Meta-learning (computer science): Difference between revisions

Revision as of 21:38, 5 December 2019

Revision as of 18:37, 11 December 2019

Line 8:

| title = Shift of bias for inductive concept learning

| journal = In R. Michalski, J. Carbonell, & T. Mitchell: Machine Learning

| pages = 163-190163–190

| year = 1986

}}</ref> This means that it will only learn well if the bias matches the learning problem. A learning algorithm may perform very well in one ___domain, but not on the next. This poses strong restrictions on the use of [[machine learning]] or [[data mining]] techniques, since the relationship between the learning problem (often some kind of [[database]]) and the effectiveness of different learning algorithms is not yet understood.

Line 68:

Some approaches which have been viewed as instances of meta learning:

* [[Recurrent neural networks]] (RNNs) are universal computers. In 1993, [[Jürgen Schmidhuber]] showed how "self-referential" RNNs can in principle learn by [[backpropagation]] to run their own weight change algorithm, which may be quite different from backpropagation.<ref name="sch1993">{{cite journal | last1 = Schmidhuber | first1 = Jürgen | year = 1993| title = A self-referential weight matrix | url= | journal = Proceedings of ICANN'93, Amsterdam | pages = 446-451446–451}}</ref> In 2001, [[Sepp Hochreiter]] & A.S. Younger & P.R. Conwell built a successful supervised meta learner based on [[Long short-term memory]] RNNs. It learned through backpropagation a learning algorithm for quadratic functions that is much faster than backpropagation.<ref name="hoch2001">{{cite journal | last1 = Hochreiter | first1 = Sepp | last2 = Younger | first2 = A. S. | last3 = Conwell | first3 = P. R. | year = 2001| title = Learning to Learn Using Gradient Descent | url= | journal = Proceedings of ICANN'01| pages = 87-9487–94}}</ref><ref name="scholarpedia" /> Researchers at [[Deepmind]] (Marcin Andrychowicz et al.) extended this approach to optimization in 2017.<ref name="marcin2017">{{cite journal | last1 = Andrychowicz | first1 = Marcin | last2 = Denil | first2 = Misha | last3 = Gomez | first3 = Sergio | last4 = Hoffmann | first4 = Matthew | last5 = Pfau | first5 = David | last6 = Schaul | first6 = Tom | last7 = Shillingford | first7 = Brendan | last8 = de Freitas | first8 = Nando | year = 2017| title = Learning to learn by gradient descent by gradient descent | url= | journal = Proceedings of ICML'17, Sydney, Australia}}</ref>

* In the 1990s, Meta [[Reinforcement Learning]] or Meta RL was achieved in Schmidhuber's research group through self-modifying policies written in a universal programming language that contains special instructions for changing the policy itself. There is a single lifelong trial. The goal of the RL agent is to maximize reward. It learns to accelerate reward intake by continually improving its own learning algorithm which is part of the "self-referential" policy.<ref name="sch1994">{{cite journal | last1 = Schmidhuber | first1 = Jürgen | year = 1994| title = On learning how to learn learning strategies | url= | journal = Technical Report FKI-198-94, Tech. Univ. Munich}}</ref><ref name="sch1997">{{cite journal | last1 = Schmidhuber | first1 = Jürgen | last2 = Zhao | first2 = J. | last3 = Wiering | first3 = M. | year = 1997| title = Shifting inductive bias with success-story algorithm, adaptive Levin search, and incremental self-improvement | url= | journal = Machine Learning | volume = 28 | pages = 105–130 | doi=10.1023/a:1007383707642}}</ref>

* An extreme type of Meta [[Reinforcement Learning]] is embodied by the [[Gödel machine]], a theoretical construct which can inspect and modify any part of its own software which also contains a general [[Automated theorem proving|theorem prover]]. It can achieve [[recursive self-improvement]] in a provably optimal way.<ref name="goedelmachine">{{cite journal | last1 = Schmidhuber | first1 = Jürgen | year = 2006| title = Gödel machines: Fully Self-Referential Optimal Universal Self-Improvers | url= | journal = In B. Goertzel & C. Pennachin, edsEds.: Artificial General Intelligence | pages = 199-226199–226}}</ref><ref name="scholarpedia" />

* ''Model-Agnostic Meta-Learning'' (MAML) was introduced in 2017 by Chelsea Finn et al.<ref name="maml" /> Given a sequence of tasks, the parameters of a given model are trained such that few iterations of gradient descent with few training data from a new task will lead to good generalization performance on that task. MAML "trains the model to be easy to fine-tune."<ref name="maml" /> MAML was successfully applied to few-shot image classification benchmarks and to policy gradient-based reinforcement learning.<ref name="maml">{{cite arxiv | last1 = Finn | first1 = Chelsea | last2 = Abbeel | first2 = Pieter | last3 = Levine | first3 = Sergey |year = 2017| title = Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks | eprint=1703.03400|class=cs.LG }}</ref>

* ''Discovering [[meta-knowledge]]'' works by inducing knowledge (e.g. rules) that expresses how each learning method will perform on different learning problems. The metadata is formed by characteristics of the data (general, statistical, information-theoretic,... ) in the learning problem, and characteristics of the learning algorithm (type, parameter settings, performance measures,...). Another learning algorithm then learns how the data characteristics relate to the algorithm characteristics. Given a new learning problem, the data characteristics are measured, and the performance of different learning algorithms are predicted. Hence, one can predict the algorithms best suited for the new problem.