arXiv subjects

Nayara Fonseca

Publications and source records attributed to Nayara Fonseca.

17 recordsLinked to original sources

arXiv2026

Decoupling Dynamical Richness from Representation Learning: Towards Practical Measurement

Dynamic feature transformation (the rich regime) does not always align with predictive performance (better representation), yet accuracy is often used as a proxy for richness, limiting analysis of their relationship. We propose a computationally efficient, performance-independent metric of richness grounded in the low-rank bias of rich dynamics, which recovers neural collapse as a special case. The metric is empirically more stable than existing alternatives and captures known lazy-torich transitions (e.g., grokking) without relying on accuracy. We further use it to examine how training factors (e.g., learning rate) relate to richness, confirming recognized assumptions and highlighting new observations (e.g., batch normalization promotes rich dynamics). An eigendecomposition-based visualization is also introduced to support interpretability, together providing a diagnostic tool for studying the relationship between training factors, dynamics, and representations.

stat.ML↗

arXiv2024

An exactly solvable model for emergence and scaling laws in the multitask sparse parity problem

Deep learning models can exhibit what appears to be a sudden ability to solve a new problem as training time, training data, or model size increases, a phenomenon known as emergence. In this paper, we present a framework where each new ability (a skill) is represented as a basis function. We solve a simple multi-linear model in this skill-basis, finding analytic expressions for the emergence of new skills, as well as for scaling laws of the loss with training time, data size, model size, and optimal compute. We compare our detailed calculations to direct simulations of a two-layer neural network trained on multitask sparse parity, where the tasks in the dataset are distributed according to a power-law. Our simple model captures, using a single fit parameter, the sigmoidal emergence of multiple new skills as training time, data size or model size increases in the neural network.

cs.LG↗

arXiv2023

Generalizing similarity in noisy setups: the DIBS phenomenon

This work uncovers an interplay among data density, noise, and the generalization ability in similarity learning. We consider Siamese Neural Networks (SNNs), which are the basic form of contrastive learning, and explore two types of noise that can impact SNNs, Pair Label Noise (PLN) and Single Label Noise (SLN). Our investigation reveals that SNNs exhibit double descent behaviour regardless of the training setup and that it is further exacerbated by noise. We demonstrate that the density of data pairs is crucial for generalization. When SNNs are trained on sparse datasets with the same amount of PLN or SLN, they exhibit comparable generalization properties. However, when using dense datasets, PLN cases generalize worse than SLN ones in the overparametrized region, leading to a phenomenon we call Density-Induced Break of Similarity (DIBS). In this regime, PLN similarity violation becomes macroscopical, corrupting the dataset to the point where complete interpolation cannot be achieved, regardless of the number of model parameters. Our analysis also delves into the correspondence between online optimization and offline generalization in similarity learning. The results show that this equivalence fails in the presence of label noise in all the scenarios considered.

cs.LG↗

arXiv2023

Probing optimisation in physics-informed neural networks

A novel comparison is presented of the effect of optimiser choice on the accuracy of physics-informed neural networks (PINNs). To give insight into why some optimisers are better, a new approach is proposed that tracks the training trajectory curvature and can be evaluated on the fly at a low computational cost. The linear advection equation is studied for several advective velocities, and we show that the optimiser choice substantially impacts PINNs model performance and accuracy. Furthermore, using the curvature measure, we found a negative correlation between the convergence error and the curvature in the optimiser local reference frame. It is concluded that, in this case, larger local curvature values result in better solutions. Consequently, optimisation of PINNs is made more difficult as minima are in highly curved regions.

cs.LG↗

arXiv2021

Probing photophobic (rel)axion dark matter

We investigate the interplay between early universe cosmology and dark matter direct detection, considering axion models with naturally suppressed couplings to photons. In the context of the cosmological relaxation of the electroweak scale, we focus on a scenario of \emph{Relaxion Dark Matter}, in which the relaxion field constitutes all the observed dark matter relic density and its allowed mass range is fixed to a few $\mathrm{keV}$ by construction. In particular, we show that a relaxion particle with mass $m_ϕ= 3.0 \,\mathrm{keV}$ which couples to electrons with $g_{ϕ, e}= 6.8 \times 10^{-14}$ is consistent with the XENON1T excess, while accounting for the observed dark matter and satisfying astro/cosmo probes. This scenario uses the electroweak scale as the link connecting the relaxion production at early times with the dark matter absorption rate in direct detection.

hep-ph↗

arXiv2020

Relaxion Fluctuations (Self-stopping Relaxion) and Overview of Relaxion Stopping Mechanisms

In implementations of the electroweak scale cosmological relaxation mechanism proposed so far, the effect of the quantum fluctuations of the homogeneous relaxion field has been ignored. We show that they can grow during the classical cosmological evolution of the relaxion field passing through its many potential barriers. The resulting production of relaxion particles can act as an efficient stopping mechanism for the relaxion. We revisit the original relaxion proposal and determine under which conditions inflation may no longer be needed as a source of friction. We review alternative stopping mechanisms and determine in detail the allowed parameter space for each of them (whether happening before, during and after inflation), also considering and severely constraining the case of friction from electroweak gauge boson production in models with large and Higgs-independent barriers.

hep-ph↗

arXiv2020

Relaxion Dark Matter

We highlight a new connection between the Standard Model hierarchy problem and the dark matter sector. The key piece is the relaxion field, which besides scanning the Higgs mass and setting the electroweak scale, also constitutes the observed dark matter abundance of the universe. The relaxation mechanism is realized during inflation, and the necessary friction is provided by particle production. Using this framework we show that the relaxion is a phenomenologically viable dark matter candidate in the keV mass range.

hep-ph↗

arXiv2020

Axion Fragmentation

We investigate the production of axion quanta during the early universe evolution of an axion-like field rolling down a wiggly potential. We compute the growth of quantum fluctuations and their back-reaction on the homogeneous zero-mode. We evaluate the transfer of kinetic energy from the zero mode to the quantum fluctuations and the conditions to decelerate the axion zero-mode as a function of the Hubble rate, the slope of the potential, the size of the barriers and the initial field velocity. We discuss how these effects impact the relaxion mechanism.

hep-ph↗

arXiv2019

The Super-Planckian Axion Strikes Back

We present a novel framework for obtaining large hierarchies in axion decay constants as well as trans-Planckian field excursions, with no need for tuning or a large number of fields. We consider a model with two or more CFTs with a common cutoff, that are linked by a gauged diagonal symmetry. This construction is dual to the geometry of a warped space with two or more throats glued at a common brane, allowing for calculability. Many applications of our setup are possible,such as ultra-light axions, natural inflation, and relaxion models.

hep-ph↗

arXiv2018

Higgs relaxation after inflation

We show that the mechanism of cosmological relaxation of the electroweak scale can take place independently of the inflation mechanism, thus relieving burdens from the original relaxion proposal. What eventually stops the (fast-rolling) relaxion field during its cosmological evolution is the production of particles whose mass is controlled by the Higgs vacuum expectation value. We first show that Higgs particle production does not work for that purpose as the Higgs field does not track the minimum of its potential in the regime where Higgs particles get efficiently produced through their coupling to the relaxion. We then focus on gauge boson production. We provide a detailed analysis of the scanning and stopping mechanism and determine the parameter space for which the relaxion mechanism can take place after inflation, while being compatible with cosmological constraints, such as the relaxion dark matter overabundance and Big Bang Nucleosynthesis. We find that the cutoff scale can be as high as two hundreds of TeV. In this approach, the relaxion sector is responsible for reheating the visible sector. The stopping barriers of the periodic potential are large and Higgs-independent, facilitating model-building. The allowed relaxion mass ranges from 200 MeV up to the weak scale. In this scenario, the relaxion field excursion is subplanckian, and is thus many orders of magnitude smaller than in the original relaxion proposal.

hep-ph↗

arXiv2018

A warped relaxion

We construct a UV completion of the relaxion in a warped extra dimension. We identify the relaxion with the zero mode of the fifth component of a bulk gauge field and show how hierarchically different decay constants for this field can be achieved by different localizations of anomalous terms in the warped space. This framework may also find applications for other axion-like fields. The cutoff of the relaxion model is identified as the scale of the IR brane where the Higgs lives, which can be as high as $10^6$ GeV, while above this scale warping takes over in protecting the Higgs mass.

hep-ph↗

arXiv2016

N-Relaxion: Large Field Excursions from a Few Site Relaxion Model

Relaxion models are an interesting new avenue to explain the radiative stability of the Standard Model scalar sector. They require very large field excursions, which are difficult to generate in a consistent UV completion and to reconcile with the compact field space of the relaxion. We propose an N-site model which naturally generates the large decay constant needed to address these issues. Our model offers distinct advantages with respect to previous proposals: the construction involves non-abelian fields, allowing for controlled high energy behaviour and more model building possibilities, both in particle physics and inflationary models, and also admits a continuum limit when the number of sites is large, which may be interpreted as a warped extra dimension.

hep-ph↗

arXiv2016

Sharing but not Caring: Dark Matter and the Baryon Asymmetry of the Universe

We consider scenarios where Dark Matter (DM) particles carry baryon and/or lepton numbers, which can be defined if there exist operators connecting the dark to the visible sector. As a result, the DM fields become intimately linked to the Standard Model (SM) ones and can be maximally asymmetric just like the ordinary matter. In particular, we discuss minimal scenarios where the DM is a complex scalar or a Dirac fermion coupled to operators with nonzero baryon and/or lepton numbers, and that consist of only SM fields. We consider an initial asymmetry stored in either the SM or the DM sector; the main role of these operators is to properly $share$ the asymmetry between the two sectors, in accordance with observations. After the chemical decoupling, the DM and SM sectors do $not$ $care$ about each other as there is only an ineffective communication between them. Once the DM mass is specified, the Wilson coefficients of these operators are fixed by the requirement of the correct transfer of the asymmetry. We study the phenomenology of this framework at colliders, direct detection and indirect detection experiments. In particular, the LHC phenomenology is very rich and can be tested in different channels such as the two same-sign leptons with two jets, monojet and monojet with a monolepton.

hep-ph↗

arXiv2016

Dark Matter, Shared Asymmetries, and Galactic Gamma Ray Signals

We introduce a novel dark matter scenario where the visible sector and the dark sector share a common asymmetry. The two sectors are connected through an unstable mediator with baryon number one, allowing the standard model baryon asymmetry to be shared with dark matter via semi-annihilation. The present-day abundance of dark matter is then set by thermal freeze-out of this semi-annihilation process, yielding an asymmetric version of the WIMP miracle as well as promising signals for indirect detection experiments. As a proof of concept, we find a viable region of parameter space consistent with the observed Fermi excess of GeV gamma rays from the galactic center.

hep-ph↗

arXiv2015

Dark Matter Constraints on Composite Higgs Models

In composite Higgs models the pseudo-Nambu-Goldstone Boson (pNGB) nature of the Higgs field is an interesting alternative for explaning the smallness of the electroweak scale with respect to the beyond the Standard Model scale. In non-minimal models additional pNGB states are present and can be a Dark Matter (DM) candidate, if there is an approximate symmetry suppressing their decay. Here we assume that the low energy effective theory (for scales much below the compositeness scale) corresponds to the Standard Model with a pNGB Higgs doublet and a pNGB DM multiplet. We derive general effective DM Lagrangians for several possible DM representations (under the SM gauge group), including the singlet, doublet and triplet cases. Within this framework we discuss how the DM observables (relic abundance, direct and indirect detection) constrain the dimension-6 operators induced by the strong sector assuming that DM behaves as a Weakly Interacting Particle (WIMP) and that the relic abundance is settled through the freeze-out mechanism. We also apply our general results to two specific cosets: $SO(6)/SO(5)$ and $SO(6)/SO(4) \times SO(2)$, which contain a singlet and doublet DM candidate, respectively. In particular we show that if compositeness is a solution to the little hierarchy problem, representations larger than the triplet are strongly disfavored. Furthermore, we find that composite models can have viable DM candidates with much smaller direct detection cross-sections than their non-composite counterparts, making DM detection much more challenging.

hep-ph↗

arXiv2013

Resonances from Quiver Theories at the LHC

We consider the collider signals of spin-one resonances present in full-hierarchy quiver theories of electroweak symmetry breaking. These four-dimensional theories result from the deconstruction of warped extra dimensional models and have very distinct phenomenological features when the number of sites is small. We study a class of generic scenarios in these theories where the color gauge group as well as the electroweak sector, propagate in the quiver diagram. These scenarios correspond to various specific models of electroweak symmetry breaking and fermion masses. We focus on the minimum resonant content and its main features: the presence of heavy and narrow spin one resonances. We derive bounds from the LHC data on the color-octet and color-singlet excited gauge bosons from their decays to jets and top pairs, and show their dependence on the number of sites in the quiver. We also compare them with the bounds derived from flavor violation.

hep-ph↗

arXiv2012

Full-hierarchy Quiver Theories of Electroweak Symmetry Breaking and Fermion Masses

We consider quiver theories in four dimensions with a large ultra-violet cutoff. These theories require that an ordered set of vacuum expectation values for the link fields develops dynamically and can be obtained from the coarse deconstruction of extra-dimensional theories in an AdS background. These full-hierarchy quiver theories form a large class which include AdS$_5$ models as a limit, but which have a distinctive phenomenology. As an example, in this paper we show that fermions can be introduced in a way that can at the same time generate the fermion mass hierarchy and have flavor violation consistent with experimental bounds, when the mass scale of the color-octect gauge excitation is above $3 {\rm TeV}$. We also show that electroweak precision constraints are satisfied by this mass scale, without the need to extend the gauge sector to protect against custodial violation.

hep-ph↗