Official Review · ICLR 2024

This paper proposes a new diffusion-based NAS algorithm called DiNAS, This method use graph diffusion models to generate architectures. Specifically, the method uses conditional diffusion model to control the properties of the generated architectures, e.g., performance and latency. The entire framework can be optimized by gradient-based approaches. The model is evaluated on several NAS benchmarks to demonstrate the effectiveness. 1. This paper is well-written and easy to follow. 2. Using graph diffusion to generate architectures is a reasonable idea. 1. The method is very straightforward. The method designs lack the relation with NAS problems, making the technical contribution limited. 2. There has been diffusion-based NAS method [1], but the paper does not mention the difference between the two methods. 3. All experiment are conducted on NAS benchmarks. In my opinion, NAS benchmarks provide quick feedback for developing new NAS methods, but NAS methods still need evaluation out of NAS benchmark to avoid overfitting on these benchmarks. 4. In table 5, the search time of DiNAS excludes the performance estimation time of architecture training. It is unfair to compare with other methods with performance estimation time, e.g. NASNET, DARTS. [1] DiffusionNAG: Task-guided Neural Architecture Generation with Diffusion Models See weaknesses.