Uploaded June 2022 | Updated September 2026, 16 hours ago
A Flexible framework for Machine Learning
Ferran Alet
In this last decade, we have seen a lot of progress in AI and Machine Learning using different variations on a single recipe: we specify a task as learning a function mapping inputs to outputs and we train a single neural network to approximate it. In this talk, I will show that this one NN per task framework can be extended in order to improve generalization. First, I will describe modular meta-learning: a method of obtaining language-like generalization by training a set of composable neural modules. By having multiple neural networks per task, and multiple tasks per neural network, we are able to reuse information and achieve bigger data and computational efficiency. In the second part of my talk, I will describe tailoring, a very general way of encoding inductive biases in neural networks by optimizing unsupervised objectives inside the prediction function, essentially having one neural network per input. Finally, I will describe my vision for more flexible, generalizable models that can learn to interpret and reason about new scenes in data-efficient ways.
A Flexible framework for Machine Learning
Ferran Alet
In this last decade, we have seen a lot of progress in AI and Machine Learning using different variations on a single recipe: we specify a task as learning a function mapping inputs to outputs and we train a single neural network to approximate it. In this talk, I will show that this one NN per task framework can be extended in order to improve generalization. First, I will describe modular meta-learning: a method of obtaining language-like generalization by training a set of composable neural modules. By having multiple neural networks per task, and multiple tasks per neural network, we are able to reuse information and achieve bigger data and computational efficiency. In the second part of my talk, I will describe tailoring, a very general way of encoding inductive biases in neural networks by optimizing unsupervised objectives inside the prediction function, essentially having one neural network per input. Finally, I will describe my vision for more flexible, generalizable models that can learn to interpret and reason about new scenes in data-efficient ways.
