Arnout Devos

Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang et al. · eth-zurich

We present Apertus, a fully open suite of large language models (LLMs) designed to address two systemic shortcomings in today's open model ecosystem: data compliance and multilingual representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of memorization, we adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. The Apertus models also expand multilingual coverage, training on 15T tokens from over 1800 languages, with ~40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivalling or surpassing open-weight counterparts. Beyond model weights, we release all scientific artifacts from our development cycle with a permissive license, including data preparation scripts, checkpoints, evaluation suites, and training code, enabling transparent audit and extension.

Arnout Devos, Yatin Dandi

In this paper, we propose a learning algorithm that enables a model to quickly exploit commonalities among related tasks from an unseen task distribution, before quickly adapting to specific tasks from that same distribution. We investigate how learning with different task distributions can first improve adaptability by meta-finetuning on related tasks before improving goal task generalization with finetuning. Synthetic regression experiments validate the intuition that learning to meta-learn improves adaptability and consecutively generalization. Experiments on more complex image classification, continual regression, and reinforcement learning tasks demonstrate that learning to meta-learn generally improves task-specific adaptation. The methodology, setup, and hypotheses in this proposal were positively evaluated by peer review before conclusive experiments were carried out.

Carlos Medina, Arnout Devos, Matthias Grossglauser

Most approaches in few-shot learning rely on costly annotated data related to the goal task domain during (pre-)training. Recently, unsupervised meta-learning methods have exchanged the annotation requirement for a reduction in few-shot classification performance. Simultaneously, in settings with realistic domain shift, common transfer learning has been shown to outperform supervised meta-learning. Building on these insights and on advances in self-supervised learning, we propose a transfer learning approach which constructs a metric embedding that clusters unlabeled prototypical samples and their augmentations closely together. This pre-trained embedding is a starting point for few-shot classification by summarizing class clusters and fine-tuning. We demonstrate that our self-supervised prototypical transfer learning approach ProtoTransfer outperforms state-of-the-art unsupervised meta-learning methods on few-shot tasks from the mini-ImageNet dataset. In few-shot experiments with domain shift, our approach even has comparable performance to supervised methods, but requires orders of magnitude fewer labels.

Anca-Nicoleta Ciubotaru, Arnout Devos, Behzad Bozorgtabar et al.

Most of the existing deep neural nets on automatic facial expression recognition focus on a set of predefined emotion classes, where the amount of training data has the biggest impact on performance. However, in the standard setting over-parameterised neural networks are not amenable for learning from few samples as they can quickly over-fit. In addition, these approaches do not have such a strong generalisation ability to identify a new category, where the data of each category is too limited and significant variations exist in the expression within the same semantic category. We embrace these challenges and formulate the problem as a low-shot learning, where once the base classifier is deployed, it must rapidly adapt to recognise novel classes using a few samples. In this paper, we revisit and compare existing few-shot learning methods for the low-shot facial expression recognition in terms of their generalisation ability via episode-training. In particular, we extend our analysis on the cross-domain generalisation, where training and test tasks are not drawn from the same distribution. We demonstrate the efficacy of low-shot learning methods through extensive experiments.

Arnout Devos, Matthias Grossglauser

We propose regression networks for the problem of few-shot classification, where a classifier must generalize to new classes not seen in the training set, given only a small number of examples of each class. In high dimensional embedding spaces the direction of data generally contains richer information than magnitude. Next to this, state-of-the-art few-shot metric methods that compare distances with aggregated class representations, have shown superior performance. Combining these two insights, we propose to meta-learn classification of embedded points by regressing the closest approximation in every class subspace while using the regression error as a distance metric. Similarly to recent approaches for few-shot learning, regression networks reflect a simple inductive bias that is beneficial in this limited-data regime and they achieve excellent results, especially when more aggregate class representations can be formed with multiple shots.