In this report we review memory-based meta-learning as a tool for building
sample-efficient strategies that learn from past experience to adapt to any
task within a target class. Our goal is to equip the reader with the conceptual
foundations of this tool for building new, scalable agents that operate on
broad domains. To do so, we present basic algorithmic templates for building
near-optimal predictors and reinforcement learners which behave as if they had
a probabilistic model that allowed them to efficiently exploit task structure.
Furthermore, we recast memory-based meta-learning within a Bayesian framework,
showing that the meta-learned strategies are near-optimal because they amortize
Bayes-filtered data, where the adaptation is implemented in the memory dynamics
as a state-machine of sufficient statistics. Essentially, memory-based
meta-learning translates the hard problem of probabilistic sequential inference
into a regression problem.
Description
[1905.03030] Meta-learning of Sequential Strategies
%0 Journal Article
%1 ortega2019metalearning
%A Ortega, Pedro A.
%A Wang, Jane X.
%A Rowland, Mark
%A Genewein, Tim
%A Kurth-Nelson, Zeb
%A Pascanu, Razvan
%A Heess, Nicolas
%A Veness, Joel
%A Pritzel, Alex
%A Sprechmann, Pablo
%A Jayakumar, Siddhant M.
%A McGrath, Tom
%A Miller, Kevin
%A Azar, Mohammad
%A Osband, Ian
%A Rabinowitz, Neil
%A György, András
%A Chiappa, Silvia
%A Osindero, Simon
%A Teh, Yee Whye
%A van Hasselt, Hado
%A de Freitas, Nando
%A Botvinick, Matthew
%A Legg, Shane
%D 2019
%K bayesian meta-learning optimization probability stats
%T Meta-learning of Sequential Strategies
%U http://arxiv.org/abs/1905.03030
%X In this report we review memory-based meta-learning as a tool for building
sample-efficient strategies that learn from past experience to adapt to any
task within a target class. Our goal is to equip the reader with the conceptual
foundations of this tool for building new, scalable agents that operate on
broad domains. To do so, we present basic algorithmic templates for building
near-optimal predictors and reinforcement learners which behave as if they had
a probabilistic model that allowed them to efficiently exploit task structure.
Furthermore, we recast memory-based meta-learning within a Bayesian framework,
showing that the meta-learned strategies are near-optimal because they amortize
Bayes-filtered data, where the adaptation is implemented in the memory dynamics
as a state-machine of sufficient statistics. Essentially, memory-based
meta-learning translates the hard problem of probabilistic sequential inference
into a regression problem.
@article{ortega2019metalearning,
abstract = {In this report we review memory-based meta-learning as a tool for building
sample-efficient strategies that learn from past experience to adapt to any
task within a target class. Our goal is to equip the reader with the conceptual
foundations of this tool for building new, scalable agents that operate on
broad domains. To do so, we present basic algorithmic templates for building
near-optimal predictors and reinforcement learners which behave as if they had
a probabilistic model that allowed them to efficiently exploit task structure.
Furthermore, we recast memory-based meta-learning within a Bayesian framework,
showing that the meta-learned strategies are near-optimal because they amortize
Bayes-filtered data, where the adaptation is implemented in the memory dynamics
as a state-machine of sufficient statistics. Essentially, memory-based
meta-learning translates the hard problem of probabilistic sequential inference
into a regression problem.},
added-at = {2019-05-09T13:27:02.000+0200},
author = {Ortega, Pedro A. and Wang, Jane X. and Rowland, Mark and Genewein, Tim and Kurth-Nelson, Zeb and Pascanu, Razvan and Heess, Nicolas and Veness, Joel and Pritzel, Alex and Sprechmann, Pablo and Jayakumar, Siddhant M. and McGrath, Tom and Miller, Kevin and Azar, Mohammad and Osband, Ian and Rabinowitz, Neil and György, András and Chiappa, Silvia and Osindero, Simon and Teh, Yee Whye and van Hasselt, Hado and de Freitas, Nando and Botvinick, Matthew and Legg, Shane},
biburl = {https://www.bibsonomy.org/bibtex/23d183aa1eaac5446a63037f917d4d1cc/kirk86},
description = {[1905.03030] Meta-learning of Sequential Strategies},
interhash = {d722300f00de73c460f54797b8531d3f},
intrahash = {3d183aa1eaac5446a63037f917d4d1cc},
keywords = {bayesian meta-learning optimization probability stats},
note = {cite arxiv:1905.03030Comment: DeepMind Technical Report (15 pages, 6 figures)},
timestamp = {2019-05-09T13:27:02.000+0200},
title = {Meta-learning of Sequential Strategies},
url = {http://arxiv.org/abs/1905.03030},
year = 2019
}