What knowledge makes you intelligent? What are the constructs used by your cognition to understand the world, interpret new experiences, and make thoughtful choices? Defining a framework that articulates the kinds of knowledge that enable understanding and higher cognition for humans or artificial intelligence (AI) will facilitate a structured discussion on ways to effectively materialize these constructs and chart a path to more intelligent machines.
Knowledge constructs that allow an AI system to organize its view of the world, comprehend meaning, and demonstrate understanding of events and tasks will likely be at the center of higher levels of machine intelligence. Machine cognition will expand beyond data to be anchored in knowledge constructs including dimensions such as descriptive knowledge, models of the world dynamics, and provenance, among others.
When studying language, we distinguish between form and meaning: form refers to the symbols — the surface expressions — used to express meaning. Each form has a particular meaning in a particular context, and forms can have different meanings in different contexts. As summarized in an article by Schölkopf, Bengio et al, “the majority of current successes of machine learning boil down to large scale pattern recognition on suitably collected independent and identically distributed (i.i.d.) data.” Systems ingest observable elements such as text characters, vocal signals and image pixels, and establish patterns and stochastic correlations, while yielding outstanding results for recognition-based tasks.
There is growing agreement that algorithms must go beyond surface correlations into meaning and understanding to achieve a higher level of machine intelligence. This categorical shift will enable what is referred to as System 2, 3rdWave, or broad generalization/flexible AI. As I outlined in the core blog ‘The Rise of Cognitive AI’, this next level of machine intelligence requires deep constructs of knowledge that can transform AI from surface correlation to comprehension of the world, representing abstractions, relations, learned experiences, insights, models and other types of structured information.
John Launchbury of DARPA calls out the aspects of AI that will see a transformational improvement in the 3rd wave of AI as abstraction (i.e., creating new meaning) and reasoning (planning and deciding). The 3rd wave itself is characterized by contextual adaptation where systems construct contextual explanatory models for classes of real-world phenomena. The framework presented here offers a perspective on how knowledge constructs will facilitate such a leap.
Two of the knowledge dimensions reflect a view of the world — the descriptive dimension with its conceptual abstractions of what is in the world, and the dynamic models of the real world and its phenomena. Stories add the human capacity to comprehend and communicate complex narratives that build on shared beliefs and mythologies. Context and source attribution as well as value and priorities are meta-knowledge dimensions that provide a condition-based overlay of validity and knowledge-about-knowledge. Finally, concept references are the structural underpinning, binding across dimensions, modalities and references. Together, these six dimensions of knowledge could bring additional depth beyond correlation of events by assuming underlying concepts that are persistent and can explain and predict past and future events, allow for planning and intervention, and consider counterfactual realities — hence the use of the term ‘deep knowledge.’
Articulating and characterizing the kinds of knowledge constructs necessary for machine intelligence can contribute to identifying the best way to implement them to bring about the next level of machine intelligence. The goal of this blog is to establish the fundamental classes of knowledge constructs deemed relevant for the development of the next level of AI cognitive capabilities.