Using Language to Gain Insight into Machines

The power of natural language makes it an excellent data source for training AI systems. Irene Terpstra and Rujul Gandhi, two master’s of engineering students at MIT, are working with mentors in the MIT-IBM Watson AI Lab to harness the potential of natural language in developing AI systems.

Terpstra, along with her mentors Anantha Chandrakasan and Xin Zhang, is focusing on using natural language to assist in chip design. They are developing an AI algorithm that can analyze language models and integrate them into the chip design process. The ultimate goal is to design chips using the combined capabilities of language models and reinforcement learning algorithms.

Gandhi, on the other hand, is working on improving communication between humans and machines. She is building a parser that converts natural language instructions into a machine-friendly form. By leveraging pre-trained encoder-decoder model T5 and a dataset of annotated English commands, Gandhi’s system can break down instructions into sub-tasks and understand logical dependencies expressed in English.

In addition to her work on communication, Gandhi is also focused on language processing for low-resource languages. She is developing speech models that can recognize word boundaries and infer word sequences in languages with limited resources. This research has applications in improving voice assistants, translation, and interpretation.