Baidu, the Chinese search engine company, said that it's into Artificial Intelligence (AI), and has built the world's most-accurate computer vision system. Called the Deep Image which runs on a series of supercomputers optimized for deep learning, Baidu claims it to be comparable to Google's.
While humans achieved an estimated error rate of 5.1 percent on the ImageNet object classification benchmark, Baidu claims Deep Image to have a 5.97 percent error rate. Google benchmarked with a 6.66 percent error rate.
What powers Deep Image is its supercomputers called Minwa, which is built to home the system. In order to achieve such capability, Baidu's deep learning researchers have used GPUs to handle computation intensities, replacing the intensive use of CPUs.
Deep Image research papers cites a study showing that 12 GPUs in a 3-machine cluster can rival the performance of a 1000-node CPU cluster that is powering Google Brain, a project that Baidu's Chief Scientist Andrew Ng has worked on.
To break it down, Minwa is a system that comprised of 36 server nodes with each having 2 six-core Intel Xeon E5-2620 processors. Each of its servers contains 4 Nvidia Tesla K40m GPUs, one FDR InfiniBand (56Gb/s) with RDMA supported. With each GPU possessing 12GB memory, the peak performance of each is 4.29TFlops.
In total, Minwa has 6.9TB host memory, 1.7TB device memory, and about 0.6 PetaFlops theoretical single precision peak performance.
With that sheer performance, Minwa is built to overcome problems that are usually associated with Deep Image algorithms.
"Given the properties of stochastic gradient decent algorithms, it is desired to have very high bandwidth and ultra low latency interconnects to minimize the communication costs, which is needed for the distributed version of the algorithm," the authors wrote.
Such powerful system allows researchers to work on different training data more efficiently and arguably, better than most other deep learning projects.
As an example, beside using 256x256-pixel images that are commonly used to teach AIs, Baidu is able to use higher resolution images (512x512 pixels), and augmented them with various effects such as color-casting, vignetting and lens distortion.
The goal for this is to let the system take in more features of smaller objects and to learn what objects really look like.
Beside AI to aid image-recognition Baidu is also investing heavily on deep learning called Deep Speech. The speech-recognition system was first introduced in December 2014.
Racing Towards The Future Of AI
Baidu is one of the company that aims for AI. Especially on the web, it's going on the same direction as Google which still holds the ImageNet record in the actual competition. Google is the company that is most associated with deep learning with features implemented deep inside its Search and other of its services.
Both Facebook and Microsoft are also having researchers in deep learning that are trying to create cutting-edge researches while releasing products that use those researches.
As other companies such as Yahoo and Twitter are also having deep learning and computer vision teams at their disposals,
While the notable IBM has saw the potential in AI long before most companies did at its time, current ones like Yahoo! and Twitter are also investing on deep learning by having their own computer vision teams at their disposal. With many aiming for a single purpose, more implementations of AI should be available in the close future.
