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

DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures (1806.08198v2)

Published 21 Jun 2018 in cs.CV and cs.LG

Abstract: Recent breakthroughs in Neural Architectural Search (NAS) have achieved state-of-the-art performances in applications such as image classification and LLMing. However, these techniques typically ignore device-related objectives such as inference time, memory usage, and power consumption. Optimizing neural architecture for device-related objectives is immensely crucial for deploying deep networks on portable devices with limited computing resources. We propose DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures, optimizing for both device-related (e.g., inference time and memory usage) and device-agnostic (e.g., accuracy and model size) objectives. DPP-Net employs a compact search space inspired by current state-of-the-art mobile CNNs, and further improves search efficiency by adopting progressive search (Liu et al. 2017). Experimental results on CIFAR-10 are poised to demonstrate the effectiveness of Pareto-optimal networks found by DPP-Net, for three different devices: (1) a workstation with Titan X GPU, (2) NVIDIA Jetson TX1 embedded system, and (3) mobile phone with ARM Cortex-A53. Compared to CondenseNet and NASNet (Mobile), DPP-Net achieves better performances: higher accuracy and shorter inference time on various devices. Additional experimental results show that models found by DPP-Net also achieve considerably-good performance on ImageNet as well.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Jin-Dong Dong (4 papers)
  2. An-Chieh Cheng (10 papers)
  3. Da-Cheng Juan (38 papers)
  4. Wei Wei (425 papers)
  5. Min Sun (108 papers)
Citations (177)

Summary

DPP-Net: Device-Aware Neural Architecture Search

The paper "DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures" offers a refined approach to Neural Architecture Search (NAS) designed to optimize deep learning models not only for accuracy and model size, but also for device-related metrics such as inference time and memory usage. This presents a comprehensive multi-objective optimization strategy that recognizes the constraints of deploying neural networks on devices with limited computational resources like mobile phones and embedded systems.

The authors introduce DPP-Net, a search methodology centered around device-aware considerations, which is particularly focused on achieving Pareto-optimal network architectures. This approach balances the trade-offs between multiple objectives, ensuring optimized architectures tailored for specific device constraints. DPP-Net operationalizes this through a progressively expanding search space reflective of mobile CNNs, leveraging operations known for efficiency like depth-wise and group convolutions, which are pivotal in mobile contexts.

A critical aspect of DPP-Net is its incorporation of a surrogate model—the use of a Recurrent Neural Network—as part of a Sequential Model-Based Optimization (SMBO) strategy. This component predicts the performance of network architectures during the search process, expediting the discovery of suitable candidates while harnessing a form of mutation that incrementally builds network complexity. The strategy effectively reveals architectures sitting on the Pareto front, ensuring optimal trade-offs between device-specific parameters like inference time and other design priorities such as accuracy.

The paper details experimental comparisons against prior NAS approaches, such as NASNet and models derived through genetic algorithms, and highlights notable improvements. On CIFAR-10, DPP-Net discovered architectures exhibiting superior accuracy and reduced inference times across diverse hardware platforms. Additionally, these architectures demonstrated robustness and adaptability when transferred to the ImageNet dataset, confirming their versatility beyond smaller image classification tasks.

Device performance evaluations further elucidated how DPP-Net identifies architectures that excel on platform-specific metrics, proving that optimizations cannot be generalized across different hardware configurations. The consideration of device heterogeneity—particularly evident in varying inference times across workstation GPUs and mobile CPUs—underscores the necessity of device-aware NAS methodologies in practical applications.

The implications of DPP-Net extend toward more efficient deployment of deep neural networks on computationally constrained devices, potentially reducing energy consumption and improving user experiences in mobile environments. Its ability to refine architectures by employing an informed search mechanism may further inspire future AI research to integrate multi-objective optimization principles into other automated design processes, broadening the scope for highly tailored and efficient AI solutions.

Future developments may explore expanded search spaces and improved integration of novel operations to further refine device-aware architectures. Continued refinement of surrogate models to predict more complex performance outcomes could enable even more precise optimization in multifaceted deployment contexts.

Overall, DPP-Net represents an evolution in NAS methodologies by integrating device-aware objectives, balancing multiple performance metrics, and uncovering optimally tailored architectures for diverse hardware environments. This careful attention to the constraints and objectives relevant to actual device deployment illustrates a significant advancement within the field of neural architecture optimization.