We introduce MetaQNN, a meta-modeling algorithm based on reinforcement learning to automatically generate high-performing CNN architectures for a given learning task. The learning agent is trained to sequentially choose CNN layers using Q-learning with an $\epsilon$-greedy exploration strategy and experience replay. The agent explores a large but finite space of possible architectures and iteratively discovers designs with improved performance on the learning task. On image classification benchmarks, the agent-designed networks (consisting of only standard convolution, pooling, and fully-connected layers) beat existing networks designed with the same layer types and are competitive against the state-of-the-art methods that use more complex layer types. We also outperform existing meta-modeling approaches for network design on image classification tasks.
In this project, we show that a simple regression model, based on support vector machines, can predict the final performance of partially trained neural network configurations using features based on network architectures, hyperparameters, and time-series validation performance data.
We use this regression model to develop an early stopping strategy for neural network configurations. With this early stopping strategy, we obtain significant speedups in both hyperparameter optimization and meta-modeling. Particularly in the context of meta-modeling, our method can learn to predict the performance of drastically different architectures and is seamlessly incorporated into reinforcement learning-based architecture selection algorithms. Finally, we show that our method is simpler, faster, and more accurate than Bayesian methods for learning curve prediction.