IBM breaks image recognition software speed record

IBM Research has unveiled new deep learning software that has enabled record-breaking image recognition capabilities - and is releasing a beta of the software for AI developers to build more accurate AI models and deliver better predictions.

The software and new record represents a milestone in making Deep Learning much more practical at scale. Namely, IBM researchers have figured out how to address the bottleneck that is occurring as GPUs become faster - and are too quick to communicate with one another.

The bottom line is that the record IBM broke slashes Deep Learning training time from days to hours, which will enable their customers to more easily address larger technical challenges significantly faster. The software responsible is available to customers through IBM Systems' colleagues via PowerAI.

Key takeaways:

-  Record communication overhead and 95% scaling efficiency on the Caffe deep learning framework over 256 GPUs in 64 IBM Power systems.

- Previous record held by Facebook AI Research: scaling efficiency 89% for a training run on Caffe2, at higher communication overhead.

- Previous record by Facebook: 1 hour

·  Using this software, IBM Research achieved a new image recognition accuracy of 33.8% for a neural network trained on a very large data set (7.5M images).

- Previous record published by Microsoft: 29.8% accuracy.