NVIDIA is a 25-year old company that started up in California’s Silicon Valley to make graphics accelerator chips to improve the quality of visual displays for personal computers. The company developed a graphic processing unit (GPU) that could enhance the central processing unit (CPU) microprocessor in personal computers by offloading compute-intensive graphic modeling tasks. The company became a driver in the computer gaming industry with its graphics accelerator printed-circuit boards that could plug into expansion ports in PC-compatible computers.
A GPU speeds up the formation of graphics artwork on a computer monitor by processing electronic instructions and data through multiple parallel paths at the same time, compared to a CPU than processes a wider variety of computing tasks through slower sequential steps.
For many decades, researchers in the realm of artificial intelligence have been seeking to duplicate human senses and cognition inside computer circuitry. The two most important areas of research were in vision recognition and natural language voice recognition, in order to duplicate the way humans use their two keenest senses to understand and navigate the world about them.
During 2012, Geoffrey Hinton, Alex Krizhensky and Ilya Sutskever from the SuperVision research lab at the University of Toronto in Canada, pioneered the use of neural network software models coupled with GPU processing systems from NVIDIA to create AlexNet. The research team conceived that convolutional manipulation of input variables and output errors could create feedback that would make ongoing changes to the weights of AlexNet’s internal processing algorithms. The result was “deep learning” by the software model that could write its own software to modify how it would process and perceive progressive sets of incoming data.
The team focused its artificial intelligence research on graphics image recognition, in order to group similar digital images of objects together through eight processing layers of visual filtering. Instead of programming the image-recognition algorithms directly, Krizhevsky enabled AlexNet to “learn” the common characteristics of similar graphic objects by feeding the neural net massive amounts of images that it could compare, while the software model adjusted its own internal software processing algorithms with each error.
During an international ImageNet Large Scale Visual Recognition Challenge in 2012, AlexNet outperformed all other existing benchmark tests of image recognition models that had been previously hand-coded by humans, while running on two NVIDIA GPU graphics processing stations. AlexNet neural net software was capable of writing its own internal software more quickly than humans could do directly, achieving a recognition error rate that was ten percentage points less than its nearest competitor.
After the competition, large corporations like Google, Facebook, Amazon, and Microsoft began duplicating the GPU technology used in this image recognition model, to improve each company’s graphic image search and tagging engines. Microsoft’s ResNet was scaled to 152 layers of image filtering, compared to just eight layers in AlexNet.
At Stanford University, AI researcher Andrew Ng applied GPU processing to natural language voice recognition algorithms and succeeded in improving the ability of computer technologies to recognize and caption spoken words. Ng has since become a chief scientist at Baidu, where he has applied this technology to translate spoken Chinese words into readable Chinese characters.
NVIDIA also worked with Tesla Motors, who used its Tegra GPU in early versions of the Model S electric car to develop Autopilot driving features. The Autopilot system could quickly process and filter both images and input data from onboard sensors, that included a forward-facing camera, forward-facing radar and a 360-degree ring of ultrasonic sensors embedded around the body of the electric car.