Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Evolving Search Space for Neural Architecture Search (2011.10904v2)

Published 22 Nov 2020 in cs.CV

Abstract: The automation of neural architecture design has been a coveted alternative to human experts. Recent works have small search space, which is easier to optimize but has a limited upper bound of the optimal solution. Extra human design is needed for those methods to propose a more suitable space with respect to the specific task and algorithm capacity. To further enhance the degree of automation for neural architecture search, we present a Neural Search-space Evolution (NSE) scheme that iteratively amplifies the results from the previous effort by maintaining an optimized search space subset. This design minimizes the necessity of a well-designed search space. We further extend the flexibility of obtainable architectures by introducing a learnable multi-branch setting. By employing the proposed method, a consistent performance gain is achieved during a progressive search over upcoming search spaces. We achieve 77.3% top-1 retrain accuracy on ImageNet with 333M FLOPs, which yielded a state-of-the-art performance among previous auto-generated architectures that do not involve knowledge distillation or weight pruning. When the latency constraint is adopted, our result also performs better than the previous best-performing mobile models with a 77.9% Top-1 retrain accuracy.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (6)
  1. Yuanzheng Ci (6 papers)
  2. Chen Lin (75 papers)
  3. Ming Sun (146 papers)
  4. Boyu Chen (30 papers)
  5. Hongwen Zhang (59 papers)
  6. Wanli Ouyang (358 papers)
Citations (39)

Summary

Evolving Search Space for Neural Architecture Search: An Analytical Review

The paper "Evolving Search Space for Neural Architecture Search" by Yuanzheng Ci et al. addresses a critical challenge in the domain of Neural Architecture Search (NAS): the adverse impact of large search spaces on the efficacy of prevalent NAS methods. Traditionally, one might hypothesize that expanding the search space would result in better-performing neural network architectures due to a richer pool of candidates. However, the authors note a counterintuitive phenomenon where enlarging the search space actually degrades the performance of existing NAS approaches like DARTS, ProxylessNAS, and SPOS. The research responds to this challenge by proposing the Neural Search-space Evolution (NSE) scheme, designed specifically to adaptively manage and evolve large search spaces in NAS.

Key Contributions

  1. Search Space Evolution: The NSE method introduces a novel approach to managing large search spaces. Instead of expanding the entire search space at once, NSE begins with a manageable subset from which it searches for an optimized subset. It then replenishes this subset with new operations from the broader search pool, evolving iteratively. This allows for productive exploration while efficiently retaining the knowledge of past trials.
  2. Multi-Branch Scheme Implementation: The paper extends the flexibility of architecture designs by incorporating a learnable multi-branch architecture model. Different from single-branch methods like DARTs, this approach allows adaptive selection of multiple operations at each step, significantly increasing the structural diversity.
  3. Performance and Efficiency: The proposed architecture achieves state-of-the-art performance, with a 77.3% top-1 accuracy on the ImageNet dataset under a FLOPs constraint of 333M, and 77.9% when considering latency constraints. These outcomes demonstrate significant improvements over prior automatic architecture designs without using techniques such as knowledge distillation or weight pruning.

Numerical Results and Evaluation

The authors extensively evaluate the performance of the NAS methods in both FLOPs and latency-constrained scenarios. The results manifest the effectiveness of NSE with considerable performance improvements over existing methods in terms of both search efficiency and final model accuracy.

  • For the FLOPs constraint, models derived using NSE exhibit higher accuracy with approximately 333M FLOPs, exceeding the performance of existing architectures like ProxylessNAS and DARTs within comparable computational constraints.
  • In latency-focused experiments, NSE-derived models demonstrate superior performance on a GTX TITAN Xp GPU with a reduced latency cost, outperforming designs from established benchmarks like MobileNetV3 and MixNet-S.

Theoretical and Practical Implications

The proposed NSE framework constitutes a pivotal advancement for NAS in large search spaces, alleviating the dependency on manual search space design which implicitly incorporates domain expertise. The iterative nature of NSE, alongside the flexibility offered by the multi-branch scheme, provides a framework for continuous improvement and adaptability. These enhancements position NSE as a viable candidate for real-world deployment where architectural and computational constraints fluctuate rapidly.

Future Directions

Several avenues for future exploration are implicated by this work. Importing concepts from continual learning can further optimize the handling of evolving search spaces. Additionally, exploring complementary techniques such as automatic model pruning and distillation could enhance the robustness and applicability of NAS frameworks under varied conditions.

In conclusion, the contribution from Ci et al. delivers a robust solution to some of the inherent limitations of current NAS practices for large search spaces, providing both theoretical insights and practical tools for the evolution of deep learning model design automation.

Youtube Logo Streamline Icon: https://streamlinehq.com