Escape Velocity
Recursive self improvement may be nearing its escape velocity: the point where it sustains itself without constant human intervention.
Louis Kirsch
My mission is to automate AI research to generate superintelligence (ASI), for the benefit of humanity and the proliferation of intelligence. Currently, I am a Research Scientist at Google DeepMind.
Latest posts
Towards General-Purpose In-Context Learning Agents
We meta-train in-context learning RL agents that generalize across domains (with different actuators, observations, dynamics, and dimensionalities) using supervised learning.
General-Purpose In-Context Learning
Transformers and other black-box models can exhibit in-context learning-to-learn that generalizes to significantly different datasets while undergoing multiple phase transitions in terms of their learning behavior.
Self-Referential Meta Learning
We investigate self-referential meta learning systems that modify themselves without the need for explicit meta optimization.
Introducing Symmetries to Black Box Meta Reinforcement Learning
We add symmetries (permutation invariance) to black-box meta reinforcement learners to increase their generalization capabilities.
Meta learning gradient-free learning algorithms
This year's NeurIPS 2021 I present one full paper and a workshop paper on meta learning gradient-free & general-purpose learning algorithms. Can the backpropagation algorithm be encoded purely in the recurrent dynamics of RNNs? How do we automatically discover novel general-purpose learning algorithms that do not need gradient descent? How can symmetries help generalization of reinforcement…
General Meta Learning and Variable Sharing
Humans develop learning algorithms that are incredibly general and can be applied across a wide range of tasks. Unfortunately, this process is often tedious trial and error with numerous possibilities for suboptimal choices. General Meta Learning seeks to automate many of these choices, generating new learning algorithms automatically. Different from contemporary Meta Learning, where the…
MetaGenRL: Improving Generalization in Meta Reinforcement Learning
Biological evolution has distilled the experiences of many learners into the general learning algorithms of humans. Inspired by this process, MetaGenRL distills the experiences of many complex agents to meta-learn a low-complexity neural objective function that affects how future individuals will learn. Unlike recent meta-RL algorithms, MetaGenRL can generalize to new environments that are…
NeurIPS 2018, Updates on the AI road map
I present an updated roadmap to AGI with four critical challenges: Continual Learning, Meta-Learning, Environments, and Scalability. I motivate the respective areas and discuss how research from NeurIPS 2018 has advanced them and where we need to go next.
Is AI progress a function of knowledge or computational resources?
TODO