The 3rd annual Machine Learning in the Real World Workshop (MLRWW) took place on November 8th, 2017 in Paris, France. The aim of this workshop is bring together Machine Learning researchers and practitioners to discuss scientific questions emerging from solving “real world” problems and gain insights into current and emerging trends. The highlights of the workshop included invited talks by leading experts in the machine learning field, poster session and panel discussions.
A well-received talk by Sathiya Keerthi focussed on the generalization of Deep Neural Networks (DNNs) despite the over parameterization. During his talk, he discussed the effect of the optimization methods used for training DNNs and discussed similar studies to explore the complex relationship between optimization and generalization.
Sathiya Keerthi: Interplay between Optimization and Generalization in Deep Neural Networks
In his talk, Alexandros Karatzoglou explored how Recurrent Neural Networks (RNNs) can be used to model session-based based recommendations in conjunction with personalized recommendations and compressed representations with bloom embeddings.
Alexandros Karatoglou: Recurrent Neural Networks for Session-based Recommendations
Thorsten Joachims discussed how deep learning techniques can be effectively used to evaluate the efficacy of online advertisements (i.e. batch learning from logged bandit feedback). He spoke about the efficacy of how training deep networks from propensity-scored bandit feedback can be used for applications ranging from visual object detection to ad placement.
Thorsten Joachims: Deep Learning from Logged Interventions
Along similar lines, Csaba Szepesvari in his talk, presented the effective use of structure and prior information for scaling up bandits to large scale problems, to deal with delayed and missing feedback.
Csaba Szepesvari: Messy Bandit Problems
To make existing machine learning algorithms amenable to the volume of big data, Chih-Jen Lin presented his work on efficient optimization methods for large-scale linear classification techniques where in, he discussed optimization methods in three environments: single-thread, shared-memory, and distributed.
Chih-Jen Lin: Training large-scale linear classifiers: status and challenges
Besides the invited talks, we had a number of spotlight presentations from up & coming student researchers which helped set the stage for the poster sessions. The links to their presentations are below:
Gaussian Embeddings for Collaborative Filtering – Ludovic De Santos
Quasi-Bayesian Learning – Benjamin Guedj
Comparative Study of Counterfactual Estimators – Thomas Nedelec
Siamese Cookie Embeddings for Cross-Device User Matching – Ugo Tanielian
Stochastic Bandit Model for Delayed Conversions – Claire Vernade
A linear reinforcement learning algorithm for non stationary actions in recommender systems – Romain Warlop
Learning from video and text via large-scale discriminative clustering – Antoine Miech
Overall, the 3rd annual Machine Learning in the Real World workshop was a big success. Our, CTO, Dan Teodosiu, summed it up nicely when he emphasized the importance of continuing collaboration with the ML community and how we at Criteo are committed to the same.
Dan Teodosiu, CTO, Criteo
The next edition of the workshop will be on June 28th, 2018 & we are looking forward to bringing you more of these great talks!