The past few decades have seen explosive growth in development and training of AI systems, which are now embodied in digital computing processes spanning several key industries. One area that has benefited from AI, and specifically Machine Learning (ML) techniques and statistical methods, is the area of Human Language Technology (HLT). However, ML suffers from a need of large amounts of annotated data for training in order to achieve the required accuracy for the various applications. ML technology is also brittle, incapable of dealing with new data sources, topics, media, and vocabularies. These weaknesses of ML as applied to natural language are due to exclusive reliance on the statistical aspects of language, with no regard for its meaning.
DARPA’s GAILA program is focused on the development of AI that can achieve childlike language acquisition and understanding from visual concepts.
Children acquire language based on their perceptions of aural and visual information about the world around them. The process of observing moving images and aligning them with auditory stimuli allows them to associate a sequence of sounds (a word) with some aspect of the concrete or abstract elements of the world that the word represents. Children learn to decipher which aspects of an observed scenario relate to the different words in the message from a tiny fraction of the examples that ML systems require. Sequencing information, variations of word forms, and other additional information help children make ever finer classifications of the concepts that they learn, about events or actions (often expressed by verbs), the objects or entities that participate in those events (typically nouns), and the attributes and relations of those entities and events (adjectives and prepositional phrases).
The Grounded Artificial Intelligence Language Acquisition (GAILA) project will focus on enabling computers to acquire language in a manner similar to the way children do. The process of observing moving images and aligning them with auditory stimuli allows them to associate a sequence of sounds (a word) with some aspect of the concrete or abstract elements of the world that the word represents.
The Defense Advanced Research Projects Agency (DARPA) is issuing an Artificial Intelligence Exploration (AIE) opportunity, inviting submissions of innovative basic research concepts in the technical domain of human language technology, cognitive science, and language acquisition.
The primary goal of this effort is to research and develop a model for grounded language acquisition and an automated language acquisition prototype that learns to understand English text and speech, for the purpose of making the information more useable by automated analytics.
GE, Siena College Scientists to Demonstrate AI Agent that Enables Machines to Acquire Language in a Classroom Style
Could industrial machines become MacGyver-like in learning and acting on the fly to solve complex problems? One of the keys will be demonstrating AI that can meaningfully learn from visual and contextual cues. This is the focus of a new research project by scientists from GE and Siena College scientists through DARPA’s Grounded Artificial Intelligence Language Acquisition (GAILA) program.
GE Research and Siena have been awarded $500,000 for Phase I of the project, which will be completed over nine months. The goal of Phase I is to demonstrate various forms of language acquisition based on visual grounding methods. Upon the successful completion of Phase I, DARPA will perform a down selection for Phase II of the program.
“Today, 99.9% of AI is based on millions of known statistical datapoints with minimal interpretation beyond what the data says,” said Peter Tu, Chief Scientist for Artificial Intelligence at GE Research who is leading the DARPA project. “Through this project, we’re aiming to create an AI agent that can learn the meaning of things, not just the statistics of things. This would unleash a whole new realm of capabilities for machines across multiple industry sectors.”
Tu added, “Today, AI integrated into wind farm operations can improve annualized energy output (AEP) by a few percentage points, which is all based on known datasets. But imagine if the turbines on these wind farms could observe and modify their operations based on entirely new situations they haven’t yet observed. If the AI was able to observe and derive meaningful actions in real-time, the energy output would be much greater.”
Tu explained that the AI agent under development is being designed to learn like a child learns when growing up. “Children pick up things from what they see and hear and spend time playing and experimenting,” Tu says. “They can acquire a familiarity and context that machines can’t replicate today. We’re hoping to change that.”
GE Research has partnered with researchers from Siena College Institute of Artificial Intelligence (SCIAI), which has a strong natural language processing program and develops research in a broad range of Artificial Intelligence areas in partnership with private industry and others.
Dr. Sharon Gower Small, director of SCIAI who is leading the effort with Dr. Ting Liu from Siena, said, “Traditionally, many aspects of Natural Language Processing have relied heavily on tools that were built on large amounts of manually annotated text. The challenge for the team at the Siena College Institute of Artificial Intelligence (SCIAI) is to develop novel techniques to acquire knowledge by combining the image analysis results from our GE partner and linguistic features generated from unsupervised machine learning techniques and vastly smaller amounts of data. These techniques will include a dialogue model that will interact with human experts to confirm/correct the learned knowledge, which mimics how children learn from their parents, teachers, and peers”,
References and Resources also include: