Publications SISSA
URI permanente per questa collezione
Sfogliare
Immissioni recenti
1 - 5 di 14954
-
PubblicazioneNoisy models of the ventral stream reveal the impact of recurrence and learned representations on information processing timescales( 2026)In vision neuroscience, the temporal dynamics of the sensory stream and of its neural representations are thought to be deeply linked to the function of the hierarchy of cortical areas that deal with object recognition, known as the visual ventral stream. Neural representations that are invariant under identity-preserving object transformations, and therefore allow for efficient learning of object identity, are theorized to emerge from a self-supervised learning process that attempts to extract “temporally stable” features from the sensory input. Conversely, invariance increases along the hierarchy, putatively implying progressively slower neural codes in higher-level areas. Recent neurophysiological evidence shows that indeed, as one moves along this cortical hierarchy, neural representations of dynamic stimuli become slower, and additionally the temporal scales of the within-trial fluctuations of these representations (called “intrinsic timescales”) increase starkly. However, while these timescale hierarchies have been reproduced in biologically grounded recurrent models, their network determinants have remained largely unexplored in image-computable models of the ventral stream. Here we investigate the temporal structure of the neural codes in a noisy, recurrent and adaptive model of the ventral visual stream. We show that, surprisingly, the organization of the representation timescales is set by the broad architectural features of the network, regardless of training, while the ordering of the intrinsic timescales across layers is sensitive to the details of the functions implemented by each layer. Our work underscores the importance of the temporal structure of the neural code as a probe for the link between structure and function in models of the vertebrate visual system.
-
PubblicazioneBuilding a FAIR-aligned data-processing pipeline for investigating research artifact availability in high-energy physics studies( 2026-06-18)In the current thesis project, a FAIR-aligned data-processing pipeline was developed to extract and classify Data Availability Statements (DAS) and Code Availability Statements (CAS) in high-energy physics scienti c publi cations. The pipeline was designed as a reusable tool organized around ve modules handling ingestion of publications, conversion to text format, ex traction, normalization and classi cation of availability statements. In total, 17,185 documents (n = 8,511 for 2024, n = 8,674 for 2025) were processed. Out of these, 17,076 were scienti c studies analyzed to verify the inclusion of availability statements and classify their content. Results showed that 63.5% of the publications analyzed included a DAS section, while only 31.1% in cluded a CAS section. Nonetheless, year-speci c results showed a marked increase in the inclusion of both DAS and CAS sections in 2025 publications, suggesting possible adaptation to newly introduced open science policies. Classi cation results showed that a majority of studies declared no data and no code associated with the manuscript. When looking at studies that clearly reported associated data or code, a majority of studies declared that such artifacts were only available upon request. Despite this, year di erences in category shares showed an increase in studies reporting both data (+15.8 pp) and code (+16.2 pp) as openly accessible in 2025 publications, accom panied by a decrease in studies reporting these artifacts as only available upon request. This might suggest a general increase in open availability of research artifacts in 2025 compared to 2024. Furthermore, data exploration revealed possible ambiguities around what is considered the proper subject of DAS and CAS sections. For code in particular, authors seem to report availability and usage of di erent types of software artifacts in these sections. This led to the development of a taxonomy di erentiating author-produced code from external software tools in order to adequately classify CAS sec tions. Methodological and policy-related implications of these ndings are discussed. The thesis describes the development of a FAIR-aligned data-processing pipeline for verifying the presence and classifying the content of Data Availability Statements and Code Availailibility Statements in high-energy physics publications. Additionally, the thesis reports analytical results for 2024 and 2025 publications funded and openly published by the Sponsoring Consortium for Open Access Publishing in Particle Physics (SCOAP3).
-
PubblicazioneEstimates for modular theory in algebraic QFT(SISSA, 2026-10-16)The general scope of this PhD thesis is to establish relations between Tomita-Takesaki modular theory, in particular modular Hamiltonians and relative modular Hamiltonians, and physically better understood quantities, such as the stress-energy tensor, energy densities or other modular data in cases where it is known explicitly. In Chapter 1, we will present the basics of operators algebras, Tomita-Takesaki theory, symplectic spaces, and link these concepts to discuss quasi-equivalence of Gaussian states. In Chapter 2, we will discuss algebraic QFTs both in the framework of the Haag-Kastler axioms and of locally covariant QFT, to then specialize to scalar fields and present the stress-energy tensor and the Hadamard condition. In Chapter 3, focussing on Gaussian states of the Weyl-CCR algebra, we study relations between quasi-equivalence of their GNS-representations, differences of the total energy (on some Cauchy surface), and differences between functions of the modular Hamiltonians. Following qualitative relations by Araki and Yamagami, Verch and Longo, we investigate quantitative relations, including in particular estimates of differences between functions of modular Hamiltonians in terms of energy differences. These results will make use of inequalities between p-norms of operators, which we will discuss in Appendix C. In Chapter 4, generalizing the picture beyond Gaussian states, we exploit locality properties of general algebraic QFTs to construct a scheme that allows to estimate relative modular Hamiltonians for suitable pairs of states on some local algebra in terms of modular Hamiltonians of a reference state on bigger or smaller algebras, which might be better understood. For pairs of states susceptible to this scheme, there exist unitaries linking them to the reference state which do not enable superluminal signalling in the sense of Sorkin's paradox. Such unitaries may not however compose the whole class of non-signalling operators, raising the question whether the existence of a non-signalling unitary is a sufficient condition to show the applicability of our estimate. To investigate the strength of our estimates we then consider coherent states for CCR systems, focussing particularly on free scalar fields, where the difference between the upper and lower bounds can be made arbitrarily small, recovering an exact result by squeezing. This yields an independent proof for the relative entropy formula in cases where the relative modular Hamiltonian has not been computed exactly and corroborates the strength of our general estimates. For these results, we will make use of several commutation properties of operators on Fock spaces, which we will present in Appendix A. Results about operator monotone functions will be used throughout this thesis and are explained in Appendix B.
-
PubblicazioneEnhancing the Sensorium Dataset for FAIR Neuroscience Research( 2026-06-29)The FAIR principles Findable, Accessible, Interoperable, and Reusable provide a framework for improving the scienti c value of research data. Although adoption is increasing across disciplines, neurophysiology datasets often do not meet these standards, which restricts their reuse beyond the initial study. This thesis uses the Sensorium dataset, released as part of a challenge to develop predictive models of the mouse primary visual cortex, as a case study in dataset FAIRi cation. The Sensorium dataset comprises calcium imaging recordings from ap proximately 8,000 neurons per mouse, collected from ten mice exposed to both natural and parametrically generated video stimuli. When the dataset was accessed for a novel analysis, several essential metadata elements were found to be missing, including stimulus labels, unique video identi ers, segment level descriptions, and data quality ags. The rst part of this thesis ad dresses these gaps by systematically enriching metadata. A custom algo rithm was developed to classify videos based on the labels reported in the associated publication. Duplicate videos within and across recordings were identi ed and assigned unique identi ers; equivalent video segments in para metric stimuli were detected and catalogued; and data quality issues, includ ing missing trials and corrupted videos, were documented. All metadata was organised into a machine-readable, extensible schema using JSON and CSV formats, and released openly with a Python toolkit for data management, ltering, and visualisation, under permissive open-source and CC0 licences. The second part of the thesis demonstrates the scienti c value of this FAIRi cation e ort by applying Topological Data Analysis (TDA), speci cally zigzag persistence, to characterise the dynamics of neural population activity in response to the di erent stimuli. Classi cation analyses show that zigzag persistence vectorisations capture meaningful and generalisable topological features of neural activity. These results were enabled by the structured metadata generated in the initial phase of the thesis. Collectively, this work highlights both the challenges of retroactive FAIR i cation and the scienti c bene ts it allows.
-
PubblicazionePerceptual misalignment of texture representations in convolutional neural networks( 2026)Mathematical modeling of visual textures traces back to Julesz's intuition that texture perception in humans is based on local correlations between image features. An influential approach for texture analysis and generation generalizes this notion to linear correlations between the nonlinear features computed by convolutional neural networks (CNNs), compiled into Gram matrices. Given that CNNs are often used as models for the visual system, it is natural to ask whether such "texture representations" spontaneously align with the textures' perceptual content, and, in particular, whether those CNNs that are regarded as better models for the visual system also possess more human-like texture representations. Here we quantify the perceptual content captured by feature correlations computed for a diverse pool of CNNs, and we compare it to the models' perceptual alignment with the mammalian visual system as measured by Brain-Score. Surprisingly, we find that there is no connection between conventional measures of CNN quality as a model of the visual system and its alignment with human texture perception. We conclude that texture perception involves mechanisms that are distinct from those that are commonly modeled using approaches based on CNNs trained on object recognition, possibly depending on the integration of contextual information.