Ferran Alet

I am a PhD student at MIT CSAIL, where I work on machine learning and robotics with Leslie Kaelbling and Tomas Lozano-Perez. I am also the organizer of the MIT Embodied Intelligence Seminar. Currently mentoring Giray Kuru, Max Thomsen (with Maria Bauza), Katie Wu (with Yilun Du), Javier Lopez-Contreras, Jan Olivetti, and Edgar Gonzalez.

Past mentees: Martin Schneider(now PhD at MIT), Erica Weng(now PhD at CMU), Patrick John Chia , ShengTong Zhang, Margaret Wang, Scott Perry, Catherine Zeng, Adarsh Keshav Jeewajee and Paolo Gentili.

Twitter  /  Email  /  CV  /  Google Scholar  /  LinkedIn

profile photo

I want intelligent agents that can solve many tasks by generalizing very broadly from past experience or requiring just a small amount of data to solve a new task. My research brings together two promising frameworks to think about this problem: meta-learning and combinatorial generalization. Meta-learning allows agents to reuse acquired knowledge from previous tasks to learn a new task more efficiently. Combinatorial generalization creates new systems that generalize more broadly by combining known concepts in novel ways. This approach resembles our linguistic ability to express infinitely many ideas by combining a finite amount of English words. In practice, I build modular architectures that look like computer programs at the top level and neural networks at the bottom level, and algorithms that do meta-learning at both levels. I envision that the combination of meta-learning and modularity can be a scalable path forward to move from simple architectures solving simple problems to more complex architectures that solve harder tasks.

Invited talks

Representative papers are highlighted.

Tailoring: encoding inductive biases by optimizing unsupervised objectives at prediction time
Ferran Alet , Kenji Kawaguchi, Maria Bauza, Nurullah Giray Kuru, Tomás Lozano-Pérez, Leslie Pack Kaelbling
pre-print, 2020

During prediction we can optimize any loss that depends only on the input, allowing us to encode inductive biases in the form of losses and train models to do well after fine-tuning on the unsupervised loss. We show its breadth of applicability on contrastive learning, physics and adversarial examples.

Meta-learning curiosity algorithms
Ferran Alet* , Martin Schneider*, Tomás Lozano-Pérez, Leslie Pack Kaelbling
ICLR, 2020
NeurIPS meta-learning and reinforcement learning workshops, 2019
code / report explaining discovered algorithms / video

By meta-learning program-like structures instead of neural network weights, we can increase meta-learning generalization. We discover new algorithms in simple environments that generalize to completely new complex domains.

Neural Relational Inference with Fast Modular Meta-learning
Ferran Alet , Erica Weng, Tomás Lozano-Pérez, Leslie Pack Kaelbling
NeurIPS, 2019  

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.

Omnipush: accurate, diverse, real-world dataset of pushing dynamics with RGB-D video
Maria Bauza, Ferran Alet Yen-Chen Lin, Tomás Lozano-Pérez, Leslie Pack Kaelbling, Phillip Isola, Alberto Rodriguez
IROS, 2019
project website / code / data / press

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

Graph Element Networks: adaptive, structured computation and memory
Ferran Alet , Adarsh K. Jeewajee, Maria Bauza, Alberto Rodriguez, Tomás Lozano-Pérez, Leslie Pack Kaelbling
ICML, 2019   (Long talk)
talk/ code

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.

Modular meta-learning
Ferran Alet , Tomás Lozano-Pérez, Leslie Pack Kaelbling
CoRL, 2018  
video/ code

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.

Finding Frequent Entities in Continuous Data
Ferran Alet , Rohan Chitnis, Tomás Lozano-Pérez, Leslie Pack Kaelbling
IJCAI, 2018  

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.

Robotic Pick-and-Place of Novel Objects in Clutter with Multi-Affordance Grasping and Cross-Domain Image Matching
Andy Zeng et al.
IROS, 2018  
talk/ project website

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

I borrowed the template from this guy.