Montreal

This work calls for a change in how we evaluate performance on reinforcement learning benchmarks. We present more reliable protocols, easily applicable with even a handful of runs, to prevent unreliable results from stagnating the field (Outstanding Paper Award at NeurIPS 2021).

We break the standard pipeline of program synthesis by examples into two stages: (a) find programs that satisfy only one IO example (PE solutions), and (b) combine the PE solutions to find a global solution that satisfies all IO examples. For the latter, we propose a neural module that automatically learns to combine the PE solutions using the program execution information.

Investigating the impact of aliasing on generalization in Deep Convolutional Networks and proposing simple architectural changes which can lead to substantial performance gains on both i.i.d. and out-of-distribution generalization.

Can we predict runtime errors in a "static" setting, where program execution is not possible? We introduce a competition dataset and task for predicting runtime errors, including the location of the error, and evaluate with IPA-GNN and generic neural models.

Hardware accelerated (GPU/TPU), batchable and differentiable optimizers in JAX.

A large-scale and diverse benchmark for measuring the competence of different image classification models in a realistic and challenging few-shot setting.

An algorithm for training sparse neural networks that identifies neurons that should be active during training, which helps the optimization process to utilize the most relevant connections and results in better sparse solutions.

Findings on what musician/developer teams needed when creating a song using machine learning, the challenges they faced, and how they leveraged and repurposed existing characteristics of such systems to overcome some of these challenges.

A TensorFlow implementation of the Implicit Gradient Transport optimizer, which reduces the variance in online optimization by transporting past gradients.