Semicon Korea 2019 Keynote Presentation
Myung-Hee Na, Distinguished Engineer at IBM Research
"AI will evolve faster through heterogeneous systems, approximate computing, analog in-memory computing, and quantum computing." Myung-Hee Na, Distinguished Engineer of IBM Research, said as a keynote speaker at Semicon Korea 2019 held at COEX, Seoul on the 23rd.
Engineer Na who presented on the theme of "The Era of AI Hardware" emphasized that we have grown past the Narrow AI age of performing simple tasks, and entered the Broad AI age. She also emphasized that problems should be solved through a combination of hardware and software. "AI is not a technology but a tool to help human beings themselves. "We will be able to increase human abilities through Broad AI,” she said.
Broad AI can actively work with multiple data. It would take more than four days for humans to edit highlights of a golf game over the course of a few days, but the AI can complete the tasks in a day or two. This is because the highlight elements are automatically determined by factors such as the movement of the athlete and the voice tone of the commentator and the viewer. Engineer Na emphasized that the hardware must be changed in order to implement this.
"I cannot go over to General AI at once, which is the next stage of Broad AI. But at the same time, we cannot just wait for the release of the next generation hardware. We need to introduce a heterogeneous system that includes a central processing unit (CPU), memory, networking, and an AI accelerator to increase hardware performance," she explained. In addition, because the graphics accelerator (GPU) used as an AI accelerator is 2.5 times more difficult to increase performance per year, the power consumption will increase rapidly. Accordingly, she also introduced approximate computing by considering efficiency.
Approximate computing imitates the human brain’s working principles. It is a principle to understand what a picture is, even if it is not perfect. Similarly, there is a word superiority effect. When a human recognizes a word, it refers to accepting each character as a collective image rather than as a set.
"Approximate computing starts with the fact that the accuracy of the AI does not need to be perfect," Engineer Na said. "It is meaningful in that the accuracy as necessary is kept even if you lower the number of bits through multiple learning,” she added. To implement approximate computing, algorithms, program models, and communication methods should be considered. This means that you need to find a way to work smoothly in real systems rather than worrying about handling too much data.
For the ultimate AI evolution, quantum computing should be mobilized. Quantum computing can process bits, the smallest digital unit, as often as desired. It measures the state of a quantum and places 0 and 1 at the desired position. "Deep quantum learning has already been proven effective and will be available in the near future. With analog memory computing, which processes and analyzes data much faster than before, it is a way to accelerate the evolution of AI,” she presented.
"It is important to realize that AI can enhance people's capabilities, not make them unnecessary," she said. "As semiconductors have evolved their industry by lowering their performance and cost, there is a new opportunity for AI that has been difficult to achieve in the past," she added.