Introduction
Neural Architecture Search (NAS) has revolutionized the design of deep learning models, frequently discovering architectures that outperform those manually crafted by human experts. However, a significant drawback of many automated search methods is the resulting "topological messiness." These auto-generated networks often feature irregular, overly complex graphs that lead to high inference latency, making them impractical for deployment on mobile devices or real-time systems. In contrast, human-designed networks like ResNet are characterized by "topological elegance"—simple, repetitive patterns optimized for hardware efficiency.
Methodology: Learning from Experts
The researchers propose IRLAS (Inverse Reinforcement Learning for Architecture Search), a framework designed to bridge the gap between automated performance and human-like structural simplicity. The core of IRLAS is the Mirror Stimuli Function, inspired by biological mirror neurons that fire both during action and observation.
Key Components
- Topological Encoding: Network structures are represented as "state feature codes" that capture operation types, kernel sizes, and connection indexes. These codes are aggregated into a "feature count" representing the overall topology.
- Inverse Reinforcement Learning (IRL): Rather than manually defining structural constraints, the system uses IRL to observe an "expert" network (such as a modified ResNet). It learns a weight vector ($w$) that characterizes the expert's structural preferences.
- Heuristic Guidance: This learned structural reward ($F_{topology}$) is combined with a standard accuracy reward ($F_{accuracy}$). This provides "soft guidance," encouraging the agent to mimic expert elegance while remaining free to explore high-performing novel structures.
- Algorithm Agnostic: While primarily tested with Q-learning, the Mirror Stimuli Function can be integrated into various NAS methods, including gradient-based approaches like DARTS.
Experimental Results
IRLAS was evaluated on industry-standard benchmarks, including CIFAR-10 and ImageNet (mobile setting). The results demonstrate that the framework successfully identifies a "sweet spot" between accuracy and efficiency.
Key Findings
- Efficiency Gains: On ImageNet, IRLAS reached 75.28% top-1 accuracy while being 2–4× faster than popular auto-generated architectures like NASNet or AmoebaNet.
- Superior to Human Experts: A specialized "fast" version of the IRLAS model achieved higher accuracy than MobileNetV2 while being 10% faster, proving that the algorithm can refine and improve upon the human wisdom it mimics.
- State-of-the-Art Accuracy: On CIFAR-10, the search discovered models with a 2.60% error rate, competitive with the best models in its class.
Discussion and Future Work
The success of IRLAS highlights that human-designed regularity is not just an aesthetic choice but a functional one that directly impacts hardware performance. By treating structural simplicity as a rewardable trait rather than a hard constraint, IRLAS maintains the exploratory power of NAS while ensuring the resulting models are deployable.
Future research may explore extracting structural priors from multiple expert networks simultaneously to create more versatile guidance. Additionally, investigating how IRLAS might reduce the overall computational cost of the search process by narrowing the search space to "elegant" regions remains a promising avenue for development.
Summary updated