Abstract
Neural Architecture Search (NAS) has shown great success in automating the design of neural networks, but the prohibitive amount of computations behind current NAS methods requires further investigations in improving the sample efficiency and the network evaluation cost to get better results in a shorter time. In this paper, we present a novel scalable Monte Carlo Tree Search (MCTS) based NAS agent, named AlphaX, to tackle these two aspects. AlphaX improves the search efficiency by adaptively balancing the exploration and exploitation at the state level, and by a Meta-Deep Neural Network (DNN) to predict network accuracies for biasing the search toward a promising region. To amortize the network evaluation cost, AlphaX accelerates MCTS rollouts with a distributed design and reduces the number of epochs in evaluating a network by transfer learning, which is guided with the tree structure in MCTS. In 12 GPU days and 1000 samples, AlphaX found an architecture that reaches 97.84% top-1 accuracy on CIFAR-10, and 75.5% top-1 accuracy on ImageNet, exceeding SOTA NAS methods in both the accuracy and sampling efficiency. Particularly, we also evaluate AlphaX on NASBench-101, a large scale NAS dataset; AlphaX is 3x and 2.8x more sample efficient than Random Search and Regularized Evolution in finding the global optimum. Finally, we show the searched architecture improves a variety of vision applications from Neural Style Transfer, to Image Captioning and Object Detection.
Introduction
Designing efficient neural network architectures manually is a tedious, trial-and-error process that relies heavily on human intuition. Neural Architecture Search (NAS) aims to automate this, but evaluating thousands of candidate architectures requires massive computational resources—often thousands of GPU days. Existing search algorithms struggle with a core trade-off: greedy algorithms (like Reinforcement Learning or Hill Climbing) often get stuck in local optima, while Random Search or Evolutionary Algorithms can be highly sample-efficient.
AlphaX is a novel NAS agent that bridges this gap by making Monte Carlo Tree Search (MCTS) practical, scalable, and highly sample-efficient. By combining MCTS with a predictive Meta-DNN and leveraging distributed computing with transfer learning, AlphaX discovers top-performing networks with significantly reduced computational overhead.
Methodology
AlphaX takes inspiration from AlphaGo, applying a similar MCTS framework to the architecture search space. The core components include:
MCTS with UCB1
AlphaX uses MCTS to navigate the architecture search space, dynamically balancing exploration (trying new, unvisited network structures) and exploitation (refining known good structures) using the UCB1 algorithm. Unlike greedy methods that might converge prematurely, MCTS systematically explores less-visited nodes, preventing the agent from getting trapped in local optima.
Meta-DNN Predictive Model
Instead of fully training every network during an MCTS rollout, AlphaX uses a multi-stage Multilayer Perceptron (MLP) to predict the accuracy of candidate architectures. This acts as a "value network" to estimate how promising a search branch is without the cost of actual training. Predicting network accuracy is challenging because many architectures cluster around specific accuracy bands; AlphaX's multi-stage approach (using dedicated MLPs for different accuracy ranges) significantly outperforms standard single-stage models.
Transfer Learning and Distributed Architecture
When actual training is required, AlphaX avoids starting from scratch. It searches the MCTS tree for a previously trained "parent" network with the minimal edit distance to the new "child" network and transfers overlapping weights. This reduces the required training time from 70 epochs to just 20. Furthermore, the system decouples the search engine from training workers, using a two-step backpropagation: a preemptive tree update using predicted accuracy, followed by a true update once actual training finishes.
Search Space and Datasets
AlphaX was evaluated across several benchmarks to test its versatility and efficiency:
- CIFAR-10 & ImageNet: Used for primary architecture search and transferability testing.
- NASBench-101: A large-scale dataset with over 420k networks, used to rigorously evaluate search efficiency without compute bottlenecks.
- MSCOCO: Used for downstream tasks like object detection and image captioning.
Results and Discussion
AlphaX demonstrates state-of-the-art performance with remarkable efficiency:
- Accuracy and Efficiency: On CIFAR-10, AlphaX found an architecture with 97.84% top-1 accuracy in just 12 GPU days (1,000 samples), matching methods like AmoebaNet while using 27x fewer samples.
- Sample Efficiency: On NASBench-101, AlphaX was 3x faster than Random Search and 2.8x faster than Regularized Evolution at finding the global optimum.
- Downstream Impact: Replacing standard backbones with the AlphaX-searched architecture improved Object Detection mAP on SSD from 20.1% to 23.7% and enhanced results in Neural Style Transfer and Image Captioning.
Conclusion
AlphaX proves that MCTS, when combined with predictive modeling and efficient weight sharing, is a powerful tool for NAS. While the Meta-DNN design is currently heuristic and search space constraints are applied for manageability, the results show a clear path forward for more sample-efficient automated architecture design. Future work will focus on refining the Meta-DNN's generalization and scaling the system to handle unconstrained search spaces.
Summary updated