Interior Grounding, Reflection, and Self-Consciousness

 

Marvin Minsky

MIT

 

In Brain, Mind and Society, Proceedings of an International Conference on Brain, Mind and Society, Graduate School of Information Sciences, Brain, Mind and Society, Tohoku University, Japan, September 2005. See http://www.ic.is.tohoku.ac.jp/~GSIS/

 

====================

 

Some computer programs are expert at some games. Other programs can recognize some words. Yet other programs are highly competent at solving certain technical problems. However, each of those programs is specialized, and no existing program today shows the common sense or resourcefulness of a typical two-year-old child—and certainly, no program can yet understand a typical sentence from a child’s first-grade storybook. Nor can any program today can look around a room and then identify the things that meet its eyes.

 

This lecture will suggest some ideas about why computer programs are still so limited. Some thinkers might say that this is because computers have no consciousness, and that nothing can be done about this, because it is in the nature of machines to only what they are programmed to do—and therefore they cannot be programmed to ‘think’.

 

Citizen: I am convinced that machines will never have thoughts or feelings like ours, because machines lack vital ingredients that can only exist in living things. So they cannot have any feelings at all, no hopes or joys or fears or pains—or motives, ambitions, or purposes. They cannot have the faintest sense of pride or shame, or of failure, achievement, or discontent, because they simply can't care about what they do, or even know they exist.

 

It seems to me that we use such statements to excuse ourselves for our failures to understand ourselves. To do this, we collect the phenomena that we can’t yet explain, and then pack them into such ‘suitcase-like’ words as sentience, spirit, or consciousness—and then describe these “vital ingredients” as entities with mysterious traits that can’t be explained in physical ways.

 

However, here I will take an opposite view. Whenever some seemingly ‘basic’ aspect of mind seems hard to explain, I will try to depict it as the product of some more complex network of processes—whose activities may sometimes cooperate, but may also have ways to conflict and compete. Then in each of the examples below, a mystery that seemed inexplicable will then be replaced by a set of several different questions and problems, each of which may still be difficult, but at least won’t seem so more intractable. We’ll start by unpacking the set of phenomena for which we have come to use the word “consciousness.” (This following section is condensed from chapter 4 of my forthcoming book, “The Emotion Machine.”)

 

∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞

 

What is Consciousness?

 

Aaron Sloman: “It is not worth asking how to define consciousness, how to explain it, how it evolved, what its function is, etc., because there's no one thing for which all the answers would be the same. Instead, we have many sub-capabilities, for which the answers are different: e.g. different kinds of perception, learning, knowledge, … self-control, etc.”— From a message in comp.ai.philosophy, 14 Dec. 1994

 

To see how many things human minds do, consider this fragment of everyday thinking.

 

Joan is part way across the street on the way to deliver her finished report. While thinking about what to say at the meeting, she hears a sound and turns her head —and sees a quickly oncoming car. Uncertain whether to cross or retreat, but uneasy about arriving late, Joan decides to sprint across the road. She later remembers her injured knee and reflects upon her impulsive decision. “If my knee had failed, I could have been killed. Then what would my friends have thought of me?”

 

It might seem natural to ask, "How conscious was Joan of what she did?" But rather than dwell on that ‘consciousness’ word, let’s look at a few of the things that Joan actually “did.”

 

Reaction: Joan reacted quickly to that sound.

Identification: She recognized it as being a sound.

Characterization: She classified it as the sound of a car.

Attention: She noticed certain things rather than others.

Indecision: She wondered whether to cross or retreat.

Imagining: She envisioned some possible future conditions.

Selection: She selected a way to choose among options.

Decision: She chose one of several alternative actions.

Planning: She constructed a multi-step action-plan.

Reconsideration: Later she reconsidered this choice.

 

In the course of doing those things, other ‘parts’ of Joan’s mind did other things.

 

Recollection: She retrieved descriptions of prior events.

Representation: She interconnected a set of descriptions.

Embodiment: She tried to describe her body's condition.

Emotion: She changed major parts of her mental state.

Expression: She constructed several verbal descriptions.

Narration: She heard them as dialogs in her mind.

Intention: She changed some of her goals’ priorities.

Apprehension: She was uneasy about arriving late.

Reasoning: She made various kinds of inferences.

 

Many of these activities involved mental processes that used descriptions of some of her other mental processes.

 

Reflection: She thought about what she’s recently done.
Self-Reflection: She reflected on her recent thoughts.

Empathy: She imagined other persons’ thoughts.

Moral Reflection: She evaluated what she has done.

Self-Awareness: She characterized her mental condition.

Self-Imaging: She made and used models of herself.

Sense of Identity: She regarded herself as an entity.

 

That’s only the start of a much longer list of aspects of how people feel and think— and if we want to understand how our minds work, we’ll need explanations for all of them. To do this, we’ll have to take each one apart, to account for the details of how it works—and then decide which of them to regard as aspect of what we call ‘consciousness.’

 

∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞

 

The mind seen as an organization of multi-level processes.

 

This section outlines a model of mind that shows how a system could reflect (to at least some extent) on what it was recently thinking about. There is not enough room to describe the whole idea here, but the reader can find more details at http://web.media.mit.edu/~minsky/E5/eb5.html.

 

My associate Push Singh and I are at present developing a prototype of a system like this. We describe more details about this in http://web.media.mit.edu/~minsky/E4/eb4.html and in http://web.media.mit.edu/~push/CognitiveDiversity.html

 

I should note that this model is consistent with some of the early views of Sigmund Freud, who saw the mind as a system for resolving (or for ignoring) conflicts between our instinctive and acquired ideas.

 

 

 

∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞

 

How do we Recognize Consciousness

 

Once we have a model in which the mind has several such layers of processes, we can start to construct hypotheses about what might be happening when a person claims to be thinking ‘consciously’. For example, this might happen when a certain process in that person’s brain detects some combination, such as this, of higher-level activities.

 

 

Similarly, we also could ask about what might cause a person to initiate such a set of activities. This might happen, for example, when a certain kind of ‘critic-like’ process detects that your thinking has got into trouble. The effect of such a critic might then help you to the sorts of things that we sometimes describe as trying to “focus’ or ‘concentrate.” The diagram below suggests one kind of process that one’s brain might use to try to switch itself into some pattern of thinking engages more high-level processes—for example, by activating resources like these:

 

 

It is important to emphasize that each of those sets of activities can be extremely complex, and also are likely to differ significantly between different individuals. This is yet one more reason why no one should expect to be able to find a simple description of what we call consciousness.

 

∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞

 

How do our mental levels develop and grow?

 

Over the past few centuries, our scientists have discovered far more about atoms and oceans and planets and stars than about the mechanics of feelings and thoughts. Those sciences progressed because those scientists were successful at discovering very small sets of “simple” and “basic” laws that explained many different phenomena in the realms of physics and chemistry.

 

Why did that strategy work less well for the science of psychology? It seems to me that one reason for this was an almost universal belief that such functions as emotions and feelings must be essentially non-mechanical— and that, therefore, they could not be explained in terms of physical processes. However, I suspect that the principal cause of this delay was the idea that psychologists, like physicists, should also seek simple “laws” of thought. In other words, it seems to me, that our psychologists and philosophers should not have tried so hard to use the methods that worked so well for those physical sciences. In fact, today we know that every human brain contains several hundred different, specialized kinds of machinery—each of which must have evolved different processes that helped our ancestors to solve the various problems that they faced in thousands of different ancient environments. So tens of thousands of different genes must be involved with how people think.

 

This suggests that modern psychologists should consider taking an opposite view, and reject the urge to base their ideas on discovering small sets of simple laws. Whenever some aspect of mind seems hard to explain (such as affection, fear, or pain), we could attempt, instead, to replace it by a more complex set of interconnected processes. In other words, we’ll take each mental phenomenon and try to depict it, not as so ‘basic’ and ‘elementary’ that it is inexplicable, but as resulting from the complex activities of big networks of different processes—which sometimes cooperate and sometimes compete. Then each mystery will begin to disappear, because of having been replaced by a several new kinds of problems. Each of those problems may still be quite hard, but because they are far less mysterious, we’ll be able to start to deal with them.

 

In other words, our main technique will be to demonstrate many seemingly separate ‘features’ of our minds are actually not single things but are aspects of what happens inside huge networks of different processes. To do this we’ll need to accumulate ideas about how some of those processes work, and then we’ll need to propose some ways that these might combine to produce the systems that we call our minds.

 

So now let us try to apply this idea to the question of how human learning works. It is easy enough to imagine machines with many levels of processes; indeed, many computer programs today are made of multiple layers of sub-programs. However, we still do not have good hypotheses about how our higher levels of brain-machinery come to do all the wonderful things that they do.

 

∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞

 

The myth of “Grounding in Experience.”

 

Most theories of human development assume that we begin by learning low-level reactions, and must wait for each stage to consolidate before we can learn to think more abstractly:

 

“Everything that we come to know—from the simplest facts to our most abstract concepts—is ultimately “grounded” on our experiences with the external world.”

 

More specifically, that ‘standard theory’ goes on to insist:

 

We begin by (somehow) learning to recognize particular sensory situations. Then we correlate our reactions with whether they lead to failure and success.

 

Then, in subsequent stages of development, we learn increasingly abstract ways to represent the objects and their relationships in the situations that we perceive.

 

However, this raises serious questions like these:

 

How do we recognize those “sensory situations”?

How do we represent them?

What determines how we react to them? (“Operants.”)

What constitute ‘success’ and ‘failure’?

How do we make those correlations?

 

To answer such questions, it seems to me, we will need many new ideas about how to design such machinery. I doubt that it will ever suffice to assume (for example) that learning is basically a matter of statistical correlations, or that high-level concepts will spontaneously form in large neural networks with simple architectures—or that we will come to understand much of human cognition by making small extensions to traditional concepts about “association of ideas” or “operant reinforcement.” One great philosopher clearly recognized that those ideas had serious deficiencies:

 

Immanuel Kant: “That all our knowledge begins with experience there can be no doubt. For how is it possible that the faculty of cognition should be awakened into exercise otherwise than by means of objects which affect our senses, and partly of themselves produce representations, partly rouse our powers of understanding into activity, to compare, to connect, or to separate these—and so to convert the raw material of our sensations into a knowledge of objects?”

 

“But, though all our knowledge begins with experience, it by no means follows that all arises out of experience. For, on the contrary, it is quite possible that our empirical knowledge is a combination of that which we receive through impressions, and [additional knowledge] altogether independent of experience … which the faculty of cognition supplies from itself, sensory impressions giving merely the occasion. [Immanuel Kant, Introduction to Critique Of Pure Reason, Second edition, April 1787]

 

For although, as Kant remarked, sensations give us occasions to learn, this cannot be what makes us able to learn: in other words, it does not seem to explain how a person first could learn to learn. Instead, you need to begin with some ‘additional knowledge’ about how to produce representations and then to connect them. This is why, it seems to me, our human brains first had to evolve the kinds of complex architectures that our neuroscientists see.

 

For example, the traditional points of view do not begin to explain why the ‘stages’ of children’s development so frequently seems highly abrupt; a child may spend an entire year expressing only “sentences” that contain no more than one or two words—and then, more complex expressions may quickly appear. This has led to a belief that has been popular for many years: that such capabilities must simply be “innate,” and are actually not “learned” at all. Accordingly, that viewpoint holds that the child needs only to “tune up” or, in some way, adapt that machinery to the language of its culture, so that it can automatically speak properly when the developmental “time is right.” The following section suggests, instead, that different levels of learning could have been proceeding simultaneously throughout that period, but do not usually appear in overt behavior until the resulting processes have become sufficiently competent.

 

∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞

 

A Theory of “Interior Grounding”

 

The old ‘physical grounding hypothesis’ assumed that no ‘higher cognitive level” could start to learn before the levels below it have learned enough. In this view, mental development must begin with processes in which the child’s lowest level reactive systems acquire some knowledge about that child’s external environment:

 

 

Only then could the next level start to learn—because (in that traditional view) the construction of each new structure must be based on the foundations of what the levels below it have learned.

 

However, we can imagine a different kind of process in which each of several levels of the brain can, at the same time, learn some ways to predict and control some of the activities in the parts of the brain to which it directly connects. In other words, each part of the brain exists inside its own ‘local world’. Then we can make a new hypothesis: evolution could have provided each of those local worlds with what we might call “mini-worlds” that genetically have been already each equipped with potentially useful kinds of behaviors.

 

A typical external mini-world might consist of the system comprising some fingers and hand; then the reactive system can learn to predict how various combinations of finger-motions lead to different palm-sensations. Such a system could learn to predict that clenching the fingers will cause a sense of pressure on the palm. Similarly, an infant reactive level could learn to predict the effects of larger-scale motions of the limbs, or motions of the tongue in the mouth, or some visual effects of moving the eyes.

 

So far, this is the conventional view, in which all of our learning is fin