3IA PhD/Postdoc Seminar #2

Presentation of the Education and Training program

Using Denoising Autoencoder for Cayley table completion task
Abstract: The talk will go through neural networks applied to the following task: given a Cayley table of a semigroup from which several cells are missing, how one reconstructs the full table? We will discuss where to get training data, what loss function and network architecture to choose, and how the process in abstract algebra is dissimilar to what usually happens in the fields heavily dominated by deep learning.

A Deep Latent Recommender System based on User Ratings and Reviews
Abstract: We introduce a deep latent recommender system named deepLTRS in order to provide users with high quality recommendations based on observed user ratings and texts of product reviews. Our approach adopts a variational auto-encoder (VAE) architecture as a generative deep latent variable model for both the ordinal matrix, encoding users scores about products, and the document-term matrix, encoding the reviews. Numerical experiments on simulated and real-world data sets demonstrate that deepLTRS outperforms the state-of-the-art, in particular in context of extreme data sparsity.

Fibre-optics, earthquakes, and a zebra: intelligent signal denoising with Deep Learning
Abstract: Distributed Acoustic Sensing (DAS) is an emerging technology that uses fibre-optic cables (the ones that deliver our everyday internet) to measure vibrations in the subsurface. These vibrations can originate from natural phenomena like earthquakes and ocean waves, or from human sources such as boats, cars, trains, and even pedestrians. With DAS, these vibration measurements can be made every 10 metres along the cable, up to distances of 50 kilometres (which would yield 5,000 measurement locations sampled every 1 millisecond!). This unprecedented sampling density offers unique opportunities to study traffic patterns, marine mammals, landslides, earthquakes and much more.
As with any measurement, DAS recordings also contain nuisance signals (noise) that obscure our signals of interest. In this talk we will detail a Deep Learning technique that leverages the spatial sampling density of DAS to separate interesting and coherent signals from incoherent noise. This method is entirely self-supervised, so no labels or “ground truth” is required, and it does not assume any statistical properties of the noise, other than that it is incoherent. We apply our model to earthquakes recorded by a submarine fibre-optic cable, but the applications of the method are completely general (and so extend to the example applications given above).

Additional questions and discussions