Pareto-generators: Hardware-aware NAS in a learned latent space for on-device VAEs and GANs

Edge-oriented NAS has matured (Benmeziane et al., TACO 2023; Ranmal et al., 2024), but is rarely applied to the unique demands of generative decoders. ENAO (Li et al., IJCNN 2024) embeds discrete architectures into a continuous latent space to accelerate search. This project proposes:

• Learning a generative “architecture autoencoder” for VAEs/GANs (generator/decoder cells), then performing evolutionary search in the latent space (ENAO-style).
• Training a Pareto rank–preserving surrogate (HW-PR-NAS) that predicts FID/KID, throughput, memory footprint, and energy across target hardware backends, guiding search toward true Pareto fronts instead of optimizing separate objectives.
• Producing families of generators tailored for on-device tasks like training-free SemCom decoding (Du et al., 2023) or privacy-preserving augmentation at the edge, with hardware calibration loops akin to ESC-NAS.
  This is novel in making “generator NAS” a first-class, hardware-aware problem, leveraging continuous latent design spaces and rank-preserving surrogates. The practical payoff is deployable, high-quality generative models on resource-constrained devices—unlocking applications from real-time SemCom decoding to in-sensor data augmentation without cloud reliance.

References:

1. Generative Al-aided Joint Training-free Secure Semantic Communications via Multi-modal Prompts. Hongyang Du, Guangyuan Liu, D. Niyato, Jiayi Zhang, Jiawen Kang, Zehui Xiong, Bo Ai, Dong In Kim (2023). IEEE International Conference on Acoustics, Speech, and Signal Processing.
2. ENAO: Evolutionary Neural Architecture Optimization in the Approximate Continuous Latent Space of a Deep Generative Model. Zheng Li, Xuan Rao, Shaojie Liu, Bo Zhao, Derong Liu (2024). IEEE International Joint Conference on Neural Network.
3. ESC-NAS: Environment Sound Classification Using Hardware-Aware Neural Architecture Search for the Edge. Dakshina Ranmal, Piumini Ranasinghe, Thivindu Paranayapa, D. Meedeniya, Charith Perera (2024). Italian National Conference on Sensors.
4. Multi-objective Hardware-aware Neural Architecture Search with Pareto Rank-preserving Surrogate Models. Hadjer Benmeziane, Hamza Ouarnoughi, K. El Maghraoui, S. Niar (2023). ACM Transactions on Architecture and Code Optimization (TACO).