Cognitive artificial intelligence

What is required to allow an artificial agent to engage in rich, human-like interactions with people? The author argues that this will require capturing the process by which humans continually create and renegotiate “bargains” with each other, about who should do what in a particular interaction, which actions are allowed or forbidden, and the momentary conventions governing communication, including language. But such bargains are far too numerous, and social interactions too rapid, to be stated explicitly. Moreover, the very process of communication presupposes innumerable momentary agreements, thus raising the threat of circularity. Thus, the improvised “social contracts” that govern people's interactions must be implicit. The author draws on the recent theory of virtual bargaining, according to which social partners mentally simulate a process of negotiation, to outline how these implicit agreements can be made, and note that viewpoint raises substantial theoretical and computational challenges. Nonetheless, the author suggests that these challenges must be met to create artificial intelligence systems that can work collaboratively alongside people, rather than being helpful special-purpose computational tools.

Artificial Intelligence emulations of reasoning have traditionally employed a representation of an environment consisting of a set of rules and facts expressed in a symbolic language. Goals are also expressed in this language and the rules and facts are used to solve them. For instance, a robot might use a representation of its environment to form plans to achieve goals.

The group argues that successful reasoning requires the representation to be fluid. Not only might the facts describing the environment need to change as the environment changes, but the rules might also change. Moreover, the language in which the rules and facts are expressed, may also have to change. For instance, a rule might acquire a new precondition when the old rule is discovered to hold only in limited circumstances. A concept in the language might also need to be divided into variants together with corresponding variants of the rules and facts that use it. 

Mathematics provides the toughest test of this thesis. Surely, its language, rules and facts remain stable during the proof of a theorem. Not so. George Pólya has written the classic guide to the art of problem-solving (Pólya, 1945). Imre Lakatos has written a fascinating rational reconstruction of the evolution of mathematical methodology (Lakatos, 1976). Although it was not their intention to do so, both these authors have implicitly provided profound evidence for our thesis that representations should be fluid. 

This talk will discuss the potential of neural network models to learn grounded and structured lexical concepts. The author will give an overview of two recent lines of work. The first asks whether neural networks trained to model the physical world can learn representations of concepts that align well to human language. The second asks whether neural networks conceptual representations are compositional and structured in the way we expect human representations to be.

Statistical machine learning typically achieves high accuracy models by employing tens of thousands of examples. By contrast, both children and
adult humans typically learn new concepts from either one or a small number of instances. The high data efficiency of human learning is not easily explained in terms of standard formal frameworks for machine learning, including Gold's Learning-in-the-Limit framework and Valiant's Probably-Approximately Correct (PAC) model. This presentation will explore ways in which this apparent disparity between human and machine learning can be reconciled by considering algorithms involving a preference for specificity combined with program minimality. The talk will show how this can be efficiently enacted using hierarchical search based on the use of certificates, logical matrices and pushdown automata to support the learning of compactly expressed maximal efficiency algorithms. Early results of a new system called DeepLog indicate that such approaches can support efficient top-down identification of relatively complex logic programs from small numbers of examples.

The talk will bring together two closely related, but distinct, notions: argument and explanation. First, it will clarify their relationship and  then provide an overview of relevant research on these notions, drawn both from the cognitive science and the AI literatures. This will then be used to identify key directions for future research, indicating areas where bringing together cognitive science and AI perspectives would be mutually beneficial.

Our use of language goes far beyond a simple decoding of the literal meaning of the speech signal. Philosophers have argued instead that language use involves complex reasoning about the knowledge and intentions of interlocutors. Yet language use is fast and effortless. How can language depend on complex reasoning while being so easy? In this talk Associate Professor Goodman will explore this question using both structured and autoregressive probabilistic models, considering cases including implicature and metaphor.