Fine-tuning of pre-trained transformer models has become the standard approach for solving common NLP tasks (Devlin et al., 2019). Most of the existing approaches rely on a randomly initialized classifier on top of such networks. We argue that this fine-tuning procedure is sub-optimal as the pre-trained model has no prior on the specific classifier labels, while it might have already learned an intrinsic textual representation of the task. In this paper, we introduce a new scoring method that casts a plausibility ranking task in a full-text format and leverages the masked language modeling head tuned during the pre-training phase. We study commonsense reasoning tasks where the model must rank a set of hypotheses given a premise, focusing on the COPA (Gordon et al., 2012), Swag (Zellers et al., 2018), HellaSwag (Zellers et al., 2019) and CommonsenseQA (Talmor et al., 2019) datasets. By exploiting our scoring method without fine-tuning, we are able to produce strong baselines (e.g. 80% test accuracy on COPA) that are comparable to supervised approaches. Moreover, when fine-tuning directly on the proposed scoring function, we show that our method provides a much more stable training phase across random restarts (e.g ×10 standard deviation reduction on COPA test accuracy) and requires less annotated data than the standard classifier approach to reach equivalent performances.

Abstract-Robust wide baseline pose estimation is an essential step in the deployment of smart camera networks. In this work, we highlight some current limitations of conventional strategies for relative pose estimation in difficult urban scenes. Then we propose a solution which relies on an adaptive search of corresponding interest points in synchronized video streams which allows us to converge robustly towards a high-quality solution. The experiments are performed using a manually annotated ground truth of a large scale scene exhibiting significant depth and perspective variation, uniform areas, repetitive patterns and homogeneous dynamic elements. The results show a fast and robust convergence of the solution, and a significant improvement, compared to single image based alternatives, of the RMSE of ground truth matches, and of the maximum absolute error.