[Paper Review] MONAS: Multi-Objective Neural Architecture Search using Reinforcement Learning

Chi-Hung Hsu, Shih-Chieh Chang+8 more|arXiv (Cornell University)|Jun 27, 2018

Adversarial Robustness in Machine Learning11 references150 citations

TL;DR

MONAS uses reinforcement learning to search for neural network architectures that balance accuracy and other objectives like power consumption, with MONAS-S extending to weight-sharing for scalability.

ABSTRACT

Recent studies on neural architecture search have shown that automatically designed neural networks perform as good as expert-crafted architectures. While most existing works aim at finding architectures that optimize the prediction accuracy, these architectures may have complexity and is therefore not suitable being deployed on certain computing environment (e.g., with limited power budgets). We propose MONAS, a framework for Multi-Objective Neural Architectural Search that employs reward functions considering both prediction accuracy and other important objectives (e.g., power consumption) when searching for neural network architectures. Experimental results showed that, compared to the state-ofthe-arts, models found by MONAS achieve comparable or better classification accuracy on computer vision applications, while satisfying the additional objectives such as peak power.

Motivation & Objective

• Motivate automatic neural architecture search to balance prediction accuracy with resource-related objectives such as power consumption.
• Propose MONAS, a reinforcement-learning-based framework that incorporates multiple objectives into the search process.
• Demonstrate adaptability to application-specific constraints and apply MONAS to AlexNet-like and CondenseNet-like model families.
• Introduce MONAS-S, a scalable extension using weight-sharing to handle larger search spaces.

Proposed method

• Use a one-layer LSTM-based robot network (RNN) as a controller to generate hyperparameters for target CNNs.
• Train the target network with the generated hyperparameters and use its validation accuracy and power consumption as rewards for the controller.
• Apply policy gradient reinforcement learning to update the controller parameters, with rewards combining accuracy and energy (or other constraints).
• Define multiple reward functions to realize different optimization goals, including mixed accuracy-energy trade-offs and constraint-based rewards.
• Extend MONAS with MONAS-S by adopting weight-sharing search (DAG) to pre-train shared weights and accelerate search in very large spaces.
• Measure rewards via GPU profiling (peak/average power) and MAC operations to estimate energy and computation costs.

Experimental results

Research questions

• RQ1 Can MONAS adapt to different multi-objective reward functions and constraints?
• RQ2 How efficiently does MONAS guide search toward regions satisfying power or accuracy constraints?
• RQ3 How does the Pareto frontier change under different reward configurations?
• RQ4 Can MONAS discover architectures that outperform state-of-the-art baselines under multi-objective goals?
• RQ5 Is MONAS-S scalable to very large search spaces while maintaining energy-aware objectives?

Key findings

• MONAS guides the search toward constraint-satisfying architectures more efficiently than random search.
• Different alpha settings in the mixed reward steer the search toward higher accuracy or lower energy, illustrating controllable trade-offs.
• MONAS discovers architectures on the Pareto frontier that can surpass CondenseNet baselines in accuracy and energy efficiency.
• MONAS-S scales to extremely large search spaces (e.g., 1.6e29 possibilities) and biases toward lower MAC operations while maintaining accuracy.
• Compared to ENAS, MONAS-S yields target networks with lower MAC operations and favorable layer-wise operation distributions.