Since “Tianji” appeared on the cover of “Nature” for more than a year, this is the third time that Tsinghua University’s brain-inspired computing research has been included in Nature magazine.
On October 14, in the latest issue of “Nature”, a breakthrough in a brain-like computing architecture appeared.
Researchers from Tsinghua University, Beijing National Research Center for Information Science and Technology, and the University of Delaware (University of Delaware) research team in the paper “A system hierarchy for brain-inspired computing” (a brain-inspired computing system hierarchy) The concept of “neuromorphic completenes” was proposed. This research is considered to accelerate the research of brain-like computing and general artificial intelligence.
Currently, there are generally two approaches to developing artificial general intelligence (AGI): neuroscience-oriented and computer science-oriented. Since the two approaches have fundamental differences in formulation and coding, they rely on different and incompatible platforms, hindering the development of AGI.
Paper link: https://www.nature.com/articles/s41586-020-2782-y
The first author of the study is Zhang Youhui, a researcher at the Department of Computer Science, Tsinghua University, and Shi Luping, a professor at Tsinghua University and director of the Center for Brain-Inspired Computing at Tsinghua University, is the corresponding author of the paper.
Taking inspiration from the biological brain, neuromorphic computing provides direction for the next wave of development in computer technology and architecture. Brain-inspired computing is different from traditional computer architectures. The latter is based on Turing’s complete and perfect von Neumann structure. The former does not currently have a generalized system hierarchy or a complete understanding of brain-inspired computing. This affects the compatibility between brain-inspired computing software and hardware, thereby hindering the development efficiency of large brain-inspired computing.
Facing this challenge, researchers at Tsinghua University and other institutions have proposed the concept of “brain-like computing completeness”, which relaxes the requirements for hardware integrity and proposes a corresponding system hierarchy, including Turing-complete software Abstract models and general abstract neuromorphic architectures.
Using this hierarchy, we can describe various programs into a unified representation and translate into equivalent executables on any neuromorphic complete hardware. This means that this system can ensure programming language portability, hardware integrity and compilation feasibility.
To support the execution of different types of programs on various typical hardware platforms, the researchers implemented a series of toolchain software, which in turn demonstrated the advantages of this system architecture.
Brand New System Hierarchy
In this study, the researchers proposed a highly versatile and pervasive brain-inspired computing system hierarchy, which includes three levels: software, hardware, and compilation.
Unlike the traditional computing system hierarchy, for the brain-inspired computing system hierarchy, the software layer refers to neuromorphic application and development frameworks (such as Nengo and PyTorch). Correspondingly, researchers propose to use POG as an intermediate representation of software, and EPG as an intermediate representation of hardware (CFG, control flow graph), and introduce a compilation tool to convert POG into EPG. For the hardware layer, the researchers proposed Abstract Neuromorphic Architecture (ANA), including scheduling unit, processing unit, memory, and interconnection network, as an abstraction of neuromorphic hardware (TrueNorth, SpiNNaker, Tianjic, and Loihi).
Considering the similarity of brain-like computing, the researchers further proposed the concept of “brain-like computing completeness”, introducing approximation equivalence and precise equivalence.
Hierarchical comparison of brain-like computer systems and traditional computer systems.
Software in the figure refers to programming languages or frameworks, and the algorithms or models built on top of them. At this level, the researchers proposed a unified and general software abstraction model – POG (programming operator graph) – to adapt to a variety of brain-inspired algorithms and model designs. POG consists of a unified description method and an event-driven parallel program execution model that integrates storage and processing. It describes what a brain-inspired program is and defines how to execute it. Since POG is Turing complete, it can support multiple applications, programming languages and frameworks to the greatest extent.
The hardware part includes all brain-inspired chips and architecture models. The researchers designed an abstract neuromorphic architecture as a hardware abstraction. It has an EPG (execution primitive graph), which is used as an interface with the upper layer to describe the programs it can execute. EPG has a hybrid “control-flow–dataflow” representation to maximize its adaptability to different hardware, while also conforming to a popular hardware trend – hybrid architecture.
Compilation is an intermediate layer that converts a program into an equivalent form supported by the hardware. In order to improve usability, the researchers proposed a set of basic hardware execution primitives, which are widely supported in mainstream brain-like chips, and proved that the hardware equipped with this set of primitives is neuromorphically complete. In addition, the researchers also took a tool chain software as an example of the compilation layer to demonstrate the feasibility, rationality and superiority of the hierarchical structure.
“This hierarchy avoids tight coupling between hardware and software, ensuring that any brain-inspired program can be represented by a Turing-complete POG and then compiled into an equivalent Executable EPG. We ensured the portability of programming, the integrity of hardware, and the feasibility of compilation, and verified the effect of system design dimension optimization introduced by neuromorphic completeness through experiments. This hierarchical structure also promotes the integration of software and hardware. Collaborative Design.”
Corresponding to the “Turing completeness” concept of today’s conventional computers and the “Von Neumann” architecture, the new brain-inspired computing completeness and the hierarchical structure of the brain-inspired computing system with decoupled software and hardware have proved its feasibility. At the same time, it expands the application scope of the brain-inspired computing system so that it can support general computing.
This research fills the gaps in completeness theory and corresponding system hierarchy for the direction of brain-inspired computing in its infancy, and is conducive to independent mastery of the core technology of new computer systems.
According to a reviewer in the journal Nature, “This is a novel idea and may prove to be a significant development in the field of neuromorphic computing and towards artificial intelligence research.”
The Links: LQ057Q3DC03 TT 250 N 18 KOF 25CN