Counting the number of items in a visual scene remains a fundamental yet challenging task in computer vision. Traditional approaches rely on domain-specific counting architectures trained on datasets with predefined object categories. However, recent progress in developing large-scale multimodal vision-language models (VLMs) suggests that these domain-general architectures may offer a flexible alternative for open-set object counting. In this study, we systematically compare the performance of state-of-the-art specialized counting architectures against VLMs across two established counting datasets and a novel benchmark designed for fine-grained control over visual properties of test images. Our findings show that most VLMs can approximately enumerate the number of items in a visual scene, matching or even surpassing the performance of specialized computer vision architectures. Notably, enumeration accuracy significantly improves when VLMs are prompted to generate intermediate representations—such as object locations and verbal labels—prior to counting. Nevertheless, none of the models can reliably count the number of objects in complex visual scenes, showing that further research is still needed to create AI systems capable of robust counting in realistic environments. We made our code publicly available.( https://github.com/Ku1-NaN/VLM_VisEnum ).

Reliably counting and generating sequences of items remain a significant challenge for neural networks, including Large Language Models (LLMs). Indeed, although this capability is readily handled by rule-based symbolic systems based on serial computation, learning to systematically deploy counting procedures is difficult for neural models, which should acquire these skills through learning. Previous research has demonstrated that recurrent architectures can only approximately track and enumerate sequences of events: it remains unclear whether modern deep learning systems, including LLMs, can deploy systematic counting procedures over sequences of discrete symbols. This paper aims to fill this gap by investigating the sequential enumeration abilities of five state-of-the-art LLMs, including proprietary, open-source, and reasoning models. We probe LLMs in sequential naming and production tasks involving lists of letters and words, adopting a variety of prompting instructions to explore the role of chain-of-thought in the spontaneous emerging of counting strategies. We also evaluate open-source models with the same architecture but increasing size to assess whether mastering counting principles follows scaling laws, and we analyze embedding dynamics during sequential enumeration to investigate the emergent encoding of numerosity. We find that some LLMs are indeed capable of deploying counting procedures when explicitly prompted, but none of them spontaneously engage in counting when simply asked to enumerate the number of items in a sequence. Our results suggest that, despite their impressive emergent abilities, LLMs cannot yet robustly and systematically deploy counting procedures, highlighting a persistent gap between neural and symbolic approaches to compositional generalization.

Functional Magnetic Resonance Imaging (fMRI) provides spatio-temporal maps of brain activity; however, extracting the rich information they contain is challenging. Traditional approaches use only summary statistics, losing details that might be hidden in the complex temporal dynamics. Deep neural networks are emerging as an apt solution in this context, given their ability to handle vast amounts of structured data. In this paper, we consider two widely studied fMRI datasets: the Human Connectome Project for connectome fingerprinting, and ABIDE for autism classification. We aim to understand how handling the temporal and spatial dimensions could influence the performance of the models and their interpretability. Specifically, we compare neural network models with architectural biases toward temporal, spatial, or combined spatio-temporal features. The results of our analysis show that existing methods exploiting the spatial dimension, or spatio-temporal hybrids, are not competitive with simpler ones considering the temporal dimension only, such as LSTM. Additionally, we propose a contrastive learning approach for connectome fingerprinting, enabling robust individual identification without requiring access to all subjects during training. Our findings suggest that explicit graph modeling of the interaction between brain regions introduces complexity without improving performance, thereby challenging current trends.

The ongoing decline in global biodiversity constitutes a critical challenge for environmental science, necessitating the prompt development of effective monitoring frameworks and conservation protocols to safeguard the structure and function of natural ecosystems. Recent progress in ecoacoustic monitoring, supported by advances in artificial intelligence, might finally offer scalable tools for systematic biodiversity assessment. In this study, we evaluate the performance of BirdNET, a state-of-the-art deep learning model for avian sound recognition, in the context of selected bird species characteristic of the Italian Alpine region. To this end, we assemble a comprehensive, manually annotated audio dataset targeting key regional species, and we investigate a variety of strategies for model adaptation, including fine-tuning with data augmentation techniques to enhance recognition under challenging recording conditions. As a baseline, we also develop and evaluate a simple Convolutional Neural Network (CNN) trained exclusively on our domain-specific dataset. Our findings indicate that BirdNET performance can be greatly improved by fine-tuning the pre-trained network with data collected within the specific regional soundscape, outperforming both the original BirdNET and the baseline CNN by a significant margin. These findings underscore the importance of environmental adaptation and data variability for the development of automated ecoacoustic monitoring devices while highlighting the potential of deep learning methods in supporting conservation efforts and informing soundscape management in protected areas.

Number sense is a core cognitive ability supporting various adaptive behaviors and is foundational for mathematical learning. Here, we study its emergence in unsupervised generative models through the lens of rate-distortion theory (RDT), a normative framework for understanding information processing under limited resources. We train β-Variational Autoencoders – which embody key formal principles of RDT – on synthetic images containing varying numbers of items, as commonly used in numerosity perception research. We systematically vary the encoding capacity and assess the models' sensitivity to numerosity and the robustness of the emergent numerical representations through a comprehensive set of analyses, including numerosity estimation and discrimination tasks, latent-space analysis, generative capabilities and generalization to novel stimuli. In line with RDT, we find that behavioral performance in numerosity perception and the ability to extract numerosity unconfounded by non-numerical visual features scale with encoding capacity according to a power law. At high capacity, the unsupervised model develops a robust neural code for numerical information, with performance closely approximating a supervised model explicitly trained for visual enumeration. It exhibits strong generative abilities and generalizes well to novel images, whereas at low capacity, the model shows marked deficits in numerosity perception and representation. Finally, comparison with human data shows that models trained at intermediate capacity levels span the full range of human behavioral performance while still developing a robust emergent numerical code. In sum, our results show that unsupervised generative models can develop a number sense and demonstrate that rate-distortion theory provides a powerful information-theoretic framework for understanding how capacity constraints shape numerosity perception.

Many animal species can approximately judge the number of objects in a visual scene at a single glance, and humans can further determine the exact cardinality of a set by deploying systematic counting procedures. In contrast, it has been observed that even state-of-the-art AI systems have very limited enumeration skills. In this work, we propose two benchmark tasks inspired by cognitive science that allow to precisely evaluate the visual enumeration capabilities of multimodal foundation models, thereby providing an objective measure of their number sense and counting level. We consider popular visual question answering models (BLIP, LLaVA and ViLT) as well as advanced image-to-text (Gemini, GPT and Qwen) and text-to-image (DALL-E, FLUX and Stable Diffusion) AI systems. Our analyses show that even the most advanced models cannot reliably name the number of objects in simple visual stimuli or generate images containing a target number of items, as indexed by their low accuracy in both types of tasks. Especially for numbers outside the subitizing range, their responses are often far from the target numerosity, and, in stark contrast with human behavior, in many cases the distribution of errors depends on the object category. We also observe some striking mistakes with small numbers. Our findings demonstrate that developing an intuitive visual understanding of number remains challenging for AI models and that merely increasing model size might not be a viable strategy to promote the emergence of systematic counting skills. We release the full code of our benchmark to facilitate the evaluation of enumeration skills in future AI systems.

Affective reactions have deep biological foundations; however, in humans, the development of emotion concepts is also shaped by language and higher-order cognition. A recent breakthrough in artificial intelligence (AI) has been the creation of multimodal language models that exhibit impressive intellectual capabilities, but their responses to affective stimuli have not been investigated. Here, we study whether state-of-the-art multimodal systems can emulate human emotional ratings on a standardized set of images, in terms of affective dimensions and basic discrete emotions. The AI judgements correlate surprisingly well with the average human ratings: given that these systems were not explicitly trained to match human affective reactions, this suggests that the ability to visually judge emotional content can emerge from statistical learning over large-scale databases of images paired with linguistic descriptions. Besides showing that language can support the development of rich emotion concepts in AI, these findings have broad implications for sensitive use of multimodal AI technology.

Hierarchical generative models can produce data samples based on the statistical structure of their training distribution. This capability can be linked to current theories in computational neuroscience, which propose that spontaneous brain activity at rest is the manifestation of top-down dynamics of generative models detached from action-perception cycles. A popular class of hierarchical generative models is that of Deep Belief Networks (DBNs), which are energy-based deep learning architectures that can learn multiple levels of representations in a completely unsupervised way exploiting Hebbian-like learning mechanisms. In this work, we study the generative dynamics of a recent extension of the DBN, the iterative DBN (iDBN), which more faithfully simulates neurocognitive development by jointly tuning the connection weights across all layers of the hierarchy. We characterize the number of states visited during top-down sampling and investigate whether the heterogeneity of visited attractors could be increased by initiating the generation process from biased hidden states. To this end, we train iDBN models on well-known datasets containing handwritten digits and pictures of human faces, and show that the ability to generate diverse data prototypes can be enhanced by initializing top-down sampling from "chimera states", which represent high-level features combining multiple abstract representations of the sensory data. Although the models are not always able to transition between all potential target states within a single-generation trajectory, the iDBN shows richer top-down dynamics in comparison to a shallow generative model (a single-layer Restricted Bolzamann Machine). We further show that the generated samples can be used to support continual learning through generative replay mechanisms. Our findings suggest that the top-down dynamics of hierarchical generative models is significantly influenced by the shape of the energy function, which depends both on the depth of the processing architecture and on the statistical structure of the sensory data.

Building neural systems that can learn to execute symbolic algorithms is a challenging open problem in artificial intelligence, especially when aiming for strong generalization and out-of-distribution performance. In this work, we introduce a general framework for learning convergent term rewriting systems using a neuro-symbolic architecture inspired by the rewriting algorithm itself. We present two modular implementations of such architecture: the Neural Rewriting System (NRS) and the Fast Neural Rewriting System (FastNRS). As a result of algorithmic-inspired design and key architectural elements, both models can generalize to out-of-distribution instances, with FastNRS offering significant improvements in terms of memory efficiency, training speed, and inference time. We evaluate both architectures on four tasks involving the simplification of mathematical formulas and further demonstrate their versatility in a multi-domain learning scenario, where a single model is trained to solve multiple types of problems simultaneously. The proposed system significantly outperforms two strong neural baselines: the Neural Data Router, a recent transformer variant specifically designed to solve algorithmic problems, and GPT-4o, one of the most powerful general-purpose large-language models. Moreover, our system matches or outperforms the latest o1-preview model from OpenAI that excels in reasoning benchmarks.

Humans share with many animal species the ability to perceive and approximately represent the number of objects in visual scenes. This ability improves throughout childhood, suggesting that learning and development play a key role in shaping our number sense. This hypothesis is further supported by computational investigations based on deep learning, which have shown that numerosity perception can spontaneously emerge in neural networks that learn the statistical structure of images with a varying number of items. However, neural network models are usually trained using synthetic datasets that might not faithfully reflect the statistical structure of natural environments, and there is also growing interest in using more ecological visual stimuli to investigate numerosity perception in humans. In this work, we exploit recent advances in computer vision algorithms to design and implement an original pipeline that can be used to estimate the distribution of numerosity and non-numerical magnitudes in large-scale datasets containing thousands of real images depicting objects in daily life situations. We show that in natural visual scenes the frequency of appearance of different numerosities follows a power law distribution. Moreover, we show that the correlational structure for numerosity and continuous magnitudes is stable across datasets and scene types (homogeneous vs. heterogeneous object sets). We suggest that considering such "ecological" pattern of covariance is important to understand the influence of non-numerical visual cues on numerosity judgements.

Functional magnetic resonance imaging (fMRI) has revolutionized neuroscience, providing insights into brain activity through the analysis of variations in blood oxygen level dependent (BOLD) contrast. Deep learning is considered a promising approach to discover intricate patterns within such high-dimensional signals, but at the same time its application poses novel methodological issues, particularly regarding the normalization of time series associated with different brain areas. This paper addresses the impact of different normalization methods on the predictive performance of deep learning models, providing the first comprehensive comparison in the neuroimaging domain. Through experiments with two predictive models on two fMRI classification tasks, we demonstrate the significant influence of data normalization on the accuracy and interpretability of deep learning models in fMRI data analysis.

Impaired numerosity perception in developmental dyscalculia (low "number acuity") has been interpreted as evidence of reduced representational precision in the neurocognitive system supporting non-symbolic number sense. However, recent studies suggest that poor numerosity judgments might stem from stronger interference from non-numerical visual information, in line with alternative accounts that highlight impairments in executive functions and visuospatial abilities in the etiology of dyscalculia. To resolve this debate, we used a psychophysical method designed to disentangle the contribution of numerical and non-numerical features to explicit numerosity judgments in a dot comparison task and we assessed the relative saliency of numerosity in a spontaneous categorization task. Children with dyscalculia were compared to control children with average mathematical skills matched for age, IQ, and visuospatial memory. In the comparison task, the lower accuracy of dyscalculics compared to controls was linked to weaker encoding of numerosity, but not to the strength of non-numerical biases. Similarly, in the spontaneous categorization task, children with dyscalculia showed a weaker number-based categorization compared to the control group, with no evidence of a stronger influence of non-numerical information on category choice. Simulations with a neurocomputational model of numerosity perception showed that the reduction of representational resources affected the progressive refinement of number acuity, with little effect on non-numerical bias in numerosity judgments. Together, these results suggest that impaired numerosity perception in dyscalculia cannot be explained by increased interference from non-numerical visual cues, thereby supporting the hypothesis of a core number sense deficit.Research Highlights A strongly debated issue is whether impaired numerosity perception in dyscalculia stems from a deficit in number sense or from poor executive and visuospatial functions. Dyscalculic children show reduced precision in visual numerosity judgments and weaker number-based spontaneous categorization, but no increasing reliance on continuous visual properties. Simulations with deep neural networks demonstrate that reduced neural/computational resources affect the developmental trajectory of number acuity and account for impaired numerosity judgments. Our findings show that weaker number acuity in developmental dyscalculia is not necessarily related to increased interference from non-numerical visual cues.

The recent scientific literature abounds in proposals of seizure forecasting methods that exploit machine learning to automatically analyze electroencephalogram (EEG) signals. Deep learning algorithms seem to achieve a particularly remarkable performance, suggesting that the implementation of clinical devices for seizure prediction might be within reach. However, most of the research evaluated the robustness of automatic forecasting methods through randomized cross-validation techniques, while clinical applications require much more stringent validation based on patient-independent testing. In this study, we show that automatic seizure forecasting can be performed, to some extent, even on independent patients who have never been seen during the training phase, thanks to the implementation of a simple calibration pipeline that can fine-tune deep learning models, even on a single epileptic event recorded from a new patient. We evaluate our calibration procedure using two datasets containing EEG signals recorded from a large cohort of epileptic subjects, demonstrating that the forecast accuracy of deep learning methods can increase on average by more than 20%, and that performance improves systematically in all independent patients. We further show that our calibration procedure works best for deep learning models, but can also be successfully applied to machine learning algorithms based on engineered signal features. Although our method still requires at least one epileptic event per patient to calibrate the forecasting model, we conclude that focusing on realistic validation methods allows to more reliably compare different machine learning approaches for seizure prediction, enabling the implementation of robust and effective forecasting systems that can be used in daily healthcare practice.

Many research efforts are being spent to discover predictive markers of seizures, which would allow to build forecasting systems that could mitigate the risk of injuries and clinical complications in epileptic patients.Although electroencephalography (EEG) is the most widely used tool to monitor abnormal brain electrical activity, no commercial devices can reliably anticipate seizures from EEG signal analysis at present.Recent advances in Artificial Intelligence, particularly deep learning algorithms, show promise in enhancing EEG classifier forecasting accuracy by automatically extracting relevant spatio-temporal features from EEG recordings.In this study, we systematically compare the predictive accuracy of two leading deep learning architectures: recurrent models based on Long Short-Term Memory networks (LSTMs) and Convolutional Neural Networks (CNNs).To this aim, we consider a data set of long-term, continuous multi-channel EEG recordings collected from 29 epileptic patients using a standard set of 20 channels.Our results demonstrate the superior performance of deep learning algorithms, which can achieve up to 99% accuracy, sensitivity, and specificity compared to more traditional machine learning approaches, which settle around 75% in all evaluation metrics.Our results also show that giving as input the recordings from all electrodes allows to exploit useful channel correlations to learn more robust predictive features, compared to convolutional models that treat each channel independently.We conclude that deep learning architectures hold promise for enhancing the diagnosis and prediction of epileptic seizures, offering potential benefits to those affected by such invalidating neurological conditions.

The brain exhibits a gradual transition in responses to visual event duration and frequency through the visual processing hierarchy: from monotonically increasing to timing-tuned responses. Over their hierarchies, properties of both response types are progressively transformed. Here, we implement simulations based on artificial neural networks to investigate the requirements of neural systems for the emergence of such responses and their properties' transformations. We see that recurrent networks develop monotonic responses whose properties' progressions over network layers resemble those over brain areas. Furthermore, recurrent networks can further develop tuned responses, but only with training, a gradual transition between monotonic and tuned responses emerges. Particularly, if this training is done on predictable sequences, the tuned properties' progressions resemble those observed in the brain. These results suggest that the emergence of visual timing-tuned responses and the subsequent hierarchical transformations of these responses result from recurrent neural computation and predictive processing of sensory event timing. ### Competing Interest Statement The authors have declared no competing interest.

Large Language Models (LLMs) have revolutionized the field of Natural Language Processing thanks to their ability to reuse knowledge acquired on massive text corpora on a wide variety of downstream tasks, with minimal (if any) tuning steps. At the same time, it has been repeatedly shown that LLMs lack systematic generalization, which allows to extrapolate the learned statistical regularities outside the training distribution. In this work, we offer a systematic benchmarking of GPT-4, one of the most advanced LLMs available, on three algorithmic tasks characterized by the possibility to control the problem difficulty with two parameters. We compare the performance of GPT-4 with that of its predecessor (GPT-3.5) and with a variant of the Transformer-Encoder architecture recently introduced to solve similar tasks, the Neural Data Router. We find that the deployment of advanced prompting techniques allows GPT-4 to reach superior accuracy on all tasks, demonstrating that state-of-the-art LLMs constitute a very strong baseline also in challenging tasks that require systematic generalization.

While several methods have been proposed to assess the influence of continuous visual cues in parallel numerosity estimation, the impact of temporal magnitudes on sequential numerosity judgments has been largely ignored. To overcome this issue, we extend a recently proposed framework that makes it possible to separate the contribution of numerical and non-numerical information in numerosity comparison by introducing a novel stimulus space designed for sequential tasks. Our method systematically varies the temporal magnitudes embedded into event sequences through the orthogonal manipulation of numerosity and two latent factors, which we designate as “duration” and “temporal spacing”. This allows us to measure the contribution of finer-grained temporal features on numerosity judgments in several sensory modalities. We validate the proposed method on two different experiments in both visual and auditory modalities: results show that adult participants discriminated sequences primarily by relying on numerosity, with similar acuity in the visual and auditory modality. However, participants were similarly influenced by non-numerical cues, such as the total duration of the stimuli, suggesting that temporal cues can significantly bias numerical processing. Our findings highlight the need to carefully consider the continuous properties of numerical stimuli in a sequential mode of presentation as well, with particular relevance in multimodal and cross-modal investigations. We provide the complete code for creating sequential stimuli and analyzing participants’ responses.

Modern neural network architectures still struggle to learn algorithmic procedures that require to systematically apply compositional rules to solve out-of-distribution problem instances. In this work, we focus on formula simplification problems, a class of synthetic benchmarks used to study the systematic generalization capabilities of neural architectures. We propose a modular architecture designed to learn a general procedure for solving nested mathematical formulas by only relying on a minimal set of training examples. Inspired by rewriting systems, a classic framework in symbolic artificial intelligence, we include in the architecture three specialized and interacting modules: the Selector, trained to identify solvable sub-expressions; the Solver, mapping sub-expressions to their values; and the Combiner, replacing sub-expressions in the original formula with the solution provided by the Solver. We benchmark our system against the Neural Data Router, a recent model specialized for systematic generalization, and a state-of-the-art large language model (GPT-4) probed with advanced prompting strategies. We demonstrate that our approach achieves a higher degree of out-of-distribution generalization compared to these alternative approaches on three different types of formula simplification problems, and we discuss its limitations by analyzing its failures.

Due to their central role in marine ecosystems, top predators such as dolphins are indispensable bioindicators in long-term ecological monitoring programs. As keystone species occupying the highest trophic levels, these predators exert top-down control by regulating prey populations and maintaining the ecosystem in balance. Their abundance, distribution and behavior are reliable indicators of ecosystem health and integrity. Several attempts have been made to quantify the behavior and abundance of dolphins by detecting their vocalizations, mostly using convolutional neural networks (CNNs) employed for pattern recognition over spectrogram images. However, the performance of automatic detection systems often strongly depends on the training distribution, making it difficult to generalize to other dolphin species, environmental conditions and recording devices. In this paper, we highlight the problem of robustness in CNN whistle detection and offer a possible direction to move forward through heatmaps: a tool to investigate which particular features in the spectrogram are guiding the CNN detection. Qualitative analysis reveals spectral and temporal features that can serve as keys to developing more robust CNN models for dolphin whistle detection.