Alex Graves - Automated Curriculum Learning for Neural Networks (2017)

Context

Learned in this study

Things to explore

• Neural architecture search where picking the right action is seen as a multi-armed bandit

Overview

Notes

1. Introduction

• One reason for the slow adoption of curriculum learning is that its effectiveness is highly sensitive to the mode of progression through the tasks
• One popular approach is to define a hand-chosen performance threshold for advancement to the next task, along with a fixed probability of returning to earlier tasks, to prevent forgetting
• We propose to treat the decision about which task to study next as a stochastic policy, continuously adapted to optimize some notion of learning progress

2. Background

• We consider supervised or unsupervised learning problems where target sequences $\textbf{b}^1, \textbf{b}^2, \dots$ are conditionally modelled given their respective input sequences $\textbf{a}^1, \textbf{a}^2, \dots$
• We suppose that the targets are drawn from a finite set $\mathcal{B}$
• As is typical for neural networks, sequences may be grouped together in batches $(\textbf{b}^{1:B}, \textbf{a}^{1:B})$ to accelerate training
• The conditional probability output by the model is
  $$p(\textbf{b}^{1:B}, \textbf{a}^{1:B}) = \prod_{i,j} p(\textbf{b}_j^i\ |\ \textbf{b}_{1:j-1}^i, \textbf{a}_{1:j-1}^i)$$

• We consider each batch as a single exemple $\textbf{x}$ from $\mathcal{X} := (\mathcal{A} \times \mathcal{B})^N$, and write $p(\textbf{x}) := p(\textbf{b}^{1:B}, \textbf{a}^{1:B})$ for its probability
• A task is a distribution $D$ over sequences from $\mathcal{X}$
• A curriculum is an ensemble of tasks $D_1, \dots, D_N$
• A sample is an example drawn from one of the tasks of the curriculum
• A syllabus is a time-varying sequence of distributions over tasks

2.1. Curriculum Learning

• We consider two related settings:
  • The multiple tasks setting, the goal is to perform as well as possible on all tasks in the ensemble
  • The target task setting, the goal is to minimize the loss on the final task, the other tasks acting as a series of stepping stones to the real problem

2.2 Adversarial Multi-Armed Bandits

• We view a curriculum containing $N$ tasks as an $N$-armed bandit, and a syllabus as an adaptive policy which seeks to maximize payoffs from this bandit
• In the bandit setting, an agent selects a sequence of arms (actions) $a_1, \dots, a_T$ over $T$ rounds of play ($a_t \in \{1, \dots, N\}$)
• After each round, the selected arm yields a payoff $r_t$; the payoffs for the other arms are not observed
• The classical algorithm for adversarial bandits is Exp3, which uses multiplicative weight updates to guarantee low regret with respect to the best arm
• On a round $t$, the agent selects an arm stochastically according to a policy $\pi_t$. This policy is defined by a set of weights $w_{t, i}$:
  $$\pi_t^{\text{EXP3}}(i) := \frac{e^{w_{t,i}}}{\sum_{j=1}^N e^{w_{t,j}}}$$

3. Learning Progress Signals

• Ideally we would like the policy to maximize the rate at which we minimize the loss, and the reward should reflect this rate
• However, it usually is computationally undesirable or even impossible to measure the effect of a training sample on the target objective, and we therefore turn to surrogate measures of progress
• Two types of measures:
  • Loss-driven, in the sense that they equate reward with a decrease in some loss
  • Complexity-driven, when they equate reward with an increase in model complexity

3.2. Complexity-driven Progress

• According to the Minimum Description Length (MDL) principle, increase in the model complexity by a certain amount is only worthwhile if it compresses the data by a greater amount
  • We would therefore expect the complexity to increase most in response to the training examples from which the network is best able to generalize

5. Conclusion

• We note that uniformly sampling from all tasks is a surprisingly strong benchmark. We speculate that this is because learning is dominated by gradients from the tasks on which the network is making fastest progress, inducing a kind of implicit curriculum, albeit with the inefficiency of unnecessary samples

See also

References

• Graves, Alex, et al. "Automated Curriculum Learning for Neural Networks." arXiv preprint arXiv:1704.03003 (2017).