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.

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.

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.

Radar sensing technologies offer several advantages over other gesture input modalities, such as the ability to reliably sense human movements, a reasonable deployment cost, insensitivity to ambient conditions such as light, temperature, and the ability to preserve anonymity. These advantages come at the price of high processing complexity mainly due to the spatio-temporal variations of gesture articulation performed by different people. Deep learning methods, such as CNN-LSTM and 3D CNN-LSTM, have a high potential to recognize radar-based gestures but usually require hundreds or thousands of labeled training samples and high processing power. Asking a lot of people to acquire a lot of gestures is particularly tedious and tiring to the point of being unrealistic. To overcome these challenges, we propose FORTE, a hand gesture recognition with few samples based on an optimized CNN architecture working on pre-processed raw data. Using a k=5-fold cross-validation, we define and compare three alternative CNNs for recognizing hand gestures acquired in a semi-mobile context of use with a portable radar attached to a smartphone. The best CNN reaches an accuracy of 94.96% with a precision of 95.92% and a recall of 96.03% for a dataset composed of solely 5 participants producing 2 samples for 20 classes covering 1 pointing, 2 pantomimic, 3 iconic, and 14 semaphoric gestures. We suggest some implications for designing radar-based gestures and we discuss the limitations of this approach.

Mathematical reasoning is one of the most impressive achievements of human intellect but remains a formidable challenge for artificial intelligence systems. In this work we explore whether modern deep learning architectures can learn to solve a symbolic addition task by discovering effective arithmetic procedures. Although the problem might seem trivial at first glance, generalizing arithmetic knowledge to operations involving a higher number of terms, possibly composed by longer sequences of digits, has proven extremely challenging for neural networks. Here we show that universal transformers equipped with local attention and adaptive halting mechanisms can learn to exploit an external, grid-like memory to carry out multi-digit addition. The proposed model achieves remarkable accuracy even when tested with problems requiring extrapolation outside the training distribution; most notably, it does so by discovering human-like calculation strategies such as place value alignment.

The accurate detection and quantification of submerged targets has been recognized as a key challenge in marine exploration, one that traditional census approaches cannot handle efficiently. Here we present a deep learning approach to detect the pattern of a moving fish from the reflections of an active acoustic emitter. To allow for real-time detection, we use a convolutional neural network, which provides the simultaneous labeling of a large buffer of signal samples. This allows to capture the structure of the reflecting signal from the moving target and to separate it from clutter reflections. We evaluate system performance both on synthetic (simulated) data, as well as on real data recorded over 50 sea experiments in a variety of sea conditions. When tested on real signals, the network trained on simulated patterns showed non-trivial detection capabilities, suggesting that transfer learning can be a viable approach in these scenarios, where tagged data is often lacking. However, training the network directly on the real reflections with data augmentation techniques allowed to reach a more favorable precision-recall trade-off, approaching an ideal detection bound. We also evaluate an alternative model based on recurrent neural networks which, despite exhibiting slightly inferior performance, could be applied in scenarios requiring on-line processing of the reflection sequence.

Complex phenomena are generally modeled with sophisticated simulators that, depending on their accuracy, can be very demanding in terms of computational resources and simulation time. Their time-consuming nature, together with a typically vast parameter space to be explored, make simulation- based optimization often infeasible. In this work, we present a method that enables the optimization of complex systems through Machine Learning (ML) techniques. We show how well-known learning algorithms are able to reliably emulate a complex simulator with a modest dataset obtained from it. The trained emulator is then able to yield values close to the simulated ones in virtually no time. Therefore, it is possible to perform a global numerical optimization over the vast multi-dimensional parameter space, in a fraction of the time that would be required by a simple brute-force search. As a testbed for the proposed methodology, we used a network simulator for next-generation mmWave cellular systems. After simulating several antenna configurations and collecting the resulting network-level statistics, we feed it into our framework. Results show that, even with few data points, extrapolating a continuous model makes it possible to estimate the global optimum configuration almost instantaneously. The very same tool can then be used to achieve any further optimization goal on the same input parameters in negligible time.

Wireless transmissions in water are mostly carried out via long-range (but low-rate) underwater acoustic communications, or short-range (but high-rate) underwater optical communications. In this paper, we are interested in finding out whether a statistical relationship exists between underwater acoustics and optics. Besides the theoretical interest of such a relationship, predicting the quality of the optical link through acoustics is also relevant in the context of a multimodal system with both acoustics and optics. Our study is based on a large data set acquired during the NATO ALOMEX 2015 expedition. During this experiment, we simultaneously measured several characteristics of the acoustic and optical links at multiple locations, reflecting a diversity of sea environments. Our results show a strong correlation between the properties of the acoustic link and the reliability of optical communications. This correlation makes it possible to predict the state of the underwater optical link at a certain depth and range. Due to the complexity of the acoustic and optical channels, we could not find the source of this correlation. This paper is, therefore, aimed to stimulate a theoretical study of the mutual properties of underwater acoustic and optical communication links. For reproducibility, we share the processed data from the experiment.

Several techniques for wireless networking, such as opportunistic spectrum access, or self-healing networks, may be seen as using a form of cognition, meaning that they mimic reasoning processes of intelligent beings. We propose to expand this cognition-based process by exploiting the parallel processing power of the infrastructure, so as to go beyond cognition as is meant by these approaches. We leverage novel approaches, taken from cognitive science and artificial intelligence, involving not only supervised but also unsupervised learning, and we envision their application to systems for video over wireless. The transmission of multimedia content, and its adaptation to the condition of the communication infrastructure, i.e., the wireless channel or the content delivery network, are envisioned as particularly critical steps for the development of latest generation mobile networks. For this scenario, we propose and evaluate a video classifier based on a Restricted Boltzmann Machine that tries to extract abstract features of videos from the analysis of the sizes of a few coded frames. These features can then be exploited by the communication network itself to optimize video transmission based on its content.