cogBooks

It's a very accessible outline of a non-representationalist and systems alternative to the mainstream representationalist tradition (and its hybrid variants) in Cognitive Science and Neuroscience. A.Chemero provides a nice analysis of why this perspective might be worth taking seriously. Ecological perception and Enaction are well explained and combined in the Radical Embodied paradigm.

0 thought experiments in the book.

Strongly recommend for broadening your views on what minds might be and how to study them!

It's a nice introduction to the embodied ideas in Cognitive Science, accompanied by the demonstrations from robotic simulations and experiments with humans. A.Clark reviews why the cognitive adaptive systems (such as human brains) are best viewed as inherently embodied, coupled to and extended by their environments, and, therefore, could not be successfully studied in isolation. He takes a "hybrid" perspective and suggests that the symbolic, representationalist approaches should be combined with the embodied paradigms to move forward in the shared goal of understanding how minds work.

The book can be divided into 2 main parts: (1) a critical analysis of the current state of AI and (2) a proposal of how to fix its problems. The first part is a nice read for people who believe in extraordinary power of the modern AI (even though it's overlong: the examples are repetitive, and the main thesis does not develop since the first chapter). The authors demonstrate how difficult to solve even the seemingly simplest problems are for machines. They often refer to "deep understanding" as the core part of what machines are lacking, however, they never specify what this fancy term means and why they believe that humans, in contrast, are able to deeply understand the world.

The solutions that Marcus and Davis propose in the second part of the book are not useful: mostly, they just state that the difficult problems reviewed in the first part should be solved. Their informative suggestion is that the hybrid symbolic-neural architectures should be used to achieve the "deeper understanding" and robustness in AI.

My opinion: I don't think that this (hybrid architectures) idea is justified: the problems that the authors outline have no clear connection with the proposed solution, except for the fact that the symbolic models have been used in Cognitive Science and Linguistics for many years. One of the main insights from the artificial neural network and robotic simulations is that our intuitions and high-level engineering solutions are always very far from the solutions that the learning system will converge to, even for the simplest problems with all the information about the learning algorithm, model, and data provided to us. There are multiple shortcuts to solve the cognitive problems without our pre-made "intelligent" solutions, which are available in the problem-specific structure of the complex environmental data. These results question the validity of the intuitive, interpretable, and human-engineered solutions to the learning problems that cognitive scientists were relying upon when they were using the symbolic algorithms. Marcus and Davis' suggestion moves us back to relying on our intuitions and old school ideas of what the cognitive system should do to perform the tasks, instead of letting it do what it can and probing different environments, learning rules and constraints for the learning. They cite the Kajal's saying, which is very appropriate to the debate, but seems to contradict their own proposal: "nature seems unaware of our intellectual need for convenience and unity, and very often takes delight in complication and diversity."

The main goal of this book is articulating a promising paradigm for uncovering mechanisms of cognitive development, which the authors contrast to popular nativist approaches where high-level cognitive abilities are argued to be inherited.

Various opponents of connectionism have argued that this paradigm cannot account for mechanisms of cognitive change. This book explains how connectionist models can be used to simulate cognitive development, and provides plenty of examples of how it has been done. The authors also debunk a myth that connectionism is an empiricist research program. On the contrary, many ways of how the innate biases can be represented in biological systems, and how those can be incorporated into artificial neural networks are discussed. Authors suggest three different classes of innate cognitive biases: representational, architectural, and chronotopic (determining timing of some developmental events). Representational innateness, most often invoked by nativist psychologists, is suggested to be unlikely.

I especially liked that the authors jointly contributed to the book's chapters, so the resulting perspective is quite integrative across approaches and views on cognitive development. Earlier chapters also provide accessible introductions to connectionist modeling, complex systems, and the necessary foundations of neuroscience to ground main arguments and examples in this book.