Establishing the foundations of Meta-learning - a Proposal

Brando Miranda

Permalink

https://hdl.handle.net/2142/109133 Copy

Description

Title

Establishing the foundations of Meta-learning - a Proposal

Author(s)

Brando Miranda

Issue Date

2020-12-23

Keyword(s)

• meta-learning
• agi
• general intelligence
• AI safety
• machine learning
• learning to learn
• ai

Abstract

General Artificial Intelligence (AGI) has the potential to be one of the most transformative technologies we are yet to develop. It can affect us in every way, from doing daily housework to advanced theorem proving. Meta-learning, also known as “learning to learn”, is the subfield of machine learning that studies the design of intelligent agents that learn to adapt to novel situations rapidly. In essence, meta-learning should be driving the progress towards AGI - since the ability to learn general skills rapidly is at the core of general intelligence. Although deep learning has deservedly been at the forefront of the current AI revolution, it's extensions to the field of meta-learning have lagged behind to deliver such AGI technologies. Instead its explosion in popularity has delivered a vast variety of methods of increasing complexity that were recently shown [1, 2, 3] be easily matched by algorithms that arguably exhibit no explicit meta-learning (e.g. pre-training a highly-parameterized multi-layer neural network and then fine tuning the final layer at test time [2]). I believe that to overcome this challenge we have to: 1) design universal metrics and benchmarks that quantify the core goal of meta-learning1 2) develop a scientific framework to understand meta-learning as novel skill acquisition and 3) incorporate AI safety metrics to the global research development cycle of meta-learning (arguably the appropriate and important place to include such a metric in an actionable way as soon as possible).

Type of Resource

text

Permalink

http://hdl.handle.net/2142/109133

Owning Collections