GPU類比大腦

GPU類比大腦

GPU計算與視覺計算讓GPU成為了21世紀最重要的處理器。 從影像遊戲到科學領域，GPU已應用於各行各業當中。現在，人們開始利用GPU來進行人腦研究了。

「作為長期專案的一部分，研究人員正在著手開發一台電腦，他們希望這台電腦有朝一日能夠類比人腦的視覺化系統。而且此類研究還有諸多實際應用，其中最具魅力的一項便是車輛自動行駛。這一應用能夠利用該項研究的視覺化敏銳度來避開路邊的危險情況以及不可通過的惡劣地形。

Nicolas Pinto是麻省理工學院專攻大腦視覺化研究的一位博士生。據他稱，該專案旨在瞭解如何才能利用先進的繪形處理器(GPU)來打造一台有朝一日能夠類比人腦視覺皮層機能的電腦。」

GPU計算就是運用高階語言與API，利用繪圖處理器(GPU)的大規模並行架構作為計算引擎。CPU的時脈不會再增長了，然而當今消費者卻比以往任何時候都更加需要PC性能的提升。為了提供使用者急需的性能以滿足他們的期望，唯一可行的辦法便是走多核或並行之路。例如，增加更多的核心並用這些核心來分擔繁重的工作負荷。由於電腦圖形的特點，GPU在同時執行多項任務這方面表現得特別出色，因此也就非常適合這一全新的計算環境。GPU為我們帶來的是，利用數以百計的核心以大規模並行方式來解決難題。

研究人員希望打造全智能蝙蝠車

Researchers hope to build autonomous 'Batmobile'

用圖形處理器電腦, 建立/模擬人腦視覺皮層機能

An early prototype of a computer built using graphics processor units in a bid to mimic the brain's visual cortex.

(Credit: Nicolas Pinto/MIT)

If you think the Batmobile is just something from the movies or comic books, researchers at MIT and Harvard University want to change your mind.

As part of a long-term project, the researchers are working on developing a computer that they hope could one day mimic the visualization systems of the human brain. And while there could be many practical applications for such research, one of the sexiest is a potential autonomous vehicle that could use its visualization acumen to navigate roadside dangers or impassable terrain.

According to Nicolas Pinto, a Ph.D. student specializing in brain visualization research at MIT, the idea behind the project is to learn how to build a computer using advanced graphics processing units (GPUs) that could some day simulate the functions of the brain's visual cortex.

Pinto, who is partnered on the project with David Cox at Harvard's Rowland Institute, explained that research indicates the 人腦有相當於20petaflops(20 x 10^15次方)運算能力, 也就是當今第一名超級電腦近20倍的運算能力brain has computing power of at least 20 petaflops, or roughly 20 times the world's most powerful supercomputer. And he added that scientists figure that the 而視覺皮層運算佔用了40%人腦運算, 也就是大約10petaflops(10 x 10^15次方)的運算能力visual cortex makes up at least 40 percent of the brain, meaning that by itself, it has nearly 10 petaflops of computing power.

The brain is massively parallel and its power comes from its ability to perform massively parallel computations, said Pinto, whose supervisor at MIT is James DiCarlo. As a research team, then, he explained that one of the biggest problems he and Cox face in trying to mimic the brain's power is simulating that parallelism.

The best solution, he said, seems to 導入可觀數量的GPU圖形處理器be implementing a significant number of GPUs and applying software tricks to them. And at the moment, Pinto said, 最好的方式就是使用當前最領先技術的NVIDIA GPU the best technology he and his partner have found for the job is state-of-the-art GPUs from Nvidia.

In part, he added, that's because, as unfunded university researchers, he and Cox are running their experiments on Linux computers, and Nvidia's GPUs are the best option for that operating system. Plus, he said, Nvidia is offering the research team a powerful software stack that helps with coding the GPUs.

In general, then, Pinto and Cox are working--with no specific time frame--on enabling their computers to be visually aware technology that, if successful, could make it possible for an autonomous vehicle to do things like recognize an object as a pedestrian and understand what a person is doing.

"My long-term dream is to build a cheap, autonomous, and intelligent robot-car--like a Batmobile--that, if successful, could greatly improve security and health (death rates from road accidents are alarming)," said Pinto, "minimize energy use--by enhancing transport efficiently and facilitating the integration of renewable energies--reduce productive time wasted [in commuting], and optimize urban space for people, not machines...If we can design a machine that sees like humans do, we may not be far from building one that drives as well, allowing for a gradual adoption using the current infrastructure."

To be sure, Pinto and Cox are not expecting to produce their Batmobile any time soon. Conservatively, he suggested that it could be at least five years and possibly as many as 10 before such technology is ready for prime time.

And there's no doubt, of course, that this isn't the only research effort that is attempting to build computers that can simulate the human brain. One notable project is an ongoing program at IBM in which the technology giant and five university partners were awarded a DARPA grant to try to achieve that goal. Already, that team has managed to mimic the brain power of felines, and said last November that they hope to simulate the human brain by about 2019.

'We don't know much about the brain'
To Pinto, one of the biggest hurdles that must be overcome in order to reach the Batmobile vision--or steps on the path towards that--is that researchers still don't know much about the brain. The constraints of understanding neural data mean that trying to build a computer that is like the brain requires, to some extent, simulating evolution. And that means that making progress in the research is slow, laborious and, likely, linear.

Still, Pinto and Cox are already making progress in developing algorithms for computer simulation of vision and they are also helped out by fortuitous developments in carmakers' steady addition of sensor systems to their vehicles. They are already imaging platforms, Pinto said of today's most cutting-edge cars. Now, it will be necessary to add intelligence to the vehicles if we want them to be visually aware.

On June 24, Geek Gestalt will kick off Road Trip 2010. After driving more than 18,000 miles in the Rocky Mountains, the Pacific Northwest, the Southwest and the Southeast over the last four summers, I'll be looking for the best in technology, science, military, nature, aviation and more throughout the American Northeast. If you have a suggestion for someplace to visit, drop me a line. In the meantime, you can follow my preparations for the project on Twitter @GreeterDan and @RoadTrip.