Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 6 appraisals
Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
A sentence is worth a thousand pictures: can large language models understand hum4n L4ngu4ge and the W0rld behind W0rds?
The current generation of large language models (LLMs) has been linked to claims about human-like linguistic performance, and their applications are hailed both as a step towards artificial general intelligence and as a major advance in understanding the cognitive and even neural basis of human language. To assess these claims, first, we analysed the contribution of LLMs as theoretically informative representations of a target cognitive system versus atheoretical mechanistic tools. Second, we evaluated the models' ability to see the bigger picture through top-down feedback from higher levels of processing, which requires grounding in previous expectations and past world experience. We hypothesize that since models lack grounded cognition, they cannot take advantage of these features and instead solely rely on fixed associations between represented words and word vectors. To assess this, we ran a novel leet task (l33t t4sk), which requires decoding sentences in which letters are systematically replaced by numbers. In line with our hypothesis, the results suggest that humans excel in this task, whereas models struggle. We interpret these results by identifying the key abilities that are still missing from the current state of development of these models, which require solutions that go beyond increased system scaling. This article is part of the theme issue 'World models in natural and artificial intelligence'.
15 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 →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 →Neural networks : the official journal of the International Neural Network Society
Semi-supervised classification and projection with adaptive flexible structure optimal graph
Graph-based semi-supervised learning (GSSL) has received more attention in recent years. Many existing methods obtain a fixed similarity graph based on the original data. Affected by the redundant information and noise of the original data, the obtained similarity graph is not optimal, which greatly affects the subsequent work. And the similarity graphs obtained by these methods are fully connected, so the local structure information of the data cannot be preserved. Moreover, many GSSL methods project the original data into a low-dimensional space through a linear mapping function. However, linear functions may not be appropriate for data embedded in a nonlinear manifold. In order to overcome these shortcomings mentioned above, we relax the linear mapping function and impose a ℓ0-norm constraint on the similarity graph to gain an adaptive flexible structure optimal graph. Furthermore, incorporating the principle of maximum separability, an efficient GSSL method is proposed, named semi-supervised classification and projection with adaptive flexible structure optimal graph (SAFSG). Combining the construction of similarity graphs and label propagation, SAFSG can simultaneously get an adaptive flexible structure optimal graph, a label prediction matrix and a projection matrix. In addition, an efficient iterative algorithm for optimizing SAFSG is proposed. Finally, we conduct experiments on more than ten benchmark datasets, and the experimental results show that SAFSG performs satisfactorily in both classification and projection.
9 May 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 →PloS one
Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection.
Deepfake (DF) content poses a major challenge to digital media authentication, that can mimic facial movements, creating realistic replicas that risk spreading misinformation and enabling harassment, as can be seen in Bangladesh. Previous studies applied residual convolutional blocks or attention mechanisms for DF detection; however, these approaches often treat these components in isolation and lack explicit consideration of fairness, out-of-distribution generalization, statistical analysis or artifact-focused detection. Our study introduces an approach, called ARC-Net, which uses a combination of attention and residual convolutional layers along with the EfficientNet B0 base, the attention mechanism of which allows the model to pay more attention to details that could resemble the ones seen in non-perfectly executed DF, which enhances the model's potential to distinguish them, addressing limitations in pre-trained and also advanced DL models. A dataset of 500 real images from Bangladesh combined with a 140k real and fake faces dataset was used to train and test our model alongside four pre-trained DL models. ARC-Net performed much better than the other traditional and state of art methods with 99% accuracy, 1.0 precision, 0.97 recall and 0.98 F1 score, reaching the highest level of reliability in spotting DF images. To assess the external reliability and generalizability of ARC-Net, the model was evaluated on the Deepfake Dataset and the Deepfake Database datasets, achieving consistently high and balanced performance across different scales. Three out-of-distribution (OOD) experiments were conducted, where the first one evaluated South Asian images, showing that incorporating fewer than one percent real Bangladeshi images reduced the false positive rate by more than half and improved probability calibration, while the remaining two cross-dataset experiments demonstrated strong transferability. An ablation study further showed the impact of different components within a model by systematically removing or modifying them and statistical significance between competing classifiers was assessed using McNemar's Statistical test. In addition, we've applied Explainable AI (XAI) techniques, Grad-CAM and LIME to offer transparency of the results, as giving attention to the facial region is also important for DF detection. This study will help advance DF detection by integrating ARC-Net's attention-residual mechanisms and XAI, offering insights for developing models in security and media forensics.
7 Apr 2026
Read appraisal →