Eric Horvitz

Many people grapple with challenging unknowns in the realms of philosophical and scientific discourse. I’ve found that the Bayesian tradition of defining a set of events of interest, and assessing probabilities to encode beliefs about the likelihood of those events, can be useful in communicating ideas and in guiding investigations of unexplained phenomena. We shall focus here on difficult questions in pursuit of understanding the foundations of conscious experience. Questions have long been posed about whether we might one day be able to create systems that experience subjective states similar to those experienced by people. The challenge of understanding and explaining conscious experience extends well beyond academic discussions among philosophers. Many people, including AI scientists pursuing principles of automated reasoning and decision making, and neurobiologists and psychologists with interests in nervous systems and cognition reflect about the nature of subjective states.

Diverse views thrive about the causal basis of consciousness in the absence of scientific methods for probing principles and machinery of conscious experience. When issues touching on “consciousness,” “awareness,” “subjective states,” “experience,” “feeling,” and “qualia” arise in conversation, people are faced with a wide spectrum of beliefs about the foundations of subjective experience, and on related beliefs about the feasibility of one day developing computational algorithms and architectures that could achieve what we might call conscious computing. Diversity in beliefs extends to scientists with expertise in cognition, including decision scientists, cognitive psychology, neurobiology, and artificial intelligence.

To promote discussion, I’ve found it useful to assess beliefs about alternate bases for conscious experience. Rather than seek commitment to a single hypothesis, I take a Bayesian perspective on assessing beliefs under uncertainty. With such an approach, we layout a broad set of distinct possibilities about the basis for conscious experience, and then assess the beliefs that individuals hold about the likelihood of each hypothesis actually being true. Such an assessment can be useful for revealing common patterns of belief and for focusing discussion.

The assessment of an explicit set of likelihoods across a space of possibilities that provide “conceptual coverage” at a broad level of abstraction supports the admission of multiple possibilities, albeit with different levels of belief, and helps to frame questions about the kinds of experiments, results, models, and observations that might be useful for updating such beliefs.

So, let’s layout a set of broad possibilities, attempting to define a mutually exclusive set of possibilities. We’ll work at a high-level of abstraction and talk about broad categories that could house more detailed hypotheses as further refinements. We’ll also force the set of possibilities to be exhaustive by considering an explicit Other category—a category containing all possible explanations that don’t fit into one of the other broad hypotheses.

For each hypothesis (including the Other hypothesis), I ask interested people to assess the probability that the correct explanation for conscious experience lays within that hypotheses. This assessed probability represents the individual’s belief about the basis for consciousness, expressed as a likelihood of the ultimate truth of the category. One way to view such probability assessments is to imagine that a clairvoyant, who knows the actual truth, will soon reveal the nature of consciousness. The likelihoods represent the probabilities of each hypothesis being the one pointed out by the clairvoyant as being the right explanation.

Figure 1. Formulation of a space of hypotheses about subjective experience. After seeking agreement on the essence of each hypothesis, we can assess beliefs, in terms of the differing likelihoods of a mutually exclusive and exhaustive set of hypotheses.

Let us now explore a set of candidate hypotheses, each representing a class of explanation. Figure 1 captures a potential space of possible explanations for conscious experience. Each of the definitions is meant to conjure up a “conceptual centroid,” targeting the essence of an hypothesis. It’s clear that there is some potential difference of interpretation of the hypotheses, and so, to be more precise, people should have the opportunity during communications about beliefs to refine and personalize the definitions of the hypotheses. Let’s review each hypothesis.

Theological Foundations includes the set of explanations that view conscious entities as having foundations in the spiritual world, outside of or in distinction to the mechanistic, scientific universe studied by physicists, chemists, and biologists. These explanations implicitly or explicitly pose conscious experience as arising as an essential property of souls or other experiencing entities founded on what one might refer to as classical theological or spiritual constructs, as captured, for example, by the notions of individuals in Buddhist, Christian, Islamic, Hindu, Jewish, and other organized religions.

Information Processing covers the set of hypotheses that consider subjective experience as being generated by some aspect of the processing of information. As an example, the hypothesis covers the assertion that subjective experience is created by, or associated with, patterns of coordinated information processing that support the real-time capabilities for sensory integration, reasoning, and action demonstrated by the human nervous system. People with strong beliefs about the Information Processing hypothesis tend to believe that conscious experience is based on some form of algorithmic activity, whether that algorithmic activity is performed by cell-based nervous activity, a set of silicon memory units and gates, a universal Turing machine, or a contraption built from Styrofoam balls, wood sticks, and rubber bands.

Mysterious Fabric spans the set of hypotheses that explain consciousness as hinging critically on some, as yet poorly understand, property or properties of the physical universe. Someone with strong belief in this hypothesis might assert that it is likely that biological nervous systems discovered and leveraged these special properties through evolutionary processes. The essential nature of the property and its role in subjective experience may one day be revealed via ongoing efforts in physics and neurobiology. Such a property may be metaphorically or in reality related to some fundamental aspect of quantum physics, or another property of matter/energy in space time.

Other refers to the set of all explanations that are not captured by the explicitly defined categories. That is, we ask people to assess the likelihood that “something else is going on,” significantly distinct from the other, explicit categories. Other includes explanations that fall outside of the explicit hypotheses, including explanations that represent reformulations, extensions, or combinations of the definitions of the explicit categories—given personal interpretations of the definitions of the explicit hypotheses. In an interactive assessment, we can break out additional explicit, exclusive hypotheses rather than merge everything else into the Other category.

Notice that we can refine the hypotheses space, by breaking out as distinct hypotheses, more precise, finer-grained explanations. For example, as portrayed in Figure 2, we refine Information Processing into IP-Competency and IP-Leap.

Figure 2. Refinement of the space of hypotheses. In this case, we refine Information Processing into the IP Leap and IP Competency classes

IP Competency is a refinement of Information Processing that explains conscious experience as being based in information processing, and furthermore asserts that such subjective states are an unavoidable aspect of increasing cognitive competency. That is, what we call consciousness is unavoidable as real-time sensory integration and information processing become increasingly competent. From the point of view of IP-Competency, searching for a special word or concept like consciousness is chasing an illusory “distinct” notion. Our self awareness and subjective experience is simply the way entities with our abilities necessarily must feel. Should an explanation in the IP-Competency category be true, there would be nothing explicit to discover about conscious experience with the increasing competency of reasoning systems

IP Leap is a refinement of Information Processing that asserts that subjective experience is based on special, potentially interesting and surprising information processing. IP-Leap explanations assert that conscious experience is the result of some interesting, novel information-processing analysis, coordination, integration, architecture, etc., that endows nervous systems with valuable capabilities that would not necessarily be obtainable or obtainable with the same overall architecture or resource limitations without such special processing or machinery. Strong believers in IP-Leap explanations might find themselves agreeing with the suggestion that there are some discoverable analytical and/or architectural aspects of cognitive systems that would enhance our insights about understanding—and potentially synthesizing—systems capable of supporting conscious experiences.

In performing assessment, we can identify different patterns of belief among the population. Figures 3 through 5 capture several sample patterns of belief about the foundations for subjective experience.

Figure 3 captures the beliefs of an individual who finds consciousness mysterious, with an assessment that the explanation is likely founded on some as yet poorly understood property of the universe. I have found that many people with strong beliefs that consciousness likely is linked to some, as yet poorly understood, essential property of the physical universe, often cite an explanation for subjective states that “may have something to do with quantum physics.” In the case captured in Figure 3, the subject asserts that the next most likely hypothesis is Other, or something outside of the explicit hypotheses, followed by IP Leap hypothesis, and then IP Competency, followed last by Theology.

Figure 3. A pattern of belief with weight on the Mysterious Fabric hypothesis.

Figure 4. A pattern of belief with weight on the Information-Processing hypotheses, with a leaning toward IP Leap.

Figure 4 captures a pattern of beliefs leaning towards the conscious experience as a very special kind of information processing, followed by some probability that something else is going on, with an even less likelihood that the subjective world is an inescapable aspect of the intellectual competency associated with human (and potentially other) nervous systems. Although this person gives some probability to the Mysterious Fabric hypothesis, only a small probability is assigned to theological hypotheses.

Figure 5. A pattern of belief with weight on the Theological Foundations hypotheses.

Figure 5 represents beliefs of a person with fairly strong beliefs that subjective experience is founded in theology. Although, the person has assigned overwhelming belief to the Theological Foundations hypothesis, significant belief is also assigned to the Mysterious Fabric hypothesis, and to the Other category. This person largely rejects the Information Processing hypotheses.

Formulating distinct hypotheses about the potential basis for poorly understood phenomena—and expending effort to characterize personal beliefs about the competing explanations—can enhance dialog and provide insights about the best courses of analysis and experimentation for refining understanding. I believe such a Bayesian assessment of hypotheses will assist with discussions and progress toward a better understanding of the foundations of conscious experience.