In IBM Research developed, providing almost linear speed scaling deep learning
The experts of IBM Research was able to develop software for distributed deep learning neural networks with very low overhead, which results in almost linear dependence of the rate of learning by increasing the performance of processors by IBM and Nvidia accelerators. Scalability reaches 95%.
The previous record — 89% — owned by the brainchild of professionals Facebook AI Research. IBM Research development has surpassed the development of Facebook at the time — for training reference model took 50 minutes to 1 hour. In addition, using the new software, IBM Research were able to break the record of Microsoft in terms of accuracy of recognition of images. On the set of the ImageNet-22k of 7.5 million images were the result of 33.8%, while the result of Microsoft equal to 29,8%.
Code IBM Research Distributed Deep Learning available in the distribution PowerAI 4.0 for IBM TensorFlow and Caffe.