We focus on multi-domain Neural Machine Translation, with the goal of developing efficient models which can handle data from various domains seen during training and are robust to domains unseen during training. We hypothesize that Sparse Mixture-of-Experts (SMoE) models are a good fit for this task, as they enable efficient model scaling, which helps to accommodate a variety of multi-domain data, and allow flexible sharing of parameters between domains, potentially enabling knowledge transfer between similar domains and limiting negative transfer. We conduct a series of experiments aimed at validating the utility of SMoE for the multi-domain scenario, and find that a straightforward width scaling of Transformer is a simpler and surprisingly more efficient approach in practice, and reaches the same performance level as SMoE. We also search for a better recipe for robustness of multi-domain systems, highlighting the importance of mixing-in a generic domain, i.e. Paracrawl, and introducing a simple technique, domain randomization.

Epilepsy is one of the most common neurological disorders, and many patients require surgical intervention when medication fails to control seizures. For effective surgical outcomes, precise localisation of the epileptogenic focus - often approximated through the Seizure Onset Zone (SOZ) - is critical yet remains a challenge. Active probing through electrical stimulation is already standard clinical practice for identifying epileptogenic areas. This paper advances the application of deep learning for SOZ localisation using Single Pulse Electrical Stimulation (SPES) responses. We achieve this by introducing Transformer models that incorporate cross-channel attention. We evaluate these models on held-out patient test sets to assess their generalisability to unseen patients and electrode placements. Our study makes three key contributions: Firstly, we implement an existing deep learning model to compare two SPES analysis paradigms - namely, divergent and convergent. These paradigms evaluate outward and inward effective connections, respectively. Our findings reveal a notable improvement in moving from a divergent (AUROC: 0.574) to a convergent approach (AUROC: 0.666), marking the first application of the latter in this context. Secondly, we demonstrate the efficacy of the Transformer models in handling heterogeneous electrode placements, increasing the AUROC to 0.730. Lastly, by incorporating inter-trial variability, we further refine the Transformer models, with an AUROC of 0.745, yielding more consistent predictions across patients. These advancements provide a deeper insight into SOZ localisation and represent a significant step in modelling patient-specific intracranial EEG electrode placements in SPES. Future work will explore integrating these models into clinical decision-making processes to bridge the gap between deep learning research and practical healthcare applications.

Non-autoregressive approaches aim to improve the inference speed of translation models by only requiring a single forward pass to generate the output sequence instead of iteratively producing each predicted token. Consequently, their translation quality still tends to be inferior to their autoregressive counterparts due to several issues involving output token interdependence. In this work, we take a step back and revisit several techniques that have been proposed for improving non-autoregressive translation models and compare their combined translation quality and speed implications under third-party testing environments. We provide novel insights for establishing strong baselines using length prediction or CTC-based architecture variants and contribute standardized BLEU, chrF++, and TER scores using sacreBLEU on four translation tasks, which crucially have been missing as inconsistencies in the use of tokenized BLEU lead to deviations of up to 1.7 BLEU points. Our open-sourced code is integrated into fairseq for reproducibility.

Adversarial training has been shown effective at endowing the learned representations with stronger generalization ability. However, it typically requires expensive computation to determine the direction of the injected perturbations. In this paper, we introduce a set of simple yet effective data augmentation strategies dubbed cutoff, where part of the information within an input sentence is erased to yield its restricted views (during the fine-tuning stage). Notably, this process relies merely on stochastic sampling and thus adds little computational overhead. A Jensen-Shannon Divergence consistency loss is further utilized to incorporate these augmented samples into the training objective in a principled manner. To verify the effectiveness of the proposed strategies, we apply cutoff to both natural language understanding and generation problems. On the GLUE benchmark, it is demonstrated that cutoff, in spite of its simplicity, performs on par or better than several competitive adversarial-based approaches. We further extend cutoff to machine translation and observe significant gains in BLEU scores (based upon the Transformer Base model). Moreover, cutoff consistently outperforms adversarial training and achieves state-of-the-art results on the IWSLT2014 German-English dataset.

Neural machine translation (NMT) models are typically trained using a softmax cross-entropy loss where the softmax distribution is compared against smoothed gold labels. In low-resource scenarios, NMT models tend to over-fit because the softmax distribution quickly approaches the gold label distribution. To address this issue, we propose to divide the logits by a temperature coefficient, prior to applying softmax, during training. In our experiments on 11 language pairs in the Asian Language Treebank dataset and the WMT 2019 English-to-German translation task, we observed significant improvements in translation quality by up to 3.9 BLEU points. Furthermore, softmax tempering makes the greedy search to be as good as beam search decoding in terms of translation quality, enabling 1.5 to 3.5 times speed-up. We also study the impact of softmax tempering on multilingual NMT and recurrently stacked NMT, both of which aim to reduce the NMT model size by parameter sharing thereby verifying the utility of temperature in developing compact NMT models. Finally, an analysis of softmax entropies and gradients reveal the impact of our method on the internal behavior of NMT models.

We describe a method for training accurate Transformer machine-translation models to run inference using 8-bit integer (INT8) hardware matrix multipliers, as opposed to the more costly single-precision floating-point (FP32) hardware. Unlike previous work, which converted only 85 Transformer matrix multiplications to INT8, leaving 48 out of 133 of them in FP32 because of unacceptable accuracy loss, we convert them all to INT8 without compromising accuracy. Tested on the new-stest2014 English-to-German translation task, our INT8 Transformer Base and Transformer Big models yield BLEU scores that are 99.3% to 100% relative to those of the corresponding FP32 models. Our approach converts all matrix-multiplication tensors from an existing FP32 model into INT8 tensors by automatically making range-precision tradeoffs during training. To demonstrate the robustness of this approach, we also include results from INT6 Transformer models. 1 Introduction We report a method for training accurate yet compact Transformer machine-translation models [ V aswaniet al., 2017 ] . Specifically, we aim these models at hardware with 8-bit integer (INT8) matrix multipliers. Compared to single-precision floating-point (FP32) matrix multiplications, INT8 matrix multiplications not only reduce both storage and bandwidth four times, but they also consume 15 times less energy [ Horowitz, 2014 ] .