Sign In

Accel-NASBench: A Sustainable Benchmark for Accelerator-Aware Neural Architecture Search

Core Concepts
The authors present a technique to efficiently construct a realistic neural architecture search (NAS) benchmark for the large-scale ImageNet2012 dataset, combined with performance metrics for various hardware accelerators including GPUs, TPUs, and FPGAs.
The paper addresses the key challenges in constructing NAS benchmarks for large-scale datasets and hardware-aware model search: Dataset Proxies and Sustainability: Existing NAS benchmarks use small datasets like CIFAR-10 or downsampled variants of ImageNet, which do not represent the true complexity of large-scale datasets. The authors propose a method to search for training proxies that can reduce the cost of benchmark construction for ImageNet2012 by significant margins. On-Accelerator Performance: Previous NAS benchmarks use simplified analytical models or proxy datasets to estimate hardware performance, leading to unrealistic evaluations. The authors construct a benchmark that uses end-to-end throughput and latency measurements on real hardware accelerators, including GPUs, TPUs, and FPGAs. The key steps are: Search for a proxified training scheme that maintains architecture rankings relative to a high-fidelity reference scheme, while reducing training cost by 5.6x. Collect datasets of 5.2k randomly sampled architectures on ImageNet2012 using the proxified scheme, along with their performance on various accelerators. Train surrogate models to predict accuracy and on-device performance, achieving high correlation with true values. The authors demonstrate the utility of the constructed benchmark, Accel-NASBench, by performing uni-objective and bi-objective NAS experiments using various optimizers and hardware platforms. The results show that the benchmark can accurately simulate real-world performance and enable the discovery of state-of-the-art models at zero cost.
The training time of the proxified scheme is approximately 5.6x faster than the reference scheme.

Key Insights Distilled From

by Afzal Ahmad,... at 04-15-2024
Accel-NASBench: Sustainable Benchmarking for Accelerator-Aware NAS

Deeper Inquiries

How can the proposed technique be extended to construct benchmarks for other large-scale datasets beyond ImageNet2012

The proposed technique for constructing benchmarks for large-scale datasets, as demonstrated with ImageNet2012, can be extended to other datasets by following a similar methodology. To construct benchmarks for different datasets, researchers can first identify the target dataset and its characteristics. Then, they can adapt the search for training proxies to find a scheme that reduces the training cost while maintaining architecture rankings relative to true evaluation. By collecting datasets for accuracy and on-device performance using the optimized training scheme, researchers can build benchmarks for other large-scale datasets. The key lies in selecting appropriate hyperparameters for the training proxies that balance computational efficiency with accurate emulation of the true evaluation process.

What are the potential limitations or drawbacks of using training proxies in the context of NAS, and how can they be addressed

While training proxies offer a cost-effective way to construct NAS benchmarks, they come with potential limitations. One drawback is the trade-off between computational efficiency and accuracy. Training proxies may not fully capture the complexities of the true evaluation process, leading to discrepancies in model performance predictions. To address this limitation, researchers can explore more sophisticated training proxy schemes that incorporate additional factors or fine-tune the existing proxies to better emulate the true evaluation. Additionally, conducting thorough validation and sensitivity analyses on the training proxies can help identify and mitigate any biases or inaccuracies introduced by using proxies in NAS benchmark construction.

How can the insights from this work on accelerator-aware NAS be applied to other domains beyond computer vision, such as natural language processing or speech recognition

The insights gained from accelerator-aware NAS, particularly in optimizing models for on-device performance metrics, can be applied to various domains beyond computer vision. In natural language processing (NLP), for example, researchers can leverage similar techniques to design neural network architectures that are optimized for inference speed and efficiency on different hardware platforms. By considering factors such as memory bandwidth, data reuse, and device-specific characteristics, NLP models can be tailored to achieve high performance on a range of accelerators. Similarly, in speech recognition tasks, applying accelerator-aware design principles can lead to the development of models that are not only accurate but also optimized for low-latency inference, making them suitable for real-time applications. By incorporating hardware-specific considerations into the design process, researchers can create efficient and effective models across various domains.