Novel AI approach integrates data-driven modeling with theoretical reasoning to construct sophisticated deep-learning architectures
CHALLENGE
There are generally two kinds of artificial intelligence (AI) applications available today: those based on knowledge, which are good at reasoning but are very narrow in scope, and those based on data, which are good at learning but have very shallow inference capability and often require very large amounts of data to produce reasonable results. Scientists and information analysts need AI that can effectively do both.
SOLUTION
SRI International’s Artificial Intelligence Center (AIC) has developed an “AI engine” based on Deep Adaptive Semantic Logic (DASL)—a new, deeply principled approach that integrates bottom-up data-driven modeling with top-down theoretical reasoning. DASL’s innovative machine learning and reasoning system facilitates “discussions” between scientists and data, which SRI researchers anticipate will ultimately accelerate data-driven knowledge.
Supporting information
Scientific and analytic knowledge are often most easily expressed in declarative, logical forms. However, this is currently difficult to do in machine learning architectures.
Machine reasoning (ML) techniques impose stringent requirements on the precision and usage of terminology, requiring an explosion in the number of assertions and a never-ending effort to fill in the missing pieces for each next experiment. In contrast, machine learning techniques either ignore subject matter expertise altogether or require labor-intensive encoding of small amounts of expertise into custom statistical models that do very little outside of a narrow scope.
DASL integrates qualitative and quantitative AI reasoning—both of which are essential for complicated efforts such as modeling dynamical systems like biological cells, which are far too complex for purely quantitative treatment.
How it works
Model theory underlies the formal inference behind current formal systems, expert systems and machine reasoning systems. These systems rely on subject matter expertise expressed in formal logic -frequently first-order logic- rather than familiar language to build up set-theoretic interpretations of the formal language. As a result, these systems can recognize when two expressions match exactly, but only when they match exactly. They are not able to make approximate matches, which severely limits usefulness to scientists and others who want to leverage AI for large, complex projects.
The DASL AI engine replaces the set-theoretic semantics of model theory with semantics based on real-valued functions, or how experts actually express ideas. It fully integrates knowledge representation and reasoning (KRR) with statistical machine learning (SML) to transform reasoning in terms of a logical language into reasoning in terms of a soft semantic representation.
DASL applies deep learning techniques so scientists and analysts can provide the model with complex background knowledge and hypotheses in familiar but formal language. The DASL model generalizes the available data in ways that best fit to the theory, and the user continues the “discussion” by querying the model for any properties expressed in the theory.
Using both empirical data and the logical rules asserted as expert knowledge to serve as training data, DASL creates semantic representations that approximate the empirical data and asserted knowledge as closely as possible.
DASL’s automated logical reasoning system gives people unprecedented ability to engage with and train AI using the semi-formal technical language they use among themselves, helping them maximize productivity and more effectively accelerate and advance data-driven knowledge.
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001118C0023. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of DARPA.
