PaperlayerView exact revision on arXiv

Source-linked AI summary

MONAS: Multi-Objective Neural Architecture Search using Reinforcement Learning

Chi-Hung Hsu, Shu-Huan Chang, Jhao-Hong Liang, Hsin-Ping Chou, Chun-Hao Liu, Shih-Chieh Chang, Jia-Yu Pan, Yu-Ting Chen, Wei Wei, Da-Cheng Juan

arXiv:1806.10332v2cs.LGcs.AIstat.ML

TL;DR

Neural architecture search has often prioritized prediction accuracy, producing models that may be unsuitable for resource-limited deployment. MONAS uses deep reinforcement learning to search architectures with accuracy and computation-related objectives, and its models achieve higher accuracy and lower energy consumption than CondenseNet while satisfying specified constraints.

  • Problem

    Accuracy-only neural architecture search can produce complex models unsuitable for computing environments with limited resources.

    Source: Introduction · p. 1

  • Method

    MONAS is a deep-reinforcement-learning framework that searches neural architectures using prediction accuracy and additional objectives such as energy consumption, with MONAS-S extending it through weight sharing.

    Source: Introduction · p. 1, Reinforcement Learning Process · p. 3

  • Results

    MONAS models achieve higher accuracy and lower energy consumption than state-of-the-art CondenseNet models while guiding searches toward specified constraints.

    Source: Introduction · p. 1, Results and Discussions · pp. 4–5

  • Takeaways & Limitations

    MONAS provides a general architecture-search framework for finding models that balance predictive performance with application-specific resource objectives.

    Source: Introduction · p. 1

    MAC-operation estimates omit fully connected layers and skip connections because their costs are treated as common or negligible across target networks.

    Source: Experimental Setup · p. 4

Abstract

from arXiv · show

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.

Introduction

MONAS searches neural architectures using multiple objectives rather than accuracy alone, incorporating application constraints such as energy or power. It is presented as a general and scalable framework whose discovered models can improve both accuracy and energy consumption relative to CondenseNet.

  • MONAS incorporates prediction accuracy and additional objectives, such as energy consumption, during architecture search.Source: Introduction · p. 1
  • Users can convert application- or platform-specific constraints into objectives that guide the search process.Source: Introduction · p. 1
  • MONAS searches across diverse CNN families, including AlexNet-like, CondenseNet-like, and ResNet-like models.Source: Introduction · p. 1
  • MONAS models achieve higher accuracy and lower energy consumption than state-of-the-art CondenseNet models.Source: Introduction · p. 1
  • MONAS-S uses weight sharing to search a design space up to 10^22 larger while running 3X faster than MONAS.Source: Introduction · p. 1
  • The paper is organized around prior NAS work, the MONAS and MONAS-S methods, experimental setup, results, and conclusions.Source: Introduction · p. 1

Background

Prior neural architecture search work largely optimized classification accuracy, while other multi-objective approaches used evolutionary algorithms or Bayesian optimization. MONAS instead treats computation cost, such as power or MAC operations, as an architecture-search objective alongside accuracy.

  • Earlier NAS methods focused primarily on discovering models with high classification accuracy without penalizing resource consumption.Source: Background · pp. 1–2
  • Existing multi-objective architecture-search studies used evolutionary algorithms or Bayesian optimization to balance objectives.Source: Background · p. 2
  • MONAS distinguishes itself by optimizing CNN accuracy together with computation cost, including power consumption or MAC operations.Source: Background · p. 2
  • MONAS can impose peak-power or accuracy prerequisites, or directly weight accuracy and computation cost to express a trade-off.Source: Background · p. 2

Proposed Framework: MONAS

MONAS uses reinforcement learning to search neural architectures while jointly optimizing accuracy and application-specific resource objectives. Its two-stage controller–target-network process supports broad CNN search spaces, multiple reward constraints, and scalable weight-sharing training.

Framework Overview

  • MONAS generates CNN hyperparameters with an RNN, trains the resulting target network, and uses its accuracy and energy consumption as reinforcement-learning rewards.The robot network updates through policy gradients after evaluating each target network.Source: Framework Overview · p. 2, Reinforcement Learning Process · p. 2

Implementation Details

  • The controller samples hyperparameters sequentially with a one-layer LSTM, using softmax probabilities and one-hot feedback to select each next value.Search spaces include AlexNet-like filter configurations and CondenseNet-like stage and growth settings.Source: Implementation Details · p. 2

Reinforcement Learning Process

  • The policy-gradient process treats each sampled architecture as a trajectory whose final reward reflects target-network accuracy and power consumption.The authors report faster convergence when updating the controller after every sampled target network, using N = 1.Source: Reinforcement Learning Process · p. 3
  • MONAS-S improves scalability by combining MONAS with ENAS-style weight sharing, pretraining a shared DAG before controller updates using multi-objective rewards.Sampled models receive rewards from validation accuracy together with additional measures such as energy consumption.Source: Reinforcement Learning Process · p. 3
  • MONAS supports mixed accuracy–energy rewards, peak-power constraints, accuracy constraints, and MAC constraints that proxy power use.Constraint handling assigns zero or negative rewards when sampled networks violate specified limits.Source: Reinforcement Learning Process · p. 3

Experimental Setup

The experiments evaluate MONAS on CIFAR-10 using validation accuracy, power and energy measurements, and MAC-operation estimates as experimental objectives and complexity measures.

  • Peak power, average power, and CUDA-kernel runtime are measured with NVIDIA nvprof to quantify target-network energy consumption.Source: Experimental Setup · p. 3
  • CIFAR-10 validation accuracy is used as an objective in the reward function, with 5,000 training images held out for validation.Source: Experimental Setup · p. 3
  • Convolutional-layer MAC operations are calculated as K*K*Cin*H*W*Cout, using kernel size, output-map dimensions, and input/output channels.Source: Experimental Setup · pp. 3–4
  • Depthwise-separable convolutional MAC operations are calculated as Cin*H*W*(K*K+Cout).Source: Experimental Setup · p. 4
  • Fully connected layers and skip connections are excluded because the fully connected output layer is shared across target networks and skip connections contribute few MAC operations.Source: Experimental Setup · p. 4

Results and Discussions

MONAS guides architecture search toward accuracy–resource trade-offs and application constraints, including power and MAC costs. Experiments show this guidance remains effective across larger search spaces, while training details specify the target-network procedures.

Results and Discussions

  • 340 of 600 target networks met the 70-watt peak-power constraint, while 498 of 600 met the minimum accuracy of 0.85.Random search explored uniformly without particular emphasis, whereas MONAS concentrated exploration in constraint-satisfying regions.Source: Results and Discussions · p. 4
  • After 200 iterations, over 60% of MONAS architectures satisfied the 70-watt constraint, compared with under 10% for random search.The result indicates faster concentration in the desirable region during limited search time.Source: Results and Discussions · p. 4
  • With α = 0.25 MONAS favors lower energy, whereas α = 0.75 favors higher accuracy; both settings outperform random search on the targeted trade-off.The Pareto-front analysis shows that reward weighting steers search tendency toward different objectives.Source: Results and Discussions · pp. 4–5
  • MONAS discovered CondenseNet architectures with higher accuracy and lower energy than the best human-designed models.The comparison is presented through the discovered Pareto frontier and completed hyperparameter results in Table 2.Source: Results and Discussions · p. 5

Scalability

  • MONAS-S guided search in a 1.6 × 10^29-network space toward models using fewer MAC operations while mostly satisfying the constraint.Compared with ENAS, its controller favored depth-wise separable convolutions and average pooling while maintaining high-level accuracy.Source: Scalability · pp. 5–6
  • MONAS-S preferentially selected MAC-efficient operations in the first four layers, where feature maps are largest and computation needs are higher.Later-layer operator preferences were more similar to ENAS, while the strongest differences appeared in the initial layers.Source: Scalability · p. 6

Training Details for Target Networks

  • AlexNet target networks were trained for 256 epochs, CondenseNet targets for 30 epochs during search and 300 epochs for final evaluation, and MONAS-S for 600 controller epochs.The controllers used one-layer LSTMs for AlexNet and CondenseNet, while MONAS-S extended ENAS with weight sharing.Source: Training Details for Target Networks · p. 6

Conclusion

MONAS uses deep reinforcement learning to search architectures under application-specific constraints and guide the search toward desirable regions. The framework’s experiments show successful multi-objective model discovery, while MONAS-S improves search scalability and reveals computation-aware operation choices.

  • MONAS uses deep reinforcement learning to adapt architecture search to application-specific constraints and guide exploration toward a region of interest.Source: Conclusion · p. 6
  • Figure 7 compares 1000 target networks sampled by ENAS and MONAS-S after both controllers were trained for 600 epochs.Source: Figure 7 · p. 7
  • The operation-distribution experiment compares 1000 sampled 12-layer networks from ENAS and MONAS-S across 12000 layer-operation instances.Source: Figure 8 · p. 7
  • MONAS-S samples MAC-efficient operations more often in the first four layers, which have the largest feature maps and computational needs.Source: Figure 9 · p. 7

Related papers

Loading 1806.10332v2…