While recent years have seen tremendous progress on tasks like automatic translation and speech recognition, current artificial intelligence systems still fall far short of humans' ability to learn language and to learn from language about the rest of the world. MIT's Language and Intelligence Group, led by Prof. Jacob Andreas, is working towards future in which everyone can interact with software using the languages they already speak.
Prof. Andreas, a member of CSAIL and the department of Electrical Engineering and Computer Science, is the X Consortium assistant professor at MIT in the EECS and in CSAIL. Before joining MIT, he earned his PhD at Berkeley, where he was a member of the Berkeley NLP group and the Berkeley AI Research Lab. He has been the recipient of a Samsung's AI Researcher of the Year award, MIT's Kolokotrones teaching award, and paper awards at NAACL and ICML.
His group's research focuses on three themes:
Building machine learning models that learn language in human-like ways. Today's language processing models are trained on massive datasets of text; humans, by contrast, learn from relatively small amounts of language paired with rich social, contextual, and environmental cues. Prof. Andreas's group has developed new machine learning approaches that can take advantage of these cues, enabling more general and efficient language technologies, especially in languages where large amounts of digitized text is not available.
Enabling the use of language as a general-purpose tool for building intelligent systems. Making it possible to build robots and image classifiers that learn not from millions of labeled examples, but from user-provided descriptions and explanations. Humans learn to bake a cake not by watching millions of videos, but by reading a recipe and watching one or two demonstrations. Prof. Andreas's group has developed new approaches for teaching machine learning models, using language, to synthesize computer programs, write text, and act in the world.
Developing techniques for understanding machine learning models using language. Many machine learning models, especially neural networks, are black boxes---their decision-making processes are difficult to explain or audit for correctness. Prof. Andreas's students have developed techniques that provide natural language descriptions of deep networks and their procedures; these tools have in turn revealed new failure modes and unexpected behaviors in widely used image classification models.
The Language and Intelligence group envisions a future built around training, explaining, and interacting with intelligent systems for users of all the world's languages.
Our goal is to understand the nature of intelligent behavior in the physical world, through the study of human intelligence and the design and implementation of intelligent robots.
Our research encompasses all aspects of NLP, from modeling basic linguistic phenomena to designing practical text processing systems, and developing new machine learning methods.
MIT-IBM Watson AI Lab researchers have developed a universal guide for estimating how large language models will perform based on smaller models in the same family.
Language models follow changing situations using clever arithmetic, instead of sequential tracking. By controlling when these approaches are used, engineers could improve the systems’ capabilities.
