让自行车“自行”，这款登上《自然》封面的“天机芯”有何“天机”?

清华大学的校园里，有一辆“神奇”的自行车：它不仅真的能够“自行”、主动避障，还会响应简单的声音命令、对前方行人进行探测和跟踪。

天机芯片单片（左）和5x5阵列扩展板（右）

驱动这辆自行车的，是一块指甲盖大小的、世界上首款面向通用人工智能的异构融合类脑计算芯片——“天机芯”(Tianjic)。它由清华大学类脑计算研究中心施路平教授团队打造。

This is not a scene from a sci-fi movie, but a feat backed by an artificial intelligence (AI) chip developed by a group of researchers from China's Tsinghua University led by Professor Shi Luping. The bicycle, installed with the AI chip dubbed "Tianjic", performs self-driving, automatically responding to different situations.

It nimbly avoids obstacles in the way when following a human. Even when the human runs in a slightly S-curved path, the bicycle tracks exactly in the S-shape. It is also able to recognize voice instructions, completing commands like "turn left," "go straight," and "speed up."

8月1日，基于该成果的论文“面向通用人工智能的异构天机芯片架构”(Towards artificial general intelligence with hybrid Tianjic chip architecture)，作为封面文章登上国际知名学术期刊《自然》(Nature)，实现中国在芯片和人工智能两大领域该期刊论文零的突破。

“天机芯”有何“天机”？

——计算机科学+神经科学推进AGI发展

人工智能奶茶冲制机器人

说到人工智能，相信大家都不陌生。从打败棋手柯洁的AlphaGo，到公司里的人脸识别考勤机、论坛上的实时翻译器，这些都没有走出“专用人工智能”的范畴。

与专用人工智能相对应的，是“通用人工智能”(Artificial General Intelligence)。人的大脑，就是最典型的通用智能系统，能处理视觉、听觉、学习、推理等多种任务，具备举一反三、融会贯通的能力。通用人工智能，是人工智能学界一直在努力的方向。

据《自然》报道，发展通用人工智能有两个主要方向：一种是神经科学导向，试图构建与大脑极为相似的回路；另一种是计算机科学导向，让计算机运行机器学习算法。

那么，是否存在某个平台，可以融合这两个方向呢？

“天机芯”做到了。

天机芯片有多个高度可重构的功能性核，可以同时支持机器学习算法和类脑电路，它由156个统一功能核心(FCore)组成，包含约40000个神经元和1000万个突触，采用28纳米工艺制程，面积为3.8×3.8平方毫米。

天机异构融合类脑计算架构

Researchers installed about 40,000 neurons and 10 million synapses in the fingernail-sized chip of 3.8x3.8 square millimeters. It's the first-ever chip that integrates computer-science-oriented and neuroscience-oriented approaches, two general ways to develop artificial general intelligence (AGI).

Usually, the two approaches, both representing some of the patterns in which the human brain processes information, rely on different platforms, which are incompatible due to their fundamental differences in the formulations and coding schemes. But a combination of the two is generally recognized as the best solution.

The research team proposed a multi-core architecture to accommodate the two approaches and developed the Tianjic chip on this basis, "can not only accommodate computer-science-based machine-learning algorithms, but also easily implement brain-inspired circuits and several coding schemes," read the research paper abstract.

自动行驶自行车演示平台

为了验证“天机芯”整合多模态信息和做出迅速决策的能力，研究团队开发了一款“无人自行车”系统。“自行车系统的语音识别、自主决策、视觉追踪功能运用了模拟大脑的模型，而目标探测、运动控制和躲避障碍功能运用了机器学习算法模型。”论文作者之一、清华大学类脑计算研究中心首位博士生邓磊介绍道。

自行车上配备了天机芯片、IMU传感器、摄像头、转向电机、驱动电机、调速电机和电池

在实验中，“Tianjic动力自行车”顺利完成了所有指定的任务，标志着AGI发展的巨大飞跃。

研究小组还指出，“通过随机将新变量实时引入环境中，可以产生高时空复杂性，例如不同的道路条件、噪声、天气因素，多种语言，更多人群对象等等。通过探索允许适应这些环境变化的解决方案，可以检查对AGI至关重要的问题——例如概括、稳健性和自主学习。

Functions of the unmanned bicycle system like voice recognition, autonomous decision-making, and visual tracking use the brain-inspired model, while the target detection, motion control, and obstacle avoidance use the machine-learning algorithm model, researcher Deng Lei told Thepaper.cn.

The paper entitled "Towards artificial general intelligence with hybrid Tianjic chip architecture" was published in the Nature journal as a cover story on Wednesday. The chip can be applied in automatic pilot systems and intelligent robots, said Deng. "Our study is expected to stimulate AGI development by paving the way to more generalized hardware platforms," the authors said.

“天机芯”的前世今生

——多学科交叉共融

论文通讯作者：清华大学精密仪器系教授、类脑计算中心主任施路平

据悉，2015年，施路平团队设计出第一代“天机芯”， 2017年第二代“天机芯”问世。相比于当前世界先进的、IBM的TrueNorth 芯片，第二代“天机芯”密度提升20%，速度提高至少10倍，带宽提高至少100倍，功能性、灵活性和扩展性更好。

“发展类脑计算芯片是个世界性难题。它真正的挑战不是科学，也不是技术，而是多学科融合。”施路平介绍。

2014年，清华大学依托精密仪器系成立了联合了计算机、电子、微电子、自动化、材料和医学院七个相关院系的类脑计算研究中心，学校从一开始就为中心制定了多学科融合引领新方向、产学研结合的发展战略。

此次论文共同作者来自六个院系，其中共同第一作者就来自精仪、生物工程、自动化和计算机四个系。

施路平强调，此次发表在《自然》的论文是一个非常初步的研究，通用人工智能是一项非常具有挑战性的工作，目前还处于起步阶段。他和团队的研究愿景是——

“发展类脑计算，

支撑通用人工智能，

赋能各行各业”。

据悉，团队已经开始研究下一代芯片，有望在明年初完成研发。研究团队还透露，未来将把“天机芯”部署在自动驾驶汽车和智能机器人领域。

The second-generation Tianjic was developed from the 2015 version. It features high-speed, high-performance and low-power consumption. Compared with the current advanced TrueNorth chip from IBM, Tianjic has more comprehensive functions, greater flexibility, and better scalability. Its density is increased by 20 percent, speed by at least 10 times, and bandwidth by at least 100 times.

AGI refers to the machine intelligence that can understand or learn any intellectual task that humans can accomplish. Some call it "strong AI." It's a vision that hasn't yet been realized. Some AI scholars believe the concept of AGI is not serious and is basically impossible to achieve in practice. Others are bullish about its development, saying it has the potential to shape the trajectory of humanity.

Though a difficult research subject, Professor Shi said at a press conference that he believes that AGI will happen, and is an inevitable trend from the perspective of future development.

“对于人工智能，中国做的可比说的多得多。”

事实上，在这次“零的突破”之前，中国在芯片和人工智能领域的多项成果早已引起国内外广泛关注：

今年4月10日，华为推出基于昇腾AI芯片的Atlas人工智能计算平台；6月20日，寒武纪宣布推出新一代云端人工智能芯片“思元270”；6月21日，华为发布人工智能手机芯片“麒麟810”；7月3日，百度推出远场语音交互芯片“鸿鹄”；7月25日，阿里巴巴发布基于RISC-V开源设计的处理器“玄铁910”……
每一次前进的脚步，都在用事实证明中国在发展通用人工智能领域的思路和努力。

正如美国市场观察网(Market Watch.com)此前援引美专家评论：“对于人工智能，中国做的可比说的多得多。”

“天机芯”的惊艳亮相，再次引起外媒关注：