Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 13 appraisals
Current biology : CB
Looking to the brain to improve energy efficiency of AI
Modern artificial intelligence (AI) systems have achieved remarkable capabilities, but at an extraordinary energy cost. Training and running large-scale models can consume vast resources, posing environmental, economic, and societal challenges. In contrast, biological brains perform lifelong learning, adaptive control, and flexible reasoning using orders of magnitude less energy for learning and adaptation over a lifetime. What accounts for this difference - and how can it guide future AI development? In this review, we identify key biological principles that support energy-efficient capacities in biological brains, and consider how they might inform the design of more sustainable artificial systems. We organize our analysis around three domains: architectural constraints, signaling strategies, and learning algorithms. In each domain, we discuss concrete observations from biology, from cell to circuit to cognitive level, and describe how current and emerging AI systems mirror or diverge from these motifs. One striking feature of biological energy optimization is often overlooked: that brains are remarkably stable in their energy usage across heterogeneous modes, suggesting they may minimize energy needs during active environmental processing through maximizing the utility of 'rest-like' background processes. Overall, rather than advocating for biomimicry for its own sake, we argue for biologically informed engineering. Understanding how natural systems minimize energetic cost while maximizing flexibility may help us build AI that is not only powerful, but also efficient, equitable, and environmentally responsible.
21 July 2026
Read appraisal →Philosophical transactions of the Royal Society of London. Series B, Biological sciences
The hippocampus as a generative model
A generative model can be defined as a model of the latent causes of sensory input that can be used to generate new data samples. By examining empirical evidence and computational theory, we propose that the hippocampus can be characterized as a generative model. The hippocampus is a brain region important for memory. Recordings of neural activity from the hippocampus have led to the view that the hippocampus represents a cognitive map by abstracting a low-dimensional representation of the external world. We extend this view to suggest the hippocampus represents the latent, unobserved causes of sensory data by virtue of the position of the hippocampus within the deep cortical hierarchy. These representations of unobserved latent causes endow the hippocampus with capacity to generate new data samples that allow exploration of future hypotheticals and provide an internally generated training signal back to the generative model. We explore how perturbations to the hippocampal generative model may explain core symptoms of neuropsychiatric disorders such as those observed in psychosis. Together, this perspective provides a unified account of hippocampal function that explains how computations performed by the hippocampus support higher-order cognition and adaptive behaviour. This article is part of the theme issue 'The role of hippocampal predictions in cognition: bridging perception and memory'.
11 July 2026
Read appraisal →PLoS computational biology
Predictive coding explains asymmetric connectivity in the brain: A neural network study
Seminal frameworks of predictive coding propose a hierarchy of generative modules, each attempting to infer the neural representation of the module one level below; the predictions are carried by top-down feedback projections, while the predictive error is propagated by reciprocal forward pathways. Such symmetric feedback connections support visual processing of noisy stimuli in computational models. However, neurophysiological studies have yielded evidence of asymmetric cortical feedback connections. We investigated the contribution of neural feedback in visual processing for computing grasp parameters, by utilizing convolutional neural network models that had been augmented with predictive feedback and were trained to compute grasp positions for real-world objects. After establishing an ameliorative effect of symmetric feedback on grasp detection performance when evaluated on noisy stimuli, we characterized the performance effects of asymmetric feedback, similar to that observed in the cortex. Specifically, we tested model variants extended with short-, medium-, long- and longer-range feedback connections (i) originating at the same source layer or (ii) terminating at the same target layer. We found that the performance-enhancing effect of predictive coding under adverse conditions was optimal for medium-range asymmetric feedback. Moreover, this effect was most prominent when medium-range feedback originated at a level of representational abstraction that was proximal to the input layer, in contrast to more distal layers. To conclude, our simulations show that introducing biologically realistic asymmetric predictive feedback improves model robustness to noisy visual stimuli in a neural network model optimized for grasp detection.
8 July 2026
Read appraisal →Journal of mathematical biology
Spiking neural models for decision-making tasks with learning.
In cognition, response times and choices in decision-making tasks are commonly modeled using Drift Diffusion Models (DDMs), which describe the accumulation of evidence for a decision as a stochastic process, specifically a Brownian motion, with the drift rate reflecting the strength of the evidence. In the same vein, the Poisson counter model describes the accumulation of evidence as discrete events whose counts over time are modeled as Poisson processes. This model has a spiking neurons interpretation as these processes are used to model neuronal activities. However, these models lack a learning mechanism and are limited to tasks where participants have prior knowledge of the categories. To bridge the gap between cognitive and biological models, we propose a biologically plausible Spiking Neural Network (SNN) model for decision-making that incorporates a learning mechanism and whose neurons activities are modeled by a multivariate Hawkes process. First, we show a coupling result between the DDM and the Poisson counter model, establishing that these two models provide similar categorizations and reaction times and that the DDM can be approximated by spiking Poisson neurons. To go further, we show that a particular DDM with correlated noise can be derived from a Hawkes network of spiking neurons governed by a local learning rule. In addition, we designed an online categorization task to evaluate the model predictions. This work provides a significant step toward integrating biologically relevant neural mechanisms into cognitive models, fostering a deeper understanding of the relationship between neural activity and behavior.
14 June 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
Unexpected benefits of self-modelling in neural systems
Self-models have been a topic of interest for decades in human cognition and more recently in machine learning. Yet what benefits do self-models confer? Here we show that when an artificial network learns to predict its internal states as an auxiliary task, the network becomes simpler, more regularized and more parameter efficient. To test the hypothesis of self-regularizing through self-modelling, we used a range of network architectures performing three classification tasks across two modalities. In all cases, adding self-modelling caused a significant reduction in network complexity. The reduction was observed in two ways. First, the distribution of weights was narrower when self-modelling was present. Second, a measure of network complexity, the real log canonical threshold (RLCT), was smaller when self-modelling was present. These results support the hypothesis that self-modelling has a restructuring effect, reducing complexity and increasing parameter efficiency. This self-regularization may help explain some of the benefits of self-models reported in recent machine learning literature, as well as the adaptive value of self-models to biological systems. In particular, these findings may shed light on the possible interaction between the ability to model oneself and the ability to be more easily modelled by others in a social or cooperative context. This article is part of the theme issue 'World models in natural and artificial intelligence'.
17 May 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
What physics offers for artificial intelligence? Lessons from the brain's inner time and its dynamics.
Computing can take advantage of physics-that is, 'whatever physics offers' (Jaeger et al. 2023 Nat. Commun. 14, 4911 (doi:10.1038/s41467-023-40533-1)). One of the key features physics provides is time: it describes rules and makes predictions about how activity patterns change over time, as governed by dynamics. We argue that computing can learn from physics in how it conceives time-namely, in terms of dynamics, i.e. the changing patterns of activity unfolding over time. In particular, we focus on the brain's intrinsic neural dynamics of spontaneous activity and how it 'uses' them for dynamic input processing and encoding in order to 'participate' in the world's physical time. By 'participate', we mean becoming part of the input dynamics: for instance, when listening or dancing to music, neural activity (via entrainment) and, consequently, mental activity follows the rhythm and dynamics of the music. This shapes the listener's experience, such as consciousness-the brain, and thus the subject, actively participates in music rhythms through temporal alignment (Northoff et al. 2023 Interface Focus13, 20220076 (doi:10.1098/rsfs.2022.0076)). Drawing on recent empirical evidence, we show that dynamic features such as the brain's scale-free activity and variability-which reflect an intrinsic temporal structure, or the brain's 'inner time'-are central to tracking and encoding input dynamics. Importantly, this processing is actively modulated rather than passively received, through the brain's own 'hidden dynamic repertoire'. Extending earlier discussions by Dreyfus and others, we argue that current computing devices, whether classical or natural (i.e. non-von Neumann machines), lack spontaneous activity and an inner time that can exert an active, rather than purely passive, influence on processing. As a result, they can neither actively process and encode input dynamics through their own inner time, nor 'use' or 'participate' in the dynamics of the world's physical time. Instead of 'being in time' and 'being in the world', current computing devices-and, by extension, artificial intelligence-are effectively 'locked out of time and world', meaning they are not directly connected to physical time. Unlike humans, they therefore cannot be characterized as 'being in time' or 'being in the world', which in turn prevents them from acquiring tacit or implicit knowledge, including the ability to navigate and behave flexibly within a continuously changing world. This article is part of the theme issue 'World models in natural and artificial intelligence'.
15 May 2026
Read appraisal →Journal of neural engineering
Uncovering relationships in multi-channel EEG data using principal Hessian directions and Ricci flow
Objective.The high-dimensional nature of multi-channel EEG data poses major challenges for downstream classification. Uncovering connectivity relationships and network structure between EEG channels can improve high-dimensional EEG signal recovery by identifying redundant inputs and guiding more targeted feature selection. In this study, we aim to demonstrate a systematic framework for extracting useful interchannel structure in EEG data.Approach.We present two complementary methods for inferring network structure and connectivity in EEG data: (1) a new supervised algorithm for inferring classification-relevant community structure based on principal Hessian directions (pHds), and (2) a discrete Ricci flow-based unsupervised community detection algorithm. We demonstrate these systematic methods on high-dimensional real-world EEG datasets involving classifying imagined digits versus non-digits and detecting emotional valence.Main results.We show that our pHd and Ricci flow methods-when combined-can detect interchannel relationships that meaningfully hold on unseen test classification data. Moreover, we demonstrate that this interchannel structure extracted via pHd and Ricci flow can enable subsequent improvements in downstream EEG signal classification.Significance.Our combined pHd-Ricci flow method expands the existing EEG preprocessing toolkit by offering a systematic framework for extracting meaningful network structure from high-dimensional EEG data. By facilitating more targeted and effective feature selection, our method has the potential to improve EEG signal recovery in real-world applications.
11 May 2026
Read appraisal →Neuropsychologia
Sparsity and memory constraints interact with training sequence to bias learning of associative maps.
Cognitive maps support inference and planning by representing associations between experiences encoded in memory. These map-like representations are thought to carry information not only about directly observed links but also about longer paths. The ability to make judgments based on multi-step associations varies with one's experience in an environment and with changes in memory abilities across the lifespan. However, it remains unclear exactly how representations of associative structure are influenced by learning curricula and memory constraints. Prior studies have suggested a tradeoff: memory representations can either be more integrated to improve inference, or more separated to recall distinct direct associations. Whether overlapping associations are experienced nearby in time (interleaved) or spaced apart (blocked) can bias memory representations toward integration or separation. However, key recent findings about how blocked versus interleaved experience bias integration or separation have been inconsistent. Here, we introduce a computational framework that helps reconcile these apparent discrepancies. Using neural network simulations of three separate memory-guided inference tasks, we show that variations in memory capacity and the sparsity of neural codes interact with learning sequence to shape network representations. Specifically, blocked learning promotes integration when memory capacity is low, while interleaved learning promotes integration when memory capacity is high. Integration is especially likely to result from representations formed when neural codes are both sparse and distributed. These results offer a principled computational account of how flexible, map-like representations can arise from experience and suggest avenues for individualized memory interventions to improve inference, generalization, and planning.
10 May 2026
Read appraisal →Neural networks : the official journal of the International Neural Network Society
Object-centric proto-symbolic behavioural reasoning from pixels
Autonomous intelligent agents must bridge computational challenges at disparate levels of abstraction, from the low-level spaces of sensory input and motor commands to the high-level domain of abstract reasoning and planning. A key question in designing such agents is how best to instantiate the representational space that will interface between these two levels-ideally without requiring supervision in the form of expensive data annotations. These objectives can be efficiently achieved by representing the world in terms of objects (grounded in perception and action). In this work, we present a novel, brain-inspired, deep-learning architecture that learns from pixels to interpret, control, and reason about its environment, using object-centric representations. We show the utility of our approach through tasks in synthetic environments that require a combination of (high-level) logical reasoning and (low-level) continuous control. Results show that the agent can learn emergent conditional behavioural reasoning, such as (A → B)∧(¬A → C), as well as logical composition (A → B)∧(A → C)⊢A → (B∧C) and XOR operations, and successfully controls its environment to satisfy objectives deduced from these logical rules. The agent can adapt online to unexpected changes in its environment and is robust to mild violations of its world model, thanks to dynamic internal desired goal generation. While the present results are limited to synthetic settings (2D and 3D activated versions of dSprites), which fall short of real-world levels of complexity, the proposed architecture shows how to manipulate grounded object representations, as a key inductive bias for unsupervised learning, to enable behavioral reasoning.
9 May 2026
Read appraisal →Epilepsy research
Machine learning in epilepsy
Epilepsy is a complex neurological disorder characterized by pathological processes that unfold across multiple biological scales, from cellular excitability and synaptic integration to large-scale network dynamics observable in electroencephalographic (EEG) recordings. While traditional analytical approaches have provided valuable insights, they often fail to capture high-dimensional and nonlinear structure of contemporary electrophysiological and clinical datasets. Consequently, machine learning (ML) has emerged as a powerful analytical framework in epilepsy research, although its rapid adoption has revealed a growing gap between algorithmic performance and biological interpretability. This review examines ML methods operating across three analytically distinct yet interconnected levels: (i) unsupervised learning for cellular-level phenotyping using high-dimensional electrophysiological data; (ii) supervised learning for EEG-based seizure detection and prediction; and (iii) multiscale modeling frameworks integrating neuronal and network dynamics. Rather than providing an exhaustive catalog of algorithms, we focus on inferential assumptions underlying ML applications, the methodological pitfalls constraining generalization and clinical relevance, and how ML-derived representations can be interpreted within established neurophysiological theory. We highlight that unsupervised ML facilitates identification of latent excitability phenotypes and trajectories obscured in traditional univariate analyses, while supervised ML has substantially advanced automated seizure detection and prediction, despite persistent challenges related to data leakage, class imbalance, and ambiguous preictal labeling. We argue that the most promising direction lies in embedding ML within multiscale mechanistic models, where data-driven inference facilitates parameter estimation and hypothesis generation rather than black-box prediction. By prioritizing interpretability, rigorous validation, and cross-scale integration, ML-enhanced multiscale frameworks offer a path toward clinically actionable models of epilepsy.
9 May 2026
Read appraisal →Journal of neural engineering
Physiologically inspired modeling of cortical dynamics through spiking neural networks
Objective. The characterization of neural activity underlying neurophysiological function presents a major challenge in computational neuroscience. Several methods have been proposed to investigate cortical network dynamics by reconstructing underlying neural activity from electroencephalography (EEG) signals. However, these methods generally pose significant mathematical challenges.Approach. This study introduces a novel framework to model the underlying brain activity network from a functional and physiologically-inspired perspective, combining spiking neural networks with EEG signal analysis. The dynamics of single neurons are described by the well-known Izhikevich model, and distinct populations of cortical inhibitory and excitatory neurons are employed to model experimental EEG recordings. Functional interactions among distinct populations are mathematically formalized through connective probabilities.Main results. The proposed framework is validated by testing it on synthetic data, as well as on two experimental datasets comprising data from 30 healthy subjects undergoing a cold-pressure test (CPT), and 36 subjects undergoing a mental arithmetic stressor. Experimental results suggest that the proposed framework provides novel and complementary insights into characterizing neuronal changes in comparison to standard EEG power analysis.Significance. The proposed framework constitutes a promising tool for functionally characterizing the underlying cortical dynamics under pathophysiological conditions.
20 Apr 2026
Read appraisal →PLoS computational biology
'Backpropagation and the brain' realized in cortical error neuron microcircuits
Neural responses to mismatches between expected and actual stimuli have been widely reported across different species. How does the brain use such error signals for learning? While global error signals can be useful, their ability to learn complex computation at the scale observed in the brain is lacking. In comparison, more local, neuron-specific error signals enable superior performance, but their computation and propagation remain unclear. Motivated by the breakthrough of deep learning, this has inspired the 'backpropagation and the brain' hypothesis, i.e., that the brain implements a form of the error backpropagation algorithm. In this work, we introduce a biologically motivated, multi-area cortical microcircuit model, implementing error backpropagation under consideration of recent physiological evidence. We model populations of cortical pyramidal cells acting as representation and error neurons, with bio-plausible local and inter-area connectivity, guided by experimental observations of connectivity of the primate visual cortex. In our model, all information transfer is biologically motivated, inference and learning occur without phases, and network dynamics demonstrably approximate those of error backpropagation. We show the capabilities of our model on a wide range of benchmarks, and compare to other models, such as dendritic hierarchical predictive coding. In particular, our model addresses shortcomings of other theories in terms of scalability to many cortical areas. Finally, we make concrete predictions, which differentiate it from other theories, and which can be tested experimentally.
19 Apr 2026
Read appraisal →PloS one
Topology-aware design of spiking neural networks via modular graph architectures
Spiking Neural Networks (SNNs) offer a biologically plausible and energy-efficient alternative to traditional artificial neural networks (ANNs), yet their design remains constrained by limited architectural flexibility and slow training dynamics. In this work, we introduce a novel SNN framework that leverages modular graph-based topologies and explicit synaptic delays to significantly enhance both training efficiency and classification performance. Our architecture, TANet-Tiny, incorporates structured graph stages with up to 32 nodes and diverse community-driven connectivity patterns derived from KMeans clustering, Louvain modularity, and Watts-Strogatz small-world models. We integrate these topologies into a topology-aware search space and explore them via a Spatio-Temporal Topology Sampling (STTS) approach, enabling the discovery of high-performing networks without exhaustive search. Experimental results on MNIST, CIFAR-10, and CIFAR-100 demonstrate that our modular designs achieve state-of-the-art accuracy while requiring 6-10 × fewer training epochs, with top-1 accuracy reaching 99.57% on MNIST and over 92% on CIFAR-10, all with reduced parameter counts. We introduce an accuracy-per-epoch metric to quantify training efficiency and show that modularity, rather than network size, is the critical driver of performance. This work lays the groundwork for scalable, interpretable, and low-latency SNN architectures suitable for deployment in neuromorphic and edge computing environments.
12 Apr 2026
Read appraisal →