Neuro-Symbolic Architectures for Incorporating Domain Knowledge based Inductive Bias
Abstract
The ability of statistical neural systems to automatically infer vast "sub-formal knowledge" from large data corpora has contributed to their considerable empirical successes. But they require large amount of training data, and are known to suffer from a lack of robustness and reliability, which makes them less suitable to high-stakes application domains. Classical AI systems on the other hand rely solely on formal domain knowledge, by which we mean the often structured, even formal specifications of the domain or task and how the target outputs relates to the inputs, and in general, how various typically symbolic concepts underlying the domain relateto each other. These classical AI systems are much more reliable since they rely on reasoning with such formal knowledge, rather than on ill-understood patterns from experience and sub-formal knowledge as the statistical AI systems essentially do. Indeed, human advances in science owe a lot to iteratively building on formal domain knowledge of the kind inherent to such classical AI systems. The caveat with such systems however is that such formal knowledge is necessarily incomplete, which results in highly underpowered models. Indeed, practitioners in complex domains often go beyond solely formal theories, or solely their instincts drawn from experience, and rely on both formal and sub-formal knowledge. So a critical question is how do we devise AI systems that can combine the rich structured body of domain knowledge, together with the inductive inference techniques from second-wave statistical AI that could capture sub-formal information from data corpora? To do so however we are faced with a levels mismatch: the conceptual level of formal domain knowledge is very different from the level of raw features or model parameters for statistical neural models. It is thus not clear how to constrain statistical neural models using such structured and formal domain knowledge. We thus need neuro-symbolic architectural innovations that go beyond standard statistical neural models, and then constrain these models via structured domain-knowledge-based inductive bias.In this proposal, we propose to make headway towards such neuro-symbolic architectures, for incorporating domain knowledge based inductive bias, in three progressive stages. In the first stage, we build a symbolic module on top of statistical neural models, as well as a reasoning module on top of the symbolic features. A critical ingredient in ensuring we recover the right symbolic features will be through constraining them via domain knowledge. In the second stage, we build an integrated neuro-symbolic architecture that combines the neural, symbolic and reasoning modules within one single model, that moreover is capable ofinferring causally relevant symbolic features. In the third stage, we move from concepts as scalars to concepts as functions, whichallows us to capture more complex forms of domain knowledge ranging over transformations and analogies, as well as conceptual primitives. We instantiate our neuro-symbolic innovations across three broad domains of vision, language, and robotics. Our preliminary results have already demonstrated success of such a neuro-symbolic approach, beating state-of-the-art by a significant margin, and placing highly in key competitions: CVPR 2020 Object Goal Navigation, ALFRED instruction navigation, and ALFRED 2022 object rearrangement challenges. Approved for Public Release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 12, 2023
- Source ID
- N000142312368
Entities
People
- Pradeep Ravikumar
Organizations
- Carnegie Mellon University
- Office of Naval Research
- United States Navy