readout. Iterative Transport Rank (ITR) describes the cumulative influence trajectory; marginal ITR describes the nonredundant influence contributed by successive applications. This view motivates MixerLoop, which repeats each Gated DeltaNet mixer while applying its dense FFN once. We compare MixerLoop with no recurrence and full-block recurrence at 15M and 110M parameters under the same data, initialization, and architecture. A finite context-off intervention tests whether later mixer applications produce distinct, non-negligible, and beneficial changes at the final language-model readout. MixerLoop surpasses FullLoop on aggregate CORE at 15M and retains 41.5

Traditional Knowledge Distillation (KD) transfers all outputs from a teacher model to a student model, often introducing knowledge redundancy. This redundancy dilutes critical information, leading to degraded student model performance. To address this, we propose Salient Feature Masking for Knowledge Distillation (SFKD), a lightweight enhancement that masks out less informative components and selectively distills only the top-K activations. SFKD is a drop-in modification applicable to both logit-based and feature-based KD, incurs negligible overhead, and sharpens the student’s learning signal. Empirically, SFKD yields consistent gains over strong KD baselines across architectures (ConvNeXt, ViT) and datasets (CIFAR-100: up to +2.43 pp; CUB-200: up to +6.39 pp; ImageNet-1K: up to +3.57 pp). We also provide intuition from the information bottleneck perspective to motivate why filtering out less salient teacher signals benefits the student. Overall, SFKD is a simple, empirically validated method for training student models that are both leaner and more accurate.

Neural population models, which predict the joint firing of many simultaneously recorded neurons forward in time, are typically evaluated by a single aggregate Pearson correlation r between predicted and actual spike counts, a number that masks critical structure. We argue that how we evaluate spike forecasting matters as much as what we build, and introduce SpikeProphecy, the first large-scale benchmark for causal, autoregressive spike-count forecasting on real electrophysiology recordings. Our core contribution is a population metric decomposition that separates aggregate performance into temporal fidelity, spatial pattern accuracy, and magnitude-invariant alignment. The decomposition surfaces aspects of the underlying data that an aggregate scalar collapses together. We apply the protocol to 105 Neuropixels sessions (Steinmetz 2019 + IBL Repeated Site; 89,800 neurons) with seven architecture baselines spanning four structural families: four SSMs (three diagonal and one non-diagonal), a Transformer, an LSTM, and a spiking network. The decomposition surfaces a brain-region predictability ranking that reproduces across all seven baselines and survives ANCOVA correction for firing-statistics constraints (region ΔR^2 = 0.018 above the firing-statistics covariates). It also exposes a sub-Poisson evaluation floor where rigorous metrics combine with genuine biophysical constraints on regular spike trains, and yields a negative result on KL-on-output-rates distillation for ANN-to-SNN transfer in this Poisson count domain.

The mimicry of the biological brain’s structure in information processing enables spiking neural networks (SNNs) to exhibit significantly reduced power consumption compared to conventional systems. Consequently, these networks have garnered heightened attention and spurred extensive research endeavors in recent years, proposing various structures to achieve low power consumption, high speed, and improved recognition ability. However, researchers are still in the early stages of developing more efficient neural networks that more closely resemble the biological brain. This development and research require suitable hardware for execution with appropriate capabilities, and field-programmable gate array (FPGA) serves as a highly qualified candidate compared to existing hardware such as central processing unit (CPU) and graphics processing unit (GPU). FPGA, with parallel processing capabilities similar to the brain, lower latency and power consumption, and higher throughput, is highly eligible hardware for assisting in the development of spiking neural networks. In this review, an attempt has been made to facilitate researchers’ path to further develop this field by collecting and examining recent works and the challenges that hinder the implementation of these networks on FPGA.

Spiking neural networks (SNNs) process information via sparse, event-driven spikes. In hardware, this sparsity can dramatically reduce the volume of data moved between memory hierarchies, which often dominates both energy and latency. This paper presents a profiler that estimates the three principal stages of data movement-between register and cache, between cache and memory, and between memory and storage-and quantifies how increasing sparsity linearly lowers dynamic traffic while leaving static weight movement unchanged. This can act as a robust surrogate for energy consumption, helping deep learning engineers evaluate the viability of SNNs in resource-constrained deployment scenarios. Furthermore, it can potentially inform of optimal hardware architectures, for example, by evaluating dataflow efficiency or the impact of GPU-style DRAM interactions.

Recent advancements in neuroscience research have propelled the development of Spiking Neural Networks (SNNs), which not only have the potential to further advance neuroscience research but also serve as an energy-efficient alternative to Artificial Neural Networks (ANNs) due to their spike-driven characteristics. However, previous studies often overlooked the multiscale information and its spatiotemporal correlation between event data, leading SNN models to approximate each frame of input events as static images. We hypothesize that this oversimplification significantly contributes to the performance gap between SNNs and traditional ANNs. To address this issue, we have designed a Spiking Multiscale Attention (SMA) module that captures multiscale spatiotemporal interaction information. Furthermore, we developed a regularization method named Attention ZoneOut (AZO), which utilizes spatiotemporal attention weights to reduce the model's generalization error through pseudo-ensemble training. Our approach has achieved state-of-the-art results on mainstream neuromorphic datasets. Additionally, we have reached a performance of 77.1\% on the Imagenet-1K dataset using a 104-layer ResNet architecture enhanced with SMA and AZO. This achievement confirms the state-of-the-art performance of SNNs with non-transformer architectures and underscores the effectiveness of our method in bridging the performance gap between SNN models and traditional ANN models.

Imaging flow cytometry (IFC) is a powerful cell analytic tool that exploits multi-parameters in single-cell images to characterise cell phenotypes and fluorescence information. It enables in-depth analysis of cell signalling, DNA repair and marker localisation. However, conventional frame-based acquisition is bound to the triangle of imaging constraints—speed, resolution and sensitivity, which has become an everlasting challenge to overcome during development. Neuromorphic photosensors detect contrast changes in a scene via individual-firing pixels, characterising superior data efficiency, temporal resolution and fluorescence sensitivity. In this work, we have developed a neuromorphic imaging cytometer (NIC) to capture fast-moving cell events, curating the first neuromorphic cell dataset with human blood cells, endothelial cells and artificial particles. Recently, this sensor has been adopted to address the limitations in IFC with prominent results in diverse modalities and machine learning approaches. Such a dataset serves as a baseline of healthy cell groups for both diagnostic and research purposes. In addition, the rich spatial information derived from cell images has exceptional uses with deep learning (DL) approaches to automate cell analysis, classification, sorting and gating strategy. We also trained a lightweight model combining the convolutional block attention module with a spiking neural network (CBAM-SNN) to automate cell analysis and classification. The proposed architecture has achieved a promising performance of 97% accuracy and F1 score with a significant reduction in computation requirements. Combining the data sparsity in neuromorphic imaging with a lightweight DL model and operation platform can enable next-generation, AI-driven cytometry to deliver point-of-care diagnostic and research solutions.

Accurate time-series forecasting is crucial in various scientific and industrial domains, yet deep learning models often struggle to capture long-term dependencies and adapt to data distribution shifts over time. We introduce Future-Guided Learning, an approach that enhances time-series event forecasting through a dynamic feedback mechanism inspired by predictive coding. Our method involves two models: a detection model that analyzes future data to identify critical events and a forecasting model that predicts these events based on current data. When discrepancies occur between the forecasting and detection models, a more significant update is applied to the forecasting model, effectively minimizing surprise, allowing the forecasting model to dynamically adjust its parameters. We validate our approach on a variety of tasks, demonstrating a 44.8% increase in AUC-ROC for seizure prediction using EEG data, and a 23.4% reduction in MSE for forecasting in nonlinear dynamical systems (outlier excluded). By incorporating a predictive feedback mechanism, Future-Guided Learning advances how deep learning is applied to time-series forecasting.

Spiking neural networks (SNNs), which draw from biological neuron models, have the potential to improve the computational efficiency of artificial neural networks (ANNs) due to their event-driven nature and sparse data flow. SNNs rely on dynamical sparsity, in that neurons are trained to activate sparsely to minimize data communication. This is critical when accounting for hardware given the bandwidth limitations between memory and processor. Given that neurons are sparsely activated, weights are less frequently accessed, and potentially can be pruned to less performance degradation in a SNN compared to an equivalent ANN counterpart. Reducing the number of synaptic connections between neurons also relaxes memory demands for neuromorphic processors. In this paper, we propose a spatio-temporal pruning algorithm that dynamically adapts to reduce the temporal redundancy that often exists in SNNs when processing Dynamic Vision Sensor (DVS) datasets. Spatial pruning is executed based on both global parameter statistics and inter-layer parameter count and is shown to reduce model degradation under extreme sparsity. We provide an ablation study that isolates the various components of spatio-temporal pruning, and find that our approach achieves excellent performance across all datasets, with especially high performance on datasets with time-varying features. We achieved a 0.69% improvement on the DVS128 Gesture dataset, despite the common expectation that pruning typically degrades performance. Notably, this enhancement comes with an impressive 98.18% reduction in parameter space and a 50% reduction in time redundancy.

While deep learning models have seen significant success across various domains, their black-box learning nature and lack of interpretability affect their reliability in safety-critical applications like medical diagnostics and autonomous vehicles. In an attempt to address these limitations, Bayesian neural networks (BNNs) offer a promising alternative by incorporating uncertainty estimation into model predictions, enhancing transparency and decision-making. However, BNN development has primarily focused on efficient, high-fidelity approximate inference and guaranteed convergence in asymptotic settings. These are unsuitable for modern high-dimensional, multi-modal, and non-asymptotic deep learning applications, undermining their theoretical advantages. To bridge this gap, this paper provides in-depth reviews on how approximate Bayesian inference leverages deep learning optimization to achieve high efficiency and fidelity in high-dimensional spaces and multi-modal loss landscapes. It also reconciles Bayesian consistency with generalization objectives in non-asymptotic settings and investigates the generalization capabilities of BNNs. Additionally, this survey examines the often-overlooked expressiveness of BNNs, emphasizing how weight uncertainty and the absence of in-between uncertainty affect their performance. This survey aims to inspire BNN practitioners to adopt a deep learning perspective and offer valuable insights to propel further advancements in the field.

This work proposes a unified neuromorphic spike-based large-language-model (NSLLM) framework to simultaneously address the challenges of high energy consumption and low interpretability in LLMs. Our framework transforms LLMs into efficient NSLLMs by converting their behaviors into neural dynamics-such as spike trains-through rigorous mathematical modeling and complemented by advanced techniques including quantization and sparsification. This transformation also enables the analysis of information encoding processes using computational neuroscience tools, thereby offering a novel neuroscientific perspective that conceptualizes LLMs as neural populations to enhance their interpretability. Leveraging a hardware-algorithm co-design paradigm, an NSLLM can completely eliminate matrix multiplication (MatMul) while maintaining high performance. We designed a custom MatMul-free hardware core on the VCK190 field-programmable gate array to validate the 1.5-billion-parameter NSLLM, achieving a dynamic power consumption of only 13.849 W and an inference throughput of 161.8 tokens per second. Compared with the A800 GPU, this implementation improves energy efficiency, memory usage and inference throughput by 19.8[Formula: see text], 21.3[Formula: see text] and 2.2[Formula: see text], respectively. This work provides a novel perspective within a unified framework to enhance both the energy efficiency and interpretability of LLMs, offering valuable insights for future neuromorphic chip designs tailored for large models.

Efficient communication is central to both biological and artificial intelligence (AI) systems. In biological brains, the challenge of long-range communication across regions is addressed through sparse, spike-based signaling, minimizing energy consumption and latency. In contrast, modern AI workloads, which keep scaling ever larger across distributed compute systems, are increasingly constrained by bandwidth limitations, creating bottlenecks that hinder scalability and energy efficiency. Inspired by the brain's efficient communication strategies, we propose SNAP, a hybrid neural network architecture combining spiking neural networks (SNNs) and artificial neural networks (ANNs) to address these challenges. SNAP integrates SNNs at bandwidth-constrained regions, such as chip boundaries, where spike-based encoding reduces data transfer overhead. Within each chip, dense ANN computations are maintained to preserve high throughput, accuracy, and robustness. Historically, SNNs have faced difficulties scaling up, with limitations in task-specific performance and reliance on specialized hardware to exploit sparsity. SNAP overcomes these barriers through an algorithm-architecture co-design leveraging learnable sparsity for die-to-die communication while limiting spiking layers to specific network partitions. This composable design integrates spike-based and non-spiking pathways, making it adaptable to diverse deep learning workloads. Our evaluations on language processing and computer vision tasks demonstrate up to 5.3x energy efficiency improvements and 15.2x reductions in inference latency, outperforming both traditional SNNs and non-spiking models. We find that as model resources scale, SNAP's improvement margins grow. By addressing the critical bottleneck of inter-chip communication, SNAP offers a scalable, biologically inspired pathway to more efficient AI systems.

Flow models are effective at progressively generating realistic images, but they generally struggle to capture long-range dependencies during the generation process as they compress all the information from previous time steps into a single corrupted image. To address this limitation, we propose integrating autoregressive modeling – known for its excellence in modeling complex, high-dimensional joint probability distributions – into flow models. During training, at each step, we construct causally-ordered sequences by sampling multiple images from the same semantic category and applying different levels of noise, where images with higher noise levels serve as causal predecessors to those with lower noise levels. This design enables the model to learn broader category-level variations while maintaining proper causal relationships in the flow process. During generation, the model autoregressively conditions the previously generated images from earlier denoising steps, forming a contextual and coherent generation trajectory. Additionally, we design a customized hybrid linear attention mechanism tailored to our modeling approach to enhance computational efficiency. Our approach, termed ARFlow, achieves 6.63 FID scores on ImageNet at 256 * 256 without classifier-free guidance, reaching 1.96 FID with classifier-free guidance 1.5, outperforming the previous flow-based model SiT's 2.06 FID. Extensive ablation studies demonstrate the effectiveness of our modeling strategy and chunk-wise attention design.

We introduce NeuroSA, a neuromorphic architecture specifically designed to ensure asymptotic convergence to the ground state of an Ising problem using a Fowler-Nordheim quantum mechanical tunneling based threshold-annealing process. The core component of NeuroSA consists of a pair of asynchronous ON-OFF neurons, which effectively map classical simulated annealing dynamics onto a network of integrate-and-fire neurons. The threshold of each ON-OFF neuron pair is adaptively adjusted by an FN annealer and the resulting spiking dynamics replicates the optimal escape mechanism and convergence of SA, particularly at low-temperatures. To validate the effectiveness of our neuromorphic Ising machine, we systematically solved benchmark combinatorial optimization problems such as MAX-CUT and Max Independent Set. Across multiple runs, NeuroSA consistently generates distribution of solutions that are concentrated around the state-of-the-art results (within 99%) or surpass the current state-of-the-art solutions for Max Independent Set benchmarks. Furthermore, NeuroSA is able to achieve these superior distributions without any graph-specific hyperparameter tuning. For practical illustration, we present results from an implementation of NeuroSA on the SpiNNaker2 platform, highlighting the feasibility of mapping our proposed architecture onto a standard neuromorphic accelerator platform.