In the study of language, some of the most interesting observations are made, not in terms of the components of language, but in terms of the way language is used, even how pauses are used, as in Jerry Seinfeld’s commentary. We have already considered some of the features of language in use when we discussed pragmatics in Chapter 10. We were, in effect, asking how it is that language-users successfully interpret what other language-users intend to convey. When we carry this investigation further and ask how we make sense of what we read, how we can recognize well-constructed texts as opposed to those that are jumbled or incoherent, how we understand speakers who communicate more than they say, and how we successfully take part in that complex activity called conversation, we are undertaking what is known as discourse analysis.

Previous chapters have explored the capacity of neural networks for modeling cognition. This chapter looks at applications to infant cognitive development. The first section reviews the trajectory of infants' understanding of object permanence and their ability to engage in physical reasoning, and how the symbolic representation theory can interpret the phenomenon. The second section introduce examples showing that neural networks can accommodate infant reasoning development without explicit rules and symbolic representations. The third section considers the relationship between symbolic models and neural network models, exploring an argument from Fodor and Pylyshyn trying to show that artificial neural networks are not genuine competitors to symbolic accounts.

When we considered the process of language acquisition, we concentrated on the fact that what is naturally acquired by most children is speech. Yet this is not the only way that a first language can be acquired. Just as most children of English-speaking or Spanish- speaking parents naturally acquire English or Spanish at a very early age, so the deaf children of deaf parents naturally acquire sign (or sign language).

In Chapter 1, we noted some of the basic features of the human vocal tract and the intricate muscle interlacing in and around the mouth that give humans the ability to produce a wide range of sounds with great speed. Yet, as they chatter away, humans do not simply produce a random selection of these sounds. Only certain sounds are selected on a regular basis as significant for communicative activity. In order to identify and describe those sounds, we have to slow down the chatter of everyday talk and focus on each individual sound segment within the stream of speech. This may seem straightforward, but it is not an easy task.

This chapter introduces machine learning in contemporary artificial intelligence. The first section looks at an expert system developed in the early days of AI research – ID3, which employs a decision-tree-based algorithm. The second section looks at advances in deep learning, which has transformed modern machine learning. We introduce a deep learning model inspired by the mammalian visual system, illustrating how it can extract hierarchical information from the raw data. The third section addresses two examples of neural networks -- autoencoders and convolutional neural networks, which can feature in layers of deep learning networks. The last section looks at a distinct type of machine learning -- reinforcement learning. We explain how deep reinforcement learning has made possible the two most spectacular milestones in artificial intelligence - AlphaGo and AlphaGo Zero.

This chapter looks at three milestones in the dawn of cognitive science. The first section introduces Winograd's computer model - SHRDLU. SHRDLU is a computer program designed to understand human language and respond to commands to perform actions in an artificial environment. SHRDLU's abilities suggest that the representation of mental activity might consist of language-like grammatical structures. The second section discusses an alternative model of representation. The mental image studies reported by Shepard and Kosslyn suggest that some forms of mental representation might be analog rather than digital. The third section looks at Marr's highly influential theory of vision, particularly the three different levels distinguished in his model of visual processing.

This chapter looks back to the pioneering studies that prefigured the emergence of cognitive science, subsequently converging into this new interdisciplinary field in the late 1970s. The first section addresses a key turning point in psychology. Whereas behaviorism holds that all explicit behaviors are the product of conditioning, it became clear that animals can manipulate representations of the environment to solve complex problems without reinforcement. The second section introduces the Turing machine developed by Turing in the 1930s, illustrating that purely mechanical procedures can process information algorithmically. The idea of a computable machine contributed to the birth of computer science and provided a model for thinking about how the mind processes information. Chomsky's transformational grammar offers a classic example of a computable model of how complex sentences convey information as a function of basic syntax rules. Finally, Miller and Broadbent's findings on attention support applying the information-processing model in psychology. These pioneering researchers were the first to lay out some of what were to be the basic concepts of cognitive science.