Intel's machine programming code similarity system enables everyone to create software

Recently, Intel released a new machine programming system developed jointly with MIT and Georgia Institute of Technology. The system is called Machine Inference Code Similarity System (MISIM), which is an automated engine for detecting software intent by identifying code structure and analyzing syntactic differences with other functionally similar codes.

Justin Gottschlich, founder, chief scientist and director of Intel Machine Programming Institute, said: "The ultimate goal of Intel Machine Programming is to enable everyone to create software. When this goal is fully realized, everyone can create software by expressing their design intent to the machine in the way they are best at, such as code, natural language or other means. This is a bold goal that requires a lot of effort, and MISIM is a solid step for us."

With the rise of heterogeneous computing, hardware and software systems are becoming more and more complex, and professional cross-architecture programmers are scarce, which has led to an increasingly prominent demand for new development methods in the industry. The term "machine programming" was first proposed in the paper "Three Pillars of Machine Programming" jointly released by Intel Research and MIT, which aims to improve development efficiency through automated tools. Among the various emerging machine programming tools, code similarity is a key technology that has the potential to accurately and efficiently automate the software development process, thereby meeting cross-architecture programming needs.

However, building accurate code similarity systems is a thorny problem in itself. These systems need to identify whether two code segments have similar characteristics and need to achieve similar goals. This is very difficult to achieve when only the source code is provided. But MISIM, jointly developed by Intel, MIT and Georgia Institute of Technology, can accurately identify whether two code segments run similar calculations, even if the two code segments use different data structures and algorithms. As Gottschlich explained, "This is an important step in realizing the great vision of machine programming."

The most critical difference between MISIM and existing code similarity systems is that it has an innovative context-aware semantic structure (CASS) that can extract the true purpose of the code. Unlike other existing methods, CASS can be configured to a specific context environment to capture information describing the code at a higher level. Therefore, CASS can provide more precise insights such as "what the code can achieve" rather than "how to achieve it". In addition, MISIM can do all this without the use of a compiler (a program that converts human-readable source code into computer-executable machine code). Therefore, compared with existing systems, MISIM has more advantages, including the ability to execute on incomplete code segments that developers are still writing, which is of great practical significance for application scenarios such as recommendation systems and automatic bug fixing.

Once the code structure is integrated into CASS, multiple neural network systems will give similarity scores to the code segments according to the design goals. In other words, if two code segments look very different in structure but perform the same function, these neural network systems will give a score of "highly similar".

By combining these principles into a unified system, researchers from Intel, MIT, and Georgia Tech found that MISIM could identify similar code segments 40 times more accurately than previous state-of-the-art systems.

Intel will continue to expand the capabilities of MISIM, which has now moved from the research stage to the demonstration stage, with the goal of creating a code recommendation engine to assist software developers in programming on Intel's various heterogeneous architectures. Such a system can identify the intent behind a developer's simple algorithm input and recommend candidate codes with similar semantics but better performance.

Intel Machine Programming Institute is also working with Intel's software department to study how to integrate MISIM into daily development work. Gottschlich, who is also an adjunct assistant professor at the University of Pennsylvania, hopes to help the software department and the entire Intel company improve productivity and eliminate monotonous work such as fixing bugs in programming. Gottschlich said: "If the machine can automatically check and fix bugs, I think most developers will be very willing to let it take on this task, at least I will."