Nicholas L. Cassimatis

Integrating cognition, perception and action through mental simulation in robots

We argue that many problems in robotics arise from the difficulty of integrating multiple knowledge representation and inference techniques. We describe an architecture that integrates disparate reasoning, planning, sensation and mobility algorithms by composing them from strategies for managing mental simulations. Since simulations are conducted by modules that include high-level knowledge representation and inference techniques in addition to algorithms for sensation and reactive mobility, cognition, perception and action are continually integrated. An implemented robot using this framework in object-tacking and human–robot interaction tasks demonstrates that knowledge representation and inference techniques enable more complex and flexible robot behavior.

A cognitive substrate for achieving human-level intelligence

Making progress toward human-level artificial intelligence often seems to require a large number of difficult-to-integrale computational methods and enormous amounts of knowledge about the world. This article provides evidence from linguistics, cognitive psychology, and neuroscience for the cognitive substrate hypothesis that a relatively small set of properly integrated data structures and algorithms can underlie the whole range of cognition required for human-level intelligence. Some computational principles (embodied in the Polyscheme cognitive architecture) are proposed to solve the integration problems involved in implementing such a substrate. A natural language syntactic parser that uses only the mechanisms of an infant physical reasoning model developed in Polyscheme demonstrates that a single cognitive substrate can underlie intelligent systems in superficially very dissimilar domains. This work suggests that identifying and implementing a cognitive substrate will accelerate progress toward human-level artificial intelligence.

Ability, Breadth, and Parsimony in Computational Models of Higher-Order Cognition

Computational models will play an important role in our understanding of human higher-order cognition. How can a models contribution to this goal be evaluated? This article argues that three important aspects of a model of higher-order cognition to evaluate are (a) its ability to reason, solve problems, converse, and learn as well as people do; (b) the breadth of situations in which it can do so; and (c) the parsimony of the mechanisms it posits. This article argues that fits of models to quantitative experimental data, although valuable for other reasons, do not address these criteria. Further, using analogies with other sciences, the history of cognitive science, and examples from modern-day research programs, this article identifies five activities that have been demonstrated to play an important role in our understanding of human higher-order cognition. These include modeling within a cognitive architecture, conducting artificial intelligence research, measuring and expanding a models ability, finding mappings between the structure of different domains, and attempting to explain multiple phenomena within a single model.

An Architecture for Adaptive Algorithmic Hybrids

We describe a cognitive architecture for creating more robust intelligent systems. Our approach is to enable hybrids of algorithms based on different computational formalisms to be executed. The architecture is motivated by some features of human cognitive architecture and the following beliefs: 1) Most existing computational methods often exhibit some of the characteristics desired of intelligent systems at the cost of other desired characteristics and 2) a system exhibiting robust intelligence can be designed by implementing hybrids of these computational methods. The main obstacle to this approach is that the various relevant computational methods are based on data structures and algorithms that are difficult to integrate into one system. We describe a new method of executing hybrids of algorithms using the focus of attention of multiple modules. The key to this approach is the following two principles: 1) Algorithms based on very different computational frameworks (e.g., logical reasoning, probabilistic inference, and case-based reasoning) can be implemented using the same set of five common functions and 2) each of these common functions can be executed using multiple data structures and algorithms. This approach has been embodied in the Polyscheme cognitive architecture. Systems based on Polyscheme in planning, spatial reasoning, robotics, and information retrieval illustrate that this approach to hybridizing algorithms enables qualitative and measurable quantitative advances in the abilities of intelligent systems.

Reasoning as simulation

The theory that human cognition proceeds through mental simulations, if true, would provide a parsimonious explanation of how the mechanisms of reasoning and problem solving integrate with and develop from mechanisms underlying forms of cognition that occur earlier in evolution and development. However, questions remain about whether simulation mechanisms are powerful enough to exhibit human-level reasoning and inference. In order to investigate this issue, we show that it is possible to characterize some of the most powerful modern artificial intelligence algorithms for logical and probabilistic inference as methods of simulating alternate states of the world. We show that a set of specific human perceptual mechanisms, even if not implemented using mechanisms described in artificial intelligence, can nevertheless perform the same operations as those algorithms. Although this result does not demonstrate that simulation theory is true, it does show that whatever mechanisms underlie perception have at least as much power to explain non-perceptual human reasoning and problem solving as some of the most powerful known algorithms.

An architectural framework for complex cognition

Any non-trivial task requires an appropriate representational formalism. Usually, for single-task or single-domain problems this choice of formalism is not explicitly made by the agent itself, but by the agent designer, and is implicit in the choice of data structures and algorithms used by the agent. However, complex cognition involves domains where the type of problems that the agent is expected to solve is not clear at the outset. Instead, at each stage of the problem solving process, the agent is expected to choose an appropriate formalism, solve the problem and integrate these results over the course of the entire problem solving episode. In this paper, we present one approach to solving two of the above problems - how does an agent choose the right representation and how can it integrate results from multiple representations over the course of problem solving? We present this approach in the context of Polyscheme, a cognitive architecture that is strongly integrated, focused on inference and adaptive to new information. We describe the representational formalisms and associated processes present in Polyscheme (propositional and spatial) and the decision cycle that allows information from multiple representations to be integrated. Using examples from complex tasks such as constraint satisfaction, language understanding and planning, we show how a Polyscheme agent can show improved performance by leveraging its multiple representations without the hindsight of representational choice.

Achieving human-level intelligence through integrated systems and research: introduction to this special issue

This special issue is based on the premise that in order to achieve human-level artificial intelligence researchers will have to find ways to integrate insights from multiple computational frameworks and to exploit insights from other fields that study intelligence. Articles in this issue describe recent approaches for integrating algorithms and data structures from diverse subfields of AI. Much of this work incorporates insights from neuroscience, social and cognitive psychology or linguistics. The new applications and significant improvements to existing applications this work has enabled demonstrates the ability of integrated systems and research to continue progress towards human-level artificial intelligence.

Artificial Intelligence and Cognitive Modeling Have the Same Problem

Cognitive modelers attempting to explain human intelligence share a puzzle with artificial intelligence researchers aiming to create computers that exhibit human-level intelligence: how can a system composed of relatively unintelligent parts (such as neurons or transistors) behave intelligently? I argue that although cognitive science has made significant progress towards many of its goals, that solving the puzzle of intelligence requires special standards and methods in addition to those already employed in cognitive science. To promote such research, I suggest creating a subfield within cognitive science called intelligence science and propose some guidelines for research addressing the intelligence puzzle.

Flexible Inference with Structured Knowledge through Reasoned Unification

Systems with human-level intelligence must be both flexible and able to reason in an appropriate time scale. These two goals are in tension, as manifested by the contrasting properties of general inference algorithms and structured knowledge-based systems. The problem of resolving ambiguous, implicit, and nonliteral references exemplifies many of these difficulties. We describe an approach, called reasoned unification, for dealing with these challenges by representing and jointly reasoning over linguistic and nonlinguistic knowledge (including structures such as scripts and frames) within the same inference framework. Reasoned unification enables a treatment of several reference resolution phenomena that to our knowledge have not previously been the subject of a unified analysis. This analysis illustrates how reasoned unification can resolve many difficult problems with using complex knowledge structures while maintaining their benefits.

Characterization of reasoning in terms of perceptual simulation

Although characterizing reasoning and natural language semantics in traditional logic captures their complexity and productivity, accounting for the grounding of logical reasoning in perception raises several challenges. These include difficulties in explaining the integration of reasoning and perceptual processing, and in accounting for the evolution of human reasoning from sensorimotor origins. Central to these problems is the fact that traditional logic includes elements such as quantifiers and negation that do not obviously occur in perceptual representations. We propose a formal framework in terms of perceptual simulation that bridges this gap. We demonstrate that perceptual simulations have the power to explain crucial elements of logical human reasoning and also allow us to provide the first unified linguistic analysis of noun phrases, negative polarity items and branching quantifiers within a single cognitively motivated formal framework.