If one wants to make a machine conscious, what ingredients should one put in? Asking this question is not only of some practical significance; it is also an excellent way of testing one's understanding of how the only machine that can surely generate consciousness - namely our brain - is capable of doing it. Obviously, an answer can be given only in the light of some theoretical notion of what consciousness is.

The theoretical approach we have developed maintains that consciousness is integrated information. High information content is an essential feature of conscious experience, in the precise sense that each conscious state is selected out of a repertoire of billions of possible ones. Moreover, this information is integrated - conscious experience is a unified whole that cannot be subdivided into independent components. We have also noted that the availability of a large amount of integrated information over a short period of time has obvious adaptive advantages.

If the notion of consciousness as integrated information is to be scientifically productive, as well as testable, one needs clearly defined concepts and corresponding measures. For this purpose, we have developed measures aimed at characterizing the degree to which a set of elements can integrate information. The main one is a general, information-theoretical definition of integrated informa-tion, called complexity. This measure, which is based on the amount of information that can be exchanged between bipartitions of subsets of elements, can be used to determine whether a subset of elements consti-tutes an integrated complex, as well as how many different states are available to it. Using simplified models, it can be shown that complexity constitutes an objective measure of the extent to which functionally specialized parts work together in an integrated fashion.

A key advantage of having a measure of integrated information is that it makes it possible to test the relationship between consciousness and integrated information in exemplary cases. At present, it seems most useful to begin where the distinctions are most obvious. We consider two basic contrasts. The first one is a contrast between two portions of the brain - the thalamocortical system and the cerebellum - which differ dramatically in their ability to give rise to conscious experience. The cerebellum has an estimated 50 billion neurons, against an estimated 30 billion for the thalamocortical system. It has probably as many connections, similar neurotransmitters and neuromodulators and, just as the thalamo-cortical system, it receives multiple sensory inputs. However, while many areas of the thalamocortical system are each essential for a different aspect of conscious experience, the entire cerebellum can be removed without significant changes in conscious experience. This simple contrast indicates that the generation of conscious experience is not merely a property of numbers of neurons and connections, of neurochemical diversity, or of biological intricacy.

A second illuminating contrast has to do with the neural activity within the thalamocortical system when we are awake and conscious vs. when we are asleep and unconscious, specifically during dreamless sleep. The anatomical orga-nization of the thalamocortical system obviously does not change between sleep and waking, and we now know that the amount of neural activity is not substantially different, except that during deep sleep it is highly synchro-nized.

We suggest that the reason the thalamocortical system can generate conscious experience and the cerebel-lum (or other parts of the brain) cannot is that the former is organized in such a way as to integrate a large amount of information, while the latter is not. Similarly, we predict that during deep sleep, while the anatomy and the amount of neuronal firing do not change appreciably, the amount of integrated information should markedly decrease.

In principle, these predictions can be tested empirically, e.g. through neuroimaging experiments. While such experiments are essential for validating the proposed concepts, they have several limitations in terms of the available spatial and temporal resolution as well as in terms of what manipulations are feasible. An equally important avenue to validate these concepts is to develop large-scale simulations of brain circuits in which all the rele-vant parameters are precisely known and every possible manipulation can be performed. Moreover, such a synthetic approach can provide guidelines about the requirements for developing an architecture endowed with conscious expe-rience in a machine.

Selected References:

1. Tononi, G., Complexity and coherency: integrating information in the brain. Trends in Cognitive Sciences, 1998. 2(12): p. 474-484.
2. Tononi, G. and G.M. Edelman, Consciousness and complexity. Science, 1998. 282(5395): p. 1846-51.
3. Edelman, G.M. and G. Tononi, A Universe of Consciousness: How matter becomes imagination. 2000, New York: Basic Books.
4. Tononi, G., Information measures for conscious experience. Archives Italiennes de Biologie, 2001 in press.