Invited keynote talk:


   A neural architecture for reasoning, decision-making, and
        episodic memory: Taking a cue from the brain
                                  Lokendra Shastri
                       International Computer Science Institute
                              Berkeley, CA 94704, USA
                            www.icsi.berkeley.edu/~shastri


ABSTRACT

The human brain is capable of encoding a large and diverse body of common sense
knowledge and performing a wide range of inferences rapidly; to wit our ability to
understand language in real-time. Consider the simple narrative: “John fell in the
hallway. Tom had cleaned it. He got hurt.” Upon reading this narrative most of us would
effortlessly infer that Tom had cleaned the hallway, therefore, the hallway floor was wet;
John slipped and fell because the floor was wet; John got hurt because of the fall. Such
“bridging” inferences help in establishing referential and causal coherence and are
essential for language understanding. Since we understand language at the rate of several
hundred words per minute, it follows that we draw such inferences within hundreds of
milliseconds.
The ability to reason effectively with a large body of knowledge is critical not only for
language understanding, but also for decision-making, problem solving, and planning.
Consequently, the development of efficient, large-scale inference systems has been a
central goal of artificial intelligence. Although notable progress has been made toward
this goal, it remains far from being achieved.
Given that the human brain is the only extant system capable of supporting a broad range
of efficient, large-scale reasoning, it seems appropriate to assume that understanding how
the brain represents knowledge and performs inferences might lead to critical insights
into the structure and design of large-scale inference systems.
In this talk I will review the current state of a long-term research project on
understanding the neural basis of knowledge representation, reasoning, and memory, with
an emphasis on the representation and processing of relational (first-order) knowledge. In
particular, I will describe SHRUTI, a neurally plausible model that demonstrates how a
suitably structured network of simple nodes and links can encode several hundred
thousand episodic and semantic facts, causal rules, entities, types, and utilities and yet
perform a wide range of explanatory and predictive inferences within a few hundred
milliseconds. I will examine some of the predictions stemming from this work about the
characteristics of common sense reasoning, and I will discuss how insights arising from
this work can be leveraged to design a scalable inference system running on conventional
computers.