Home > Kadri > Marvin Minsky: The Emotion Machine

Whenever our usual way to think fail, we can start to think about our own thoughts and this is reflective thinking, that can help us develop new way of thinking.

A typical brain contains a great many parts that we’ll call resources. We will try to describe mental activities like anger, love or embarrassment as result of the activities of a certain collection of mental resources.

Feeling, emotions or consciousness are not definite processes, they are effects of large networks of processes inside our brains.

Our brains must be equipped with resources called Critics – each of which is specialized to recognize some certain condition and then to activate a specific collection of resources.

Many of our emotional state results when certain particular Ways of Think start to suppress our use of certain resources. Our minds are always affected by our assumptions, values and purposes.

Marcel Proust: “Each reader reads only what is already inside himself.”[1]


Love is a kind of suitcase-like word. We need better way to understand how we think. Your Sled sees the world by using your senses. Then it stores what it learns in your memory. It originates all your desires and goals – and then solves all your problem for you, by exploiting your “intelligence”.

Emotion is one of the most suitcase like words that we use to conceal the complexity of very large ranges of different things whose relationships we don’t yet comprehend.

To understand infant emotions, we can simulate machine of their behavior. First, we need to make a list of goals that machine would need to achieve. We than need to build separate instinct-machine to meet those needs. And then package them into a single body-box. Inside instinct-machine we have three kinds of resources: some ways to recognize situation, some knowledge about how to react to these and some muscles or motors to execute actions.

If we already know in advance all the situations we might face, simple if and do model will work. Rule-based reaction machine is enough. In the beginning this will do, but later in life, we learn to use actions that change the external world.

Author is using resource in a hazy way to refer to all sorts of structures and processes that range from perception and action to Ways to Think about bodies of knowledge.

If some mental processes get stuck, we can split problems into smaller parts and try to remember if we already solved them in the past or to make several attempts and then compare and evaluate them.

If we deal with hard problems then we need to upgrade Rule-based reaction machines into Critics-selectors machine. Critics recognize a problem-type and Selector activate a Way to think.

We can categorize levels of mental procedure in six levels starting from instinctive behavioral system going all the way to values, censors and ideals:

  • Instinctive reactions
  • Learned reactions
  • Deliberative thinking
  • Reflective thinking
  • Self-reflective thinking
  • Self-conscious thinking

Each person has multiple Ways to Think and what we call “emotional” states are merely different examples of these. The concept of Rational Thinking is incomplete – because logic can only help us to draw conclusions from the assumptions that we happen to make – but logic, alone, says nothing about which assumptions we ought to make.

Attachments and Goals

It is one thing to learn how to get what you want and another to learn what you ought to want. Author is suggesting that self-conscious emotions like Pride and Shame play special roles: they help us learn ends instead of means. Where trial and error teach us new ways to achieve the goal already maintain – attachment-related blame and praise teach us which goals we should discard or retain. Our values and goals are greatly influenced by the people to whom we become “attached”.

You say that you want a certain thing when you have an active mental process that works to reduce the difference between your present situation and one in which you possess that thing.

The most common conception of how people learn is that our reactions are reinforced by success. But how do we learn what we ought to want. An Imprimer can be a person to whom a child has become attached. Imprimer’s praise can change a goal into value. Living up to one’s own internalized set of standards – or failing to live up to them – forms the basis of some very complex emotions.

There are several different ways in which our children might change themselves:

  • Positive experience
  • Negative experience
  • Aversion learning
  • Attachment praise
  • Attachment censure
  • Internal imprinting

We can define attachment-based learning scheme: if you detect Praise and an Imprinter is present, then “elevate” your present goal.

When a person learns, it is not just a matter of “making connections” but is also a matter of making the structures that then get connected.

Public Imprimers possess something called charisma. At first Imprimers must be near to us, but then we can only use models even when Imprimers are away and eventually these models become what we call conscience, ideals and moral codes.

Pain and suffering

Sometimes a pain is just a pain; if it doesn’t last long or it’s not too intense, then it won’t escalate into suffering. Pain is good to protect our bodies from harm. A major component of suffering is the frustration that comes with the loss of your options.

It is easy to recognize (as opposed to describe) a particular feeling or mental condition because you may only need to detect a few of its characteristic features.

Pain describes sensation that arrive quickly after an injury. Hurting is about what happens when pain elevates the goal to get rid of the pain. Suffering describes the states that result when this escalates into a large-scale cascade that disrupts all one’s usual Ways to Think.

When people talk about “feeling bad” they are referring to the disruption of their other goals and to the various conditions that result from this.

Suffering could be just a programing bug. Based on previous episodes of pain.

We are so focus on the physical actions that people do, while ignoring questions about what people do not do. Negative experience is according to author a very large part of every person’s precious collection of commonsense knowledge. Much of what we come to know is based on learning from our mistakes.

For learning to recognize some particular kind of potential mistakes, we have at least three types of Critics:

  • A Corrector declares that you are doing something dangerous.
  • A Suppressor interrupts before you begin the action you’re planning to take.
  • A Censor acts yet earlier, to prevent that idea from occurring to you.

Humans can switch from among different Ways of Think, but this is also a source of many of our conditions like tempers, moods and dispositions.

Critics are not only resources that detect mistakes, but also recognize successes and promising opportunities. These Critics are positive and can be called Encouragers. Some Critics are always on, others are only active on special occasions.

It would be dangerous if we could direct our Critics and switch and turn them on as we would like. Since this could mean that we could let’s say turn Hunger off and that could lead to our death.


A clear definition can make things worse, until we’re sure that our ideas are right. Consciousness is a suitcase-like word that we use to refer to many different mental activities, which don’t have a single cause of origin. Once we imagine a mind as made of smaller parts, we can replace the single, big problem by many smaller, more solvable ones.

We like to divide the functions of mind into pairs. But it makes more sense to split a mind into larger number of different parts. We can have A-brain, B-brain, C-brain all of them being responsible for certain number of functions.

All those parts need some levels and organization in order to work properly together. The Organism Principle: when a system evolves to become more complex, this always involves a compromise: if parts become too separate, then the system’s abilities will be limited – but if there are too many interconnections, then each change in one part will disrupt many others.

The processes we call consciousness it is as though the higher levels of our minds sit at mental terminals, steering great engines in our brains, not by knowing how that machinery works but by clicking on symbols from menu lists that appear on our mental screen displays.

We must try to design – as opposed to define – machines that can do what human minds can do.

We like to classify our activities into ones that we do intentionally, as opposed to actions we do unconsciously.

Any reflective systems might need at least four constituents:

  • Self-Models: we all construct mental models that describes our various mental states, bodies of knowledge about our abilities, depictions of our acquaintances and collections of stories about our pasts.
  • Serial processes: if we talk about parallel paradox, whenever one splits a problem into parts and tries to think about them at once, one’s intellect will get dispersed and leave less cleverness for each task. We have to do things sequentially.
  • Symbolic description: we can talk about connectionist network, which uses numbers to indicate how closely related are various pairs of parts or about semantic network, which uses three-way links to indicate that different components of the arch have different kinds of relationships.
  • Recent memories: we usually think of consciousness as being about what’s happening now. However, it would always take some amount of time for any particular part of a brain or machine to find out what other parts have recently done.

The Immanence Illusion: For most of the questions you would otherwise ask, some answers will have already arrived before the higher levels of your mind have had enough time to ask for them. Our sense of constant contact with the world is a form of the Immanence Illusion. If we take consciousness to mean “awareness of our internal processes”, it doesn’t live up to its reputation.

The truth is, that no mind is much employed upon the present: recollection and anticipation fill up almost all our moments. Although people talk about being conscious of what is happening now, that’s the one thing you cannot be conscious of – because, as we have mentioned before, each brain resources can know, at most, only what a few others were doing some moments ago.

Levels of mental activities

We can categorize levels of mental procedure in six levels starting from instinctive behavioral system going all the way to values, censors and ideals:

  • Instinctive reactions: we are born with instincts that help us survive.
  • Learned reactions: certain conditions demand specific ways to react.
  • Deliberative thinking: we can consider several alternatives and we can try to decide which would be the best.
  • Reflective thinking: we can react not to external events but to what is happening in our brain.
  • Self-reflective thinking: we can think about plans that we are making.
  • Self-conscious thinking: we can think about how well our actions fit our ideals.

Each species evolved machinery that enables its newborns off-spring to do many things without any prior experience. They start out with some built-in if/do reaction-rules. But sometimes simple if/do rules don’t work. For more complex situations we need more powerful, three part if + do / then rules. If (situation), do (action), then (result).

Learning by reinforcement is not the best approach. To achieve more complex goals, we need to make more elaborate plans by using all sorts of knowledge that we’ve gained from things that we’ve done in the past. We need to use some deliberation and we need to check alternatives. If we need to look for more than few steps ahead, complexity will grow exponentially. Our most effective human ways to solve hard problems are not based on making extensive searches.

Reflections on what already happened can help us to reduce our alternative searches. We can see that each level of thinking can observe and use descriptions of what happens in the level below it. Our reflective thinking most often begins when our usual systems start to fail.

It is hard enough to describe a thing that keeps changing its shape before your eyes – and surely it is harder yet to describe things that change when you thing about them.

There is an universal tendency among mankind to conceive all beings like themselves.

Instinctive Reactions keep our bodies and brains alive. The Learned Reaction level includes extensions of these that are learned after birth. The Deliberate and Reflective levels are engaged to help solve more difficult problems. Self-reflection enters when those problems require us to involve the models that we make of ourselves or our views of our possible futures. Self-Conscious Reflection that enables us to think about our “higher” values and ideals.

Today we know that visual systems in our brains receive many more signals from the rest of the brain than signals that come in from our eyes. The Immanence Illusion applies not only to imagined scenes; we never see real scenes “all at once” either.

In order for us to use higher level if thinking, we also need to have some kind of suppressor bands, mechanisms that will make sure our actions are not triggered by our vision of potential future until we decide to pursue certain line of actions.

Common sense

Why computers are not on a human level yet:

  • No present-day programs have Commonsense Knowledge. Each present-day program is equipped with no more knowledge than it needs for solving some particular problem.
  • Present-day programs do not have explicit Goals. Today, we only tell programs some things to do. Without telling them why we want them done.
  • Few present-day programs are Resourceful enough. A typical program will simply quit when it lacks some knowledge it needs or when the method it’s using fails.

We each use terms like commonsense for the things that we expect other people to know and regard as obvious. Every word, event, idea or thing can have many different meanings to us.

We can talk about different realms:

  • The Physical Realm
  • The Dominion Realm
  • The Procedural Realm
  • The Social Realm
  • The Economic Realm
  • The Conversational Language Realm
  • The Sensory and Motor Realm
  • Kinesthetic, Tactile and Haptic Realm
  • Cognitive Realm
  • The Self-Knowledge Realm

We can classify knowledge into two kinds:

  • Knowing what – declarative or explicit
  • Knowing how – procedural or tacit

Attempts to create a machine that can learn as babies were not successful. You cannot learn things that you can’t represent. John McCarthy believes that if one wants to be able to discover an abstraction, machine must be able to represent this abstraction in some relatively simple way.[2]

Some problems that researchers tackle in order to build that kind of machine:

  • The optimization paradox: The better a system already works, the more likely each change will make it worse.
  • The investment principle: The better a certain process works, the more we will tend to rely on it and the less we will be further inclined to develop new alternatives.
  • The complexity barrier: The more that the parts of a system interact, the more likely each change will have unexpected side effects.

Evolution is often described as a process of selecting beneficial changes, but most of evolution’s work involves rejecting changes that have bad effects.

Every child develops ways to link each new fragment of knowledge to some particular goals it might help to achieve.

One of important questions we need to clarify, when we think about thinking, is what are goals. A system will seem to have a goal when it persists at applying different techniques until the present situation changes into certain other conditions. Persistence, aim and resourcefulness. This particular triplet of properties could explain the function of what we call motives and goals.

In 1957 Allen Newell, Clifford Shaw and Herbert Simon create General Problem Solver machine also called Difference-Engine. Having a goal mean that a Difference-Engine is actively working to remove those differences.

We have goals because that’s how our brains evolved: the people without goals became extinct because they simply could not compete.

As goals become complex, we need subgoals. In Difference-Engine every difference it needs to reduce becomes another subgoal for it. Person discovers which subgoals are required to accomplish a job either by: trial and error, doing experiment in his head or by recalling some prior experience. When we do it for a long time doing is converted into a script or sequence of actions. The synapses between the cells of the brain become better conductors when they are more used. That is why many achievers become less able to teach others to imitate their techniques.

Marcel Proust: “Each reader reads only what is already inside himself”.[3]

What should you do when what you have does not exactly match what you need? You will want to find some substitute that is different, but not too dissimilar. Patrick Winston is talking about organizing some bodies of knowledge into what he called Difference-Networks.

Free will is not a process we use to make decision, but one that we use to stop other processes.

We can’t usually say what anything is; we can only describe what something is like. This kind of thinking is important because it helps us to deal with new situations. Reasoning by analogy is our most usual way to deal with problems. New ways to look at things is one of our most powerful commonsense processes.

We can also sometimes disable a process without directly suppressing it, by arousing one that competes with it.

Reasons why only rewarding successful attempts may not be a very good strategy:

  • Reinforcement can lead to rigidity
  • Reinforcement can have bad side effects
  • Papert’s Principle – some of the most crucial steps in our mental growth are based not simply on acquiring new skills, but on developing better high-level resources to help us select which already existing skills to use

While pleasure helps us learn easy things, we must learn to “enjoy” some suffering when it comes to learning things that need larger-scale changes in how we think.


Each person has many different ways to think. We start to think about a goal and then we encounter some obstacles. A new destructive goal or idea or a change in one’s Critics or Ways to Think could gain control of so much of a person’s resources and time that we would seem to be seeing a different mind.

Critic recognize a certain species of Problem Type, then it will activate what we shall call a Selector which tries to start up a set of resources that is has learned is likely to act as a Way to Think that may help in this situation.

K-lines (knowledge line) are structures that our brain could use to construct new mental objects and processes by combining parts of older ones.

Our model of mind should include Selectors and Critics at every of six levels of thinking.

Some useful Ways to Think:

To find some systematic means to classify the ways we try to overcome different types of obstacles. There are two extremes: knowing how and searching extensively. And then we have other Ways to Think that lie between.

  • Reasoning by Analogy
  • Dividing and conquering
  • Reformulating
  • Planning

Other techniques:

  • Simplifying
  • Elevating
  • Changing the subject

Some other ways:

  • Wishful thinking – imagine having unlimited time. If you still can’t envision solving the problem, then you should reformulate it.
  • Self-reflection – ask what makes that problem seem hard.
  • Impersonation – imagine someone better at this and try to do what that person would do.
  • Logical contradiction – try to prove that problem can’t be solved and then look for a flaw in that argument.
  • Logical reasoning – chains of deductions. Be aware of validity of assumptions.
  • External Representations
  • Imagination – what would happen if.
  • Cry for help.
  • Ask for help.
  • Resignation.

Few types of Critics that people use:

  • Innate Reactions and Built-in Alarms
  • Learned Reactive Critics
  • Deliberative Critics
  • Reflective Critics
  • Self-reflective Critics
  • Self-conscious Critics

Critic often has a negative quality. It would be hard to describe our Correctors, Suppressors and Censors without using negative terms like inhibit, prevent or terminate.

We can sometimes improve our Ways to Think by creating higher-level Selectors and Critics that help to reduce the sizes of the searchers we make.

Sigmund Freud and Henri Poincare were among the first to develop ideas about “unconscious” goals and processes. Poincare talks about four stages:

  • Preparation – in order to prepare for problem, we need to clear our minds from other goals.
  • Incubation – considering large numbers of combinations.
  • Revelation – incubation ends when mind can embrace totality of elements while realizing the details.
  • Evaluation – it is safe to trust our intuitions.

For many people, thinking and learning is largely a social activity.

Different parts of each person’s brain are involved with different forms of memory – which are sometimes classified under such names as sensory, episodic, autobiographical, semantic, declarative and procedural. Imagine that all those records are stored in various parts of what we’ll call the “context box”.

There is evidence that it takes hours or days for short-term memories to be converted to longer-term ones.

It is easy to both walk and talk because these use such different sets of resources. It is much harder to both speak and write (or to listen and read).

Every Way to Think will need at least some ability to keep other processes from stopping it. We can do that to some extent by controlling which Critics are active.

It is important that we classify the Types of Problems, that people recognize, the Ways to Think that we develop and how we learn which Ways to Think can help us to deal with each of those different Problem Types.


In 1936 Alan Turing created Universal machine that included all other machines. Today all modern computers user that very same trick of storing descriptions of other machines. In fact, those are just what computer programs are.

If you understand something in only one way, then you scarcely understand it at all, because when something goes wrong, you’ll have no place to go.

We have seen how useful it is to know many different ways to achieve the same goal. However, switching between alternatives could slow us down, unless we had ways to do it rapidly. Term panalogy describes a scheme in which corresponding features of different meaning are connected to the very same parts of just one larger structure.

We mainly react to what we expect and tend to represent the things that we see as though they remain the same as we move around. The architecture of our brain has evolved to have structures that make it easy for us to link every fragment of knowledge we learn to ones that we already know.

A panalogy can serve as a way to use very same structure to server several purposes.

Our short-term memories are limited because they use expensive resources. Transfer to long-term memory process sometimes takes a whole day or more and may require substantial intervals of sleep.

Reasons why the formation of longer-term memories require so much time and processing:

  • Retrieval
  • Credit assignment
  • The Real-Estate Problem for Long-Term Memories
  • Copying Complex Description

As children, we not only learn particular things, but we also acquire new thinking techniques.

A person can learn so quickly and without doing so many repetitions. We use higher-level processes to decide what to learn from each incident.

The quality of our credit assignments could be important aspects of the suitcase of traits that people call intelligence.

When we estimate capabilities of extraordinary personalities, we can see that they have unusual combinations of otherwise common ingredients:

  • They are highly proficient in their fields.
  • They have more than usual self-confidence.
  • They often persist where others would quit.
  • They accumulate more Ways to Think.
  • They habitually think in novel ways.
  • They have better systems for self-control.
  • They reject many popular myths and beliefs.
  • They tend to keep thinking more of the time
  • They excel at explaining what they’ve done.
  • They tend to make better credit assignments.

The investment principle: if you know two different ways to achieve the same goal, you’ll usually start with the method that you know best. Then, over time, that method may gain so much additional strength that you’ll tend to use it exclusively, even if you have been told that the other techniques is the better one.

We humans seem to be almost unique in being able to treat ideas as though they were things or in other words to conceptualize. Knowledge is not composed of ideas that exist as separate entities that float around in some mental world; they also need to be interconnected. Our most familiar way to represent an incident is to recount it as a story or script that depicts a sequence of events in time. Semantic networks are a collection of symbols that are linked by labeled connection links. They are among the most versatile forms of representation, because each connection-link could itself refer to yet another type of representation. The term trans-frame is used to name a pair of representations of the conditions before and after some action was done.

The idea of default assumption could help to explain how you can so quickly access your commonsense knowledge.

K-line can act as a sort of snapshot of a mental state because, when you later activate it, this will put you into s similar state.

Resourceful people find ways to decide (using higher-level strategies) which policy might be best to apply in various kinds of situation.

Our brains need some larger-scale organization for interconnecting our multiple ways to represent knowledge. The simplest such arrangement would be a hierarchical one. Structure goes like that: micronemes, neural networks, K-lines and K-trees, semantic networks, frames, trans-frames and narrative stories.

Although sensations give us occasions to learn, this cannot be what makes us able to learn, because we first must have the additional knowledge that our brains would need, as Kant has said, to produce representations and then to connect them. We must be born with primitive forms of structures like K-lines, frames and semantic networks.

It makes no sense to seek a single best way to represent knowledge.

The Self

We sometimes use Self for the features and traits that distinguish each person from everyone else. Perhaps our most common self-model begins by representing a person as having two parts – namely, a body and a mind.

Instead of asking about our Identity, we should ask which of my models of myself best serves my present purposes.

Why should personal traits exist at all? Here are some possible causes for the appearance of such uniformities:

  • Inborn characteristics
  • Learned characteristics
  • Investment principle
  • Archetypes and self-ideals
  • Self-control

Although our trait-based description is frequently wrong and always incomplete, they help to make things seem simpler and more understandable.

We use words like Me and I to keep us from thinking about what we are. Some other ways in which that Single-Self concept is useful to us in everyday life:

  • A localized body
  • A private mind
  • Explaining our mind
  • Moral responsibility
  • Centralized economy
  • Causal attribution
  • Attention and focus
  • Social relations

What we call pleasure is a suitcase-word for quite a few different processes that we don’t often recognize. If we try to catalog some of the feelings and activities that make concept of pleasure.

  • Satisfaction
  • Exploration
  • Goal-suppression
  • Relief

When we visually perceive the world, we do not just process information; we have a subjective experience of color, shape and depth. Experiences associated with other senses, with bodily sensations, with mental imagery, with emotion and with the stream of conscious thought.

One can never really describe what anything is; one can only describe what that thing is like.

The mind can be portrayed as being based on a scheme that deals with various situations by activating certain sets of resources, so that each such selection will function as a somewhat different Way to Think.

No single way to think should be permitted to gain too much control over the systems we use for remembering.

Our unrivaled human resourcefulness developed over three vastly different scales of times:

  • Genetic endowment
  • Cultural heritage
  • Individual experience

Resourcefulness can usually find alternatives:

  • We have multiple descriptions of things
  • We make memory-records of what we’ve done
  • Whenever one of our Ways to Think fails, we can switch to another
  • We manage to control our minds with all sorts of bribes, incentives and threats
  • We have many different ways to learn and can also learn new ways to learn

[1] In the book on page 7

[2] In the book on page 180

[3] In the book on page 199

You may also like
Bill Schmarzo: The Economics of Data, Analytics and Digital transformation
Introducing AI through ladder approach (Rob Thomas – IBM, published by O’Reilly media)
Thomas H. Davenport: The AI advantage, How to Put the Artificial Intelligence Revolution to Work
Paul R. Daugherty, H. James Wilson: Human + machine: Reimagining Work in the Age of AI