MIT Aims to Ease AI Programming
A new AI programming language seeks to ease the process of writing inference algorithms and other predictive models without the hassle of grinding out complicated equations and code.
Among the goals of the probabilistic programming system dubbed “Gen” is making it easier for coding novices to write models and algorithms for broader AI applications such as computer vision and robotics.
Presented in a paper by MIT researchers at a recent technical conference, the “general purpose” language is based on the Julia programming language aimed at high-end numerical analysis and computational science. The 1.0 version of Julia was released last year.
The researchers reported incorporating several modeling languages into Julia, which was also developed at MIT. In pursuit of an all-purpose programming scheme, each language was geared toward a different AI modeling approach.
Among the ways Gen delivers modeling flexibility was the inclusion of an inference library that provides novice users with a higher level of abstraction for implementing inference algorithms. The researchers claim Gen outperforms other AI programming approaches for applications like object tracking and inferring the structure of time-series data.
Gen also automates parts of the coding process to enable developers with less math experience the ability to build and test prototype AI systems. In one demonstration, the researchers used Gen to simplify data analytics tasks by automatically generating statistical models used to predict patterns in large data sets.
The researchers noted their work builds on open source libraries APIs like Google’s (NASDAQ: GOOGL) TensorFlow intended to spur development of machine learning models. Those automation tools are focused on higher-end deep learning models. At the other end of the spectrum, AI techniques such as statistical models and simulation engines are more flexible but less efficient.
By contrast, Gen is intended to combine automation, flexibility and speed into a general-purpose language to do for inference algorithms what TensorFlow did for deep learning. The hope is that Gen will accelerate development in disciplines ranging from cognitive science and natural language processing as well as robotics and computer vision. Hence, the MIT researchers noted, Gen seeks to achieve “both modeling expressiveness and inference efficiency, producing fast and accurate inference results in diverse problem domains.”
MIT said corporate partners such as Intel (NASDAQ: INTC) are collaborating on efforts such as 3D pose estimation, a complex inference task for computer vision systems that can be used in autonomous vehicles and augmented reality systems.
“Gen is the first system that’s flexible, automated and efficient enough to cover those very different types of examples in computer vision and data science” said Vikash Mansinghka, a researcher in MIT’s Department of Brain and Cognitive Sciences.
Some of the research funding for the AI programming effort was provided by the Defense Advanced Research Projects Agency.