Andre Martins

Nuno Gonçalves, Hugo Pitorro, Vlad Niculae et al.

Sparse attention has been proposed as a way to alleviate the quadratic cost of transformers, a central bottleneck in long-context training. A promising line of work is $α$-entmax attention, a differentiable sparse alternative to softmax that enables input-dependent sparsity yet has lagged behind softmax due to the computational overhead necessary to compute the normalizer $τ$. In this paper, we introduce AdaSplash-2, which addresses this limitation through a novel histogram-based initialization that reduces the number of iterations needed to compute $τ$ to typically 1--2. The key idea is to compute a coarse histogram of attention scores on the fly and store it in on-chip SRAM, yielding a more accurate initialization that enables fast forward and backward computation. Combined with a sparsity-aware GPU implementation that skips zero blocks with low overhead, AdaSplash-2 matches or improves per-step training time relative to FlashAttention-2 when block sparsity is moderate-to-high (e.g., $>$60\%), which often occurs at long-context lengths. On downstream tasks, models trained with our efficient $α$-entmax attention match softmax baselines at short-context lengths and achieve substantial gains in long-context settings.

Dennis Fucci, Marco Gaido, Matteo Negri et al.

Recent studies on interpreting the hidden states of speech models have shown their ability to capture speaker-specific features, including gender. Does this finding also hold for speech translation (ST) models? If so, what are the implications for the speaker's gender assignment in translation? We address these questions from an interpretability perspective, using probing methods to assess gender encoding across diverse ST models. Results on three language directions (English-French/Italian/Spanish) indicate that while traditional encoder-decoder models capture gender information, newer architectures -- integrating a speech encoder with a machine translation system via adapters -- do not. We also demonstrate that low gender encoding capabilities result in systems' tendency toward a masculine default, a translation bias that is more pronounced in newer architectures.

Sophia Sklaviadis, Andre Martins, Mario Figueiredo

From a variational perspective, many statistical learning criteria involve seeking a distribution that balances empirical risk and regularization. In this paper, we broaden this perspective by introducing a new general class of variational methods based on Fenchel-Young (FY) losses, treated as divergences that generalize (and encompass) the familiar Kullback-Leibler divergence at the core of classical variational learning. Our proposed formulation -- FY variational learning -- includes as key ingredients new notions of FY free energy, FY evidence, FY evidence lower bound, and FY posterior. We derive alternating minimization and gradient backpropagation algorithms to compute (or lower bound) the FY evidence, which enables learning a wider class of models than previous variational formulations. This leads to generalized FY variants of classical algorithms, such as an FY expectation-maximization (FYEM) algorithm, and latent-variable models, such as an FY variational autoencoder (FYVAE). Our new methods are shown to be empirically competitive, often outperforming their classical counterparts, and most importantly, to have qualitatively novel features. For example, FYEM has an adaptively sparse E-step, while the FYVAE can support models with sparse observations and sparse posteriors.