PUBLICACIONES

Green AI aims to develop accurate AI models that are also sustainable without compromising the environment, especially in terms of carbon emissions. There are few studies on this topic in recommender systems, so we analyzed the trade-offs between recommendation performance and carbon footprint in session-based recommender systems. We use five public e-commerce datasets to predict the next item a user will interact with based solely on their past click events. The GRU4Rec algorithm and five unofficial reimplementations in different deep learning frameworks (Theano, PyTorch, TensorFlow, Keras, and Reckpack) are evaluated. The results indicate a strong effect of the loss function and dataset size on the carbon footprint without significantly affecting the accuracy metrics. We show evidence that the implementation choice for the same algorithm strongly affects the CO emitted, and optimized implementations do not sacrifice recommendation efficiency, which should be considered when choosing a framework or implementation for an algorithm.

A current problem in the philosophy of neuroscience consists in determining how to individuate cognitive capacities using neurobiological evidence. One recent proposal grounded on fundamental insights from mechanistic philosophy is an iterative strategy that cycles between neural mechanisms and cognitive capacities, using the former to individuate the latter and vice versa (Francken et al. Synthese, 200(5):378, 2022). However, this view cannot be applied to a fundamental aspect of research on cognitive capacities. Understanding a capacity requires delineating its behavioral profile by identifying the different effects or phenomena associated with it in different task settings. If mechanisms are necessary for integrating these phenomena into a single capacity in a bottom-up way, the iterative cyclic strategy requires a criterion for assessing the across-context identity of the mechanism, which is missing from this view. However, we argue that introducing this criterion turns the strategy into something else. It requires substituting its bottom-up phase with a bi-dimensional stage in which the capacity and its mechanism are individuated simultaneously through the identification of the generalization that connects them and the assessment of how the generalization behaves across contexts.

A key question in the analysis of discrete models for material defects, such as vortices in spin systems and superconductors or isolated dislocations in metals, is whether information on boundary energy for a domain can be sufficient for controlling the number of defects in the interior. We present a general combinatorial dipole-removal argument for a large class of discrete models including XY systems and screw dislocation models, allowing to prove sharp conditions under which controlled flux and boundary energy guarantee that minimizers with zero or one charges in the interior exist. The argument uses the max-flow min-cut theorem in combination with an ad-hoc duality for planar graphs, and is robust with respect to changes of the function defining the interaction energies.

According to social studies of artificial intelligence (AI), public AI controversies tend to dissipate relatively quickly despite well-documented risks and harms. The reasons for this lack of controversiality are beginning to be studied. Drawing on the framework of sociotechnical controversies, we analyze the de-escalation of contentious discussions observed in the AI legislative process by Chile's National Congress. Utilizing a qualitative approach, we tracked the deliberations hosted by the Chamber of Deputies and the Senate of Chile across 51 sessions between 2023 and 2024. We describe three processes of cooling down in the AI debates: (1) deflection of technology liability, (2) instrumentalization of technology policy, and (3) moralization of technology use. However, constructive exchanges appear in some circumstances, which allow us to foresee some favorable conditions for participation in the debates on AI regulation. This paper contributes to AI controversy studies by outlining cooling-down processes and conditions that foster dialogue and providing a critical perspective on the formation of AI regulation.

Deep neural networks (DNNs) struggle at systematic generalization (SG). Several studies have evaluated the possibility of promoting SG through the proposal of novel architectures, loss functions, or training methodologies. Few studies, however, have focused on the role of training data properties in promoting SG. In this work, we investigate the impact of certain data distributional properties, as inductive biases for the SG ability of a multi-modal language model. To this end, we study three different properties. First, data diversity, instantiated as an increase in the possible values a latent property in the training distribution may take. Second, burstiness, where we probabilistically restrict the number of possible values of latent factors on particular inputs during training. Third, latent intervention, where a particular latent factor is altered randomly during training. We find that all three factors significantly enhance SG, with diversity contributing an 89% absolute increase in accuracy in the most affected property. Through a series of experiments, we test various hypotheses to understand why these properties promote SG. Finally, we find that Normalized Mutual Information (NMI) between latent attributes in the training distribution is strongly predictive of out-of-distribution generalization. We find that a mechanism by which lower NMI induces SG is in the geometry of representations. In particular, we find that NMI induces more parallelism in neural representations (i.e., input features coded in parallel neural vectors) of the model, a property related to the capacity of reasoning by analogy.

Gait, a complex process unique to everyone, has been extensively studied in fields such as medicine, rehabilitation, and biomechanics. Gait laboratories, like those at Teletón centers, play a crucial role in analyzing and diagnosing human movement patterns during locomotion. This report focuses on standardizing data from different gait laboratories. This initiative addresses the lack of uniformity, which hinders the comparison and exchange of information between different rehabilitation centers. The proposal is part of the Movement Analysis Network project, an initiative by Politecnico di Milano, Oritel, and Universidad de Concepción, promoting collaboration, research, and continuous improvement in movement analysis. The proposed methodology particularly focuses on standardizing data acquired from BTS GaitLab and Vicon Systems. To achieve an effective comparison between data from these two systems, it is essential to define uniform labels and automate the processes that allow for the comparison of the information contained in the files. For this purpose, our script takes C3D and EMT files as input and generates table-formatted files separated by study type as output. Then, those files are used as input for a NoSQL database. This database facilitates further data comparison. This work enables the creation of standardized databases, which can be used to improve the effectiveness and reliability of diagnoses and treatments related to gait pathologies. In a future work, machine learning techniques could be used for gait pattern classification, creating new opportunities for more precise and personalized medical care, thereby positively impacting patients' quality of life by tailoring treatments to their specific needs.

"Generating differentially private (DP) synthetic data that closely resembles the original private data is a scalable way to mitigate privacy concerns in the current data-driven world. In contrast to current practices that train customized models for this task, we aim to generate DP Synthetic Data via APIs (DPSDA), where we treat foundation models as blackboxes and only utilize their inference APIs. Such API-based, training-free approaches are easier to deploy as exemplified by the recent surge in the number of API-based apps. These approaches can also leverage the power of large foundation models which are only accessible via their inference APIs. However, this comes with greater challenges due to strictly more restrictive model access and the need to protect privacy from the API provider. In this paper, we present a new framework called Private Evolution (PE) to solve this problem and show its initial promise on synthetic images. Surprisingly, PE can match or even outperform state-of-the-art (SOTA) methods without any model training. For example, on CIFAR10 (with ImageNet as the public data), we achieve FID <= 7.9 with privacy cost {\epsilon} = 0.67, significantly improving the previous SOTA from {\epsilon} = 32. We further demonstrate the promise of applying PE on large foundation models such as Stable Diffusion to tackle challenging private datasets with a small number of high-resolution images. The code and data are released at this https URL."

"Delle Rose et al. (COLT’23) introduced an effective version of the Vapnik-Chervonenkis dimension, and showed that it characterizes improper PAC learning with total computable learners. In this paper, we introduce and study a similar effectivization of the notion of Littlestone dimension. Finite effective Littlestone dimension is a necessary condition for computable online learning but is not a sufficient one—which we already establish for classes of the effective Littlestone dimension 2. However, the effective Littlestone dimension equals the optimal mistake bound for computable learners in two special cases: a) for classes of Littlestone dimension 1 and b) when the learner receives as additional information a bound on the numbers to be guessed. Interestingly, finite effective Littlestone dimension also guarantees that the class consists only of computable functions. Keywords: Online learning, Littlestone dimension, computable machine learning"

Blood oxygen saturation (SpO2) is vital for patient monitoring, particularly in clinical settings. Traditional SpO2 estimation methods have limitations, which can be addressed by analyzing photoplethysmography (PPG) signals with artificial intelligence (AI) techniques. This systematic review, following PRISMA guidelines, analyzed 183 unique references from WOS, PubMed, and Scopus, with 26 studies meeting the inclusion criteria. The review examined AI models, key features, oximeters used, datasets, tested saturation intervals, and performance metrics while also assessing bias through the QUADAS-2 criteria. Linear regression models and deep neural networks (DNNs) emerged as the leading AI methodologies, utilizing features such as statistical metrics, signal-to-noise ratios, and intricate waveform morphology to enhance accuracy. Gaussian Process models, in particular, exhibited superior performance, achieving Mean Absolute Error (MAE) values as low as 0.57% and Root Mean Square Error (RMSE) as low as 0.69%. The bias analysis highlighted the need for better patient selection, reliable reference standards, and comprehensive SpO2 intervals to improve model generalizability. A persistent challenge is the reliance on non-invasive methods over the more accurate arterial blood gas analysis and the limited datasets representing diverse physiological conditions. Future research must focus on improving reference standards, test protocols, and addressing ethical considerations in clinical trials. Integrating AI with traditional physiological models can further enhance SpO2 estimation accuracy and robustness, offering significant advancements in patient care.

"Chemical Shift Imaging (CSI) or Chemical Shift Encoded Magnetic Resonance Imaging (CSE-MRI) enables the quantification of different chemical species in the human body, and it is one of the most widely used imaging modalities used to quantify fat in the human body. Although there have been substantial improvements in the design of signal acquisition protocols and the development of a variety of methods for the recovery of parameters of interest from the measured signal, it is still challenging to obtain a consistent and reliable quantification over the entire field of view. In fact, there are still discrepancies in the quantities recovered by different methods, and each exhibits a different degree of sensitivity to acquisition parameters such as the choice of echo times. Some of these challenges have their origin in the signal model itself. In particular, it is non-linear, and there may be different sets of parameters of interest compatible with the measured signal. For this reason, a thorough analysis of this model may help mitigate some of the remaining challenges, and yield insight into novel acquisition protocols. In this work, we perform an analysis of the signal model underlying CSI, focusing on finding suitable conditions under which recovery of the parameters of interest is possible. We determine the sources of non-identifiability of the parameters, and we propose a reconstruction method based on smooth non-convex optimization under convex constraints that achieves exact local recovery under suitable conditions. A surprising result is that the concentrations of the chemical species in the sample may be identifiable even when other parameters are not. We present numerical results illustrating how our theoretical results may help develop novel acquisition techniques, and showing how our proposed recovery method yields results comparable to the state-of-the-art. "