The Power of Natural Language in AI Development
Using Natural Language to Build AI Systems
Master’s of engineering students Irene Terpstra and Rujul Gandhi are working with mentors in the MIT-IBM Watson AI Lab to use the power of natural language to build AI systems. They are working on developing AI algorithms that can assist in chip design and convert natural language instructions into a machine-friendly form for robots.
Terpstra’s AI Chip Design System
Terpstra’s team is creating an AI system that can iterate on different designs using large language models like ChatGPT, Llama 2, and Bard. They are using an open-source circuit simulator language called NGspice and a reinforcement learning algorithm. The goal is to combine the reasoning powers and the knowledge base of large language models with the optimization power of reinforcement learning algorithms to have the AI design the chips themselves.
Gandhi’s Natural Language Instruction Conversion System
Gandhi’s system converts natural language instructions into a machine-friendly form for robots. They are leveraging the linguistic structure encoded by the pre-trained encoder-decoder model T5 and a dataset of annotated, basic English commands for performing certain tasks to achieve this. This approach allows the system to understand logical dependencies expressed in English and also allows for flexibility in the way instructions are phrased.
Gandhi is also working on developing speech models for low-resource languages, allowing people to interact with software and devices in their native language or dialect and improving voice assistants and translation applications.