PPP-Net: Platform-aware Progressive Search for Pareto-optimal Neural Architectures
Jin-Dong Dong, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, Min Sun
Recent breakthroughs in Neural Architectural Search (NAS) have achieved state-of-the-art performances in many applications such as image recognition. However, these techniques typically ignore platform-related constrictions (e.g., inference time and power consumptions) that can be critical for portable devices with limited computing resources. We propose PPP-Net: a multi-objective architectural search framework to automatically generate networks that achieve Pareto Optimality. PPP-Net employs a compact search space inspired by operations used in state-of-the-art mobile CNNs. PPP-Net has also adopted the progressive search strategy used in a recent literature (Liu et al. (2017a)). Experimental results demonstrate that PPP-Net achieves better performances in both (a) higher accuracy and (b) shorter inference time, comparing to the state-of-the-art CondenseNet.
@misc{ dong2018ppp-net:, title={PPP-Net: Platform-aware Progressive Search for Pareto-optimal Neural Architectures}, author={Jin-Dong Dong and An-Chieh Cheng and Da-Cheng Juan and Wei Wei and Min Sun}, year={2018}, url={https://openreview.net/forum?id=B1NT3TAIM} }