Reg comments

Baidu Research grills AI models on deep learning

Tool aims to reduce machine-training time

Baidu Research has launched DeepBench, a new tool for AI researchers interested in assessing deep learning operations across hardware platforms.

The announcement was made at the O’Reilly Artificial Intelligence conference by Greg Diamos and Sharan Narang, researchers working at Baidu’s AI lab in Silicon Valley.

Deep learning is a technique used in machine learning that is rapidly gaining popularity amongst the AI community. It uses neural networks and applies a set of algorithms with parameters fine-tuned during the learning phase to allow input data to trickle through multiple processing layers before computing the output.

It is being used to tackle a variety of problems, including natural language processing and machine vision.

Deep learning frameworks such as Theano, a software that allows developers to perform numerical calculations, or Google’s Tensorflow, which has tools for language understanding, are used to build deep learning models. The model is trained using neural networks on hardware such as NVIDIA GPUs or Intel's Xeon Phi processor.

“Because every deep learning model uses these operations with different parameters, the optimization space for hardware and software targeting deep learning is large and underspecified,” Baidu said in a blog post.

With DeepBench, Baidu Research hopes that measuring performance on deep learning models will allow developers to find ways to reduce wasting computing power during training.

“Deep learning developers and researchers want to train neural networks as fast as possible. Right now we are limited by computing performance, said Dr. Diamos.

“The first step in improving performance is to measure it, so we created DeepBench and are opening it up to the deep learning community. We believe that tracking performance on different hardware platforms will help processor designers better optimize their hardware for deep learning applications.” ®


Biting the hand that feeds IT © 1998–2017