Mellanox Enables Machine Learning at Baidu

Big Data, , , ,

A soon to be released film, CHAPPiE, tackles the subject matter of artificial intelligence with an experimental robot built and designed to learn and feel.  In the near future, crime is patrolled by an oppressive mechanized police force. When one police droid, CHAPPiE, is stolen and given new programming, he becomes the first robot with the ability to think and feel for himself. He must fight back against forces planning to take him down.


While you may see it as a fictional work of scientific vision, the path toward artificial intelligence isn’t quite so far away.  The building block of artificial intelligence is machine learning.  Deep learning is a new area of machine learning. This new area has the objective of moving machine learning closer to artificial intelligence.  Multiple organizations are investing significant resources into deep learning include Google, Microsoft, Yahoo, Facebook, Twitter and DropBox.




One such company tackling the challenge is Baidu, Inc., a web services company headquartered in Beijing, China.  The company offers many services including a Chinese language-search engine for websites, audio files and images.  The company offers multimedia content including MP3 music, and movies, and is the first company in China to offer wireless access protocol and PDA-based mobile search to users.  Baidu has seen an ever increasing percentage of voice and image searches on its platform.


The company is now investing heavily in deep learning.  They have built a state-of-the art image recognition system, aptly named “Deep Image”.  The key components of the system run on a Mellanox FDR 56G InfiniBand connected GPU-based supercomputer optimized for deep learning algorithms named Minwa (this project goes along with a speech recognition system called “Deep Speech”).Utilizing this system, Baidu claims a 5.98 percent error rate on the ImageNet object classification benchmark which bests the best results of 6.66 percent error rate that won the the 2014 ImageNet challenge.


Published findings about the Deep Image system are listed in “Deep Image: Scaling Up Image Recognition.” According to the authors of the research at Baidu Institute of Deep Learning:


We would like to train very large deep neural networks without worrying about the capacity limitation of a single GPU or even a single machine, and so scaling up is a required condition.  Given the properties of stochastic gradient decent algorithms, it is desired to have very high bandwidth and ultra-low latency interconnects to minimize the communications costs, which is needed for the distributed version of the algorithm.”


Minwa is comprised of 36 server nodes, each with 2 six-core Intel Xeon E5-2620 processor.  Each server contains 4 NVIDIA Tesla K40m GPUs and  Mellanox FDR InfiniBand (56Gb/s) which is a high-performance low-latency interconnect that supports RDMA.


Baidu built Minwa to help overcome the problems associated with the types of algorithms used in deep learning.  This powerful system allows researchers to work with better training data than most deep learning projects.  Baidu used higher resolution images (512×512 pixels) and augmented them with various effects.  Using larger images improves recognition accuracy as well as improves the classification accuracy.  The goal was to let the system take in more features of smaller objects and learn what the object is like without being thrown off by other extraneous factors.


We are excited to work with companies such as Baidu as they made new innovations in the area of deep learning and take the next step toward making the world of CHAPPiE a reality.

About Gilad Shainer

Gilad Shainer has served as Mellanox's Vice President of Marketing since March 2013. Previously, he was Mellanox's Vice President of Marketing Development from March 2012 to March 2013. Gilad joined Mellanox in 2001 as a design engineer and later served in senior marketing management roles between July 2005 and February 2012. He holds several patents in the field of high-speed networking and contributed to the PCI-SIG PCI-X and PCIe specifications. Gilad holds a MSc degree (2001, Cum Laude) and a BSc degree (1998, Cum Laude) in Electrical Engineering from the Technion Institute of Technology in Israel.

Comments are closed.