ML would like to specifically thank Tian Xia for many insightful discussions on Riemannian geometry and stochastic analysis on manifolds. We also thank Blair Bilodeau, Philippe Casgrain, Christopher Kennedy, Justin Ko, Yasaman Mahdaviyeh, Jeffrey Ne…

It is a pleasure to thank Daniel Park, Ethan Dyer, Edgar Dobriban, and Pokey Rule for a number of insightful conversations about data augmentation. B.H. was partially supported by NSF Grant DMS-1855684 and ONR MURI “Theoretical Foundations of Deep L…

This work was realized during a six-month internship at Logitech.

This paper addresses the problem of deep learning in a decentralized setting using momentum accelerated consensus. We establish a consensus-based decentralized learning algorithm using the stochastic heavy ball approach that can assist in finding th…

Highly accessible and user-friendly computer tools are difficult to design. Furthermore, due to the wide spectrum of visual disabilities, it is impossible to tailor an ideal program for each user. As a result, designers and engineers must employ the…

Our experimental results yielded the following interesting points of discussion.

In this article, we proposed a novel method to penalize the rotation of the ICP algorithm with Lie algebra. Using preliminary real-world experiments, we showed that our method leads to increased accuracy and stability of the ICP. Future work will fo…

Learning point signature for 3D shapes using a meta-learning based method introduced in this work significantly improves the learned point signature quality over the baseline model. Our two-level networks, i.e., a base-learner and a meta-learner, ca…

In the last couple of years, AI has gained a lot of attention in industry, due to the latest research developments in the field, e.g. deep learning LeCun-Nature-2015. Indeed, an increasing amount of companies have started to integrate neural models …

In this section, we first report our results on LibriSpeech dataset and compare with other streaming ASR networks. We next study the hyper-parameter λ in FastEmit to balance transducer loss and regularization loss. Finally, we conduct large-scale ex…

We introduced M2M-100, a new Many-to-Many multilingual translation model that can translate between the 9,900 directions of 100 languages. The underlying dataset was mined from CommonCrawl using a novel strategy which exploits language groupings to …

In this paper, we investigated the general problem of conditioned dialogue generation. We assume that conditioned dialogue for any condition can be done in a similar way by providing a set of dialogue data and a set of extra texts corresponding to t…

We propose that sentence boundary errors are a neglected area of study for NMT robustness, especially in the context of speech translation. We quantitatively demonstrate that poor sentence segmentation degrades performance almost twice as much as tr…

We present a full pipeline of cluster search using a CNN, which is able to infer cluster’s parameters (age, mass, extinctions, and size), as well as provide inference reliability estimates. The CNN is trained on mock star clusters superimposed on im…

Our investigations have identified several directions in which modern text-to-SQL models that rely on relation-aware transformers can be simplified. In particular, we have shown that a pure transformer decoder with vanilla self- and memory-attention…

This study tested a UAV LiDAR system on urban road infrastructure mapping with a case study on semantic segmentation of a multi-layer stack interchange. An automated end-to-end point cloud semantic segmentation network utilizing stacked convolutions…