arXiv ScienceSearch

arXiv subjects

Shiqi Su

Publications and source records attributed to Shiqi Su.

2 recordsLinked to original sources

Generalised Balanced Softmax: A Finite-Data Perspective on Logit Adjustment for Long-Tailed Recognition

Models trained on long-tailed data using standard softmax tend to exhibit higher training error and a larger generalisation gap for classes with fewer training samples. We characterise this class-wise disparity as the preference issue and quantify it using a new metric, the model imbalance level $I$. To understand this issue, we analyse how imbalanced training data adversely affects class-wise gradients under standard softmax training. This paper then develops a finite-data Generalised Balanced Softmax (GBS) framework for analysing and mitigating the preference issue. The framework uses the training-time logit adjustment $z_{nc}+β\log|N_c|$, which is algebraically identical to the training-time logit-adjusted loss of Menon et al. (2021) when $τ=β$. The case $β=1$ also coincides with Balanced Softmax and with the unit adjustment supported by the Fisher-consistency argument under the true data distribution, corresponding to an idealised infinite-data setting. Building on this existing loss family, this paper uses a heuristic power-law assumption to motivate the adjustable coefficient and studies how $β$ affects trained models. Across the evaluated long-tailed benchmarks, $β=1$ does not attain the highest average testing recall on most datasets, showing that a different coefficient can be preferable when training on finite data. The selected values of $β$ reduce $I$ and improve average testing recall relative to the $β=1$ reference, while retaining negligible computational overhead and compatibility with existing representation-learning frameworks.

cs.LG

Emulating CO Line Radiative Transfer with Deep Learning

Modelling carbon monoxide (CO) line radiation is computationally expensive for traditional numerical solvers, especially when applied to complex, three-dimensional stellar atmospheres. We present COEmuNet, a 3D convolutional neural network (CNN)-based surrogate model that emulates CO line radiation transport with high accuracy and efficiency. It consists of an asymmetric encoder-decoder design that takes 3D hydrodynamical models as inputs and generates synthetic observations of evolved stellar atmospheres. The model is trained on data from hydrodynamic simulations of Asymptotic Giant Branch (AGB) stars perturbed by a companion. Given a set of input parameters, including velocity fields, kinetic temperature distribution, and CO molecular number densities, the COEmuNet model emulates spectral line observations with a median relative error of ~7% compared to a classical numerical solver of the radiative transfer equation, measured over seven frequency channels and arbitrary viewing directions. Besides, COEmuNet delivers a 1000 times speedup, enabling efficient model fitting to observational datasets, real-time visualization of simulations and progress toward integration in large-scale cosmological simulations.

astro-ph.IM