Many AI experts believe humanlike artificial general intelligence (AGI) is but a far-fetched dream, while others find their inspiration in the quest for AGI. Speaking at last November’s AI Frontiers Conference, OpenAI Founder and Research Director Ilya Sutskever said “We (OpenAI) have reviewed progress in the field over the past few years. Our conclusion is near-term AGI should be taken as a serious possibility.”
Today, respected scientific journal Nature boosted the case for AGI with a cover story on a new research paper, Towards artificial general intelligence with hybrid Tianjic chip architecture, which aims to stimulate AGI development by adopting generalized hardware platforms.
Typically, researchers have taken one of two paths in pursuit of AGI — proceeding either via computer science or via neuroscience. Each approach however requires its own unique and incompatible platforms, and this has stalled overarching AGI research and development. With an eye on closing that gap, researchers from Tsinghua University, Beijing Lynxi Technology, Beijing Normal University, Singapore University of Technology and Design (SUTD) and University of California Santa Barbara have introduced the Tianjic chip. The revolutionary chip can adopt various core architectures, reconfigurable building blocks and so on, to accommodate both computer-science-based machine-learning algorithms and neuroscience-oriented schemes such as brain-inspired circuits.
A key innovation from the research team is Tianjic’s unified function core (FCore) which combines essential building blocks for both artificial neural networks and biologically networks — axon, synapse, dendrite and soma blocks. The 28-nm chip consists of 156 FCores, containing approximately 40,000 neurons and 10 million synapses in an area of 3.8×3.8 mm2.
Tianjic delivers an internal memory bandwidth of more than 610 gigabytes (GB) per second, and a peak performance of 1.28 tera operations per second (TOPS) per watt for running artificial neural networks. In the biologically-inspired spiking neural network mode, Tianjic achieves a peak performance of about 650 giga synaptic operations per second (GSOPS) per watt. The research team also showcased the superior performance of Tianjic compared to GPU, where the new chip achieves 1.6 – 100 times better throughput and 12 – 10000 times better power efficiency.
The research team designed a self-driving bicycle experiment to evaluate the chip’s capability for integrating multimodal information and making prompt decisions. Equipped with the Tianjic chip and IMU sensor, a camera, steering motor, driving motor, speed motor and battery, the bicycle was tasked with performing functions such as real-time object detection, tracking, voice-command recognition, riding over a speed bump, obstacle avoidance, balance control and decision making.
The research team developed a variety of neural networks (CNN, CANN, SNN and MLP networks) to enable each task. The models were pretrained and programmed onto the Tianjic chip, which can process the models in parallel and enable seamless on-chip communication across different models.
In experiments, the Tianjic-powered bicycle smoothly performed all assigned tasks, signaling a huge leap towards the acceleration of AGI development.
The research team also noted that “high spatiotemporal complexity can be generated by randomly introducing new variables into the environment in real time, such as different road conditions, noises, weather factors, multiple languages, more people and so on. By exploring solutions that allow adaptation to these environmental changes, issues critical to AGI — such as generalization, robustness and autonomous learning — can be examined.”
The research team told Chinese media they expect the Tianjic chip to be deployed in autonomous vehicles and smart robots. They have already started research on the next-generation chips and expect to close the R&D stage early next year.
Further information can be found in the paper Towards artificial general intelligence with hybrid Tianjic chip architecture.
Journalist: Tony Peng & Fangyu Cai | Editor: Michael Sarazen