Ferran Alet

I am a Research Scientist at Google DeepMind. I recently graduated from MIT CSAIL, advised by Leslie Kaelbling and Tomas Lozano-Perez, and Josh Tenenbaum.

Research: I aim to better understand and improve machine learning generalizability. To accomplish this, I leverage techniques from meta-learning, learning to search, program synthesis, and insights from mathematics and the physical sciences. I enjoy building collaborations to work across the entire theory-application spectrum.

Twitter  /  Email  /  CV  /  Google Scholar  /  LinkedIn

We suggest there is a contradiction in how we model noise in ML and derive a framework for losses in function space.

We propose to encode symmetries as conservation tailoring losses and meta-learn them from raw inputs in sequential prediction problems.

We optimize unsupervised losses for the current input. By optimizing where we act, we bypass generalization gaps and can impose a wide variety of inductive biases.

We generate a large quantity of diverse real programs by running code instruction-by-instruction and obtain I/O pairs for 200k subprograms.

By meta-learning programs instead of neural network weights, we can increase meta-learning generalization. We discover new algorithms in simple environments that generalize to complex ones.

We frame neural relational inference as a case of modular meta-learning and speed up the original modular meta-learning algorithms by two orders of magnitude, making them practical.

Diverse dataset of 250 objects pushed 250 times each, all with RGB-D video. First probabilistic meta-learning benchmark.

We learn to map functions to functions by combining graph networks and attention to build computational meshes and show this new framework can solve very diverse problems.

We propose to do meta-learning by training a set of neural networks to be composable, adapting to new tasks by composing modules in novel ways, similar to how we compose known words to express novel ideas.

People often find entities by clustering; we suggest that, instead, entities can be described as dense regions and propose a very simple algorithm for detecting them, with provable guarantees.

Description of the system for the Amazon Robotics Challenge 2017 competition, in which we won the stowing task.

He is a generous guy with a cool website.