Is “emotion” ever used in AI?

Is "emotion" ever used in AI?

Psychologists have a lot to say about emotion and it's functional utility for survival - but I've never seen any AI research that uses something resembling "emotion" inside an algorithm. (Yes, there's some work done on trying to classify human emotions, called "emotional intelligence", but that's extrememly different from /using/ emotions within an algorithm) For example, you could imagine that a robot might need fuel and be "very thirsty" - causing it to prioritize different tasks (seeking fuel). Emotions also sometimes don't just focus on objectives/priorities - but categorize how much certain classifications are "projected" into a particular emotions.
For example, maybe a robot that needs fuel might be very "afraid" of going towards cars because it's been hit in the past - while it might be "frustrated" at a container that doesn't open properly. It seems very natural that these things are helpful for survival - and they are likely "hardcoded" in our genes (since some emotions - like sexual attraction - seem to be mostly unchangeable by "nurture") - so I would think they would have a lot of general utility in AI.

• Simple google search on "emotion AI" gives plenty of results including those I've seen during the last several years. – aleck Jan 8 at 22:47
• I very specifically address this in the question. I am aware of "Emotion AI" and this is fundamentally different from what I am asking. – Steven Sagona Jan 8 at 23:52
• To be honest our emotions are fueled by instincts and chemistry mostly, so it's more of a label only. In your example being "afraid" running out of power is just a label of power level being low. Same for us being hungry, etc. – aleck Jan 9 at 2:04
• @aleck, yes emotions are just "labels" for states, but the states that these labels describe are VERY different, and the state evolution is not necessarily symmetric. If I put you in a state associated with "happiness" I doubt you will be so quick to switch with a state associated with "pain," and I doubt that the response to happyness is similar in any way to the response to pain. We have developed a language for these properties in part because of how these experiences very differently effect our behavior - and this is likely not a trivial matter but is to true intelligence. – Steven Sagona May 1 at 22:32

2 Answers

Current Simulation of Emotional Behavior

Emotion is used in AI in very limited ways in leading edge natural language systems. For instance, advanced natural language systems can calculate the probability that a particular segment of speech originates from an angry human. This recognition can be trained using labels from bio-monitors. However, the mental features of a human with soft skills tuned from years of experience with people is not nearly simulated in computers as of this writing.

We will not see computers becoming counselors (as once believed) or directors of movies or courtroom judges or customs officials any time soon. Nonetheless, the processes behind emotion are not entirely undiscovered, and there is definite interest in simulating them in computers. Much of that work is company confidential.

The emergence of emotional sophistication in computers likely to begin in the context of sexuality, primarily because flirtation is powerful and primordial emotional expression will probably be easier to simulate in natural language than higher emotional expressions such as love or chaotic ones like rage. Sexy AI will likely be exploited by what businesses might consider legitimate marketing activity.

It is also going to be exploited by the sex industry. The ethical and moral analysis of sexy AI beyond the scope of the question but will probably gain the attention of public media as it unfolds, and that has already begun on FaceBook, originating from third party attacks using fictitious identities.

The Science of Emotion

Emotion isn't a scientific quantity. From an AI perspective, emotion is a quality an individual might recognize through visual and audio queues, specifically through the natural language and affect of another individual. (Affect is a visual clue about a person's emotional and general mental state.)

An individual can also learn to recognize those clues in her or his self. They can be detected by replaying one's own speech as heard through the ear, by linguistic analysis of thoughts not spoken, or through the detection of muscle tension or vital signs. Those skilled in meditation can detect emotional predecessors closer to their causal centers in the brain and control them more directly before emotions even arise.

In the brain, emotion is not in a single geometric location. We cannot say, "That emotion of compassion comes from this group of neurons in Jerome's brain." We cannot say, "Sheri is angry at this 3D coordinate in her cerebral cortex." Emotions are also not strictly system wide either. An individual can be annoyed without going into rage, leaving most of the brain chemistry and electrical signaling unaffected.

Emotions are not entirely electrical and not entirely chemical. On the electric side, emotional states can occur simultaneously through separate circuit pathway connecting distinct and only distantly related regions of the brain. On the chemical side, there is the synaptic chemistry that is part of the electrical signal pathways. There are also many regional signaling systems using specialized pathways that are neither circulatory (blood) nor primary electrical (neuron) pathways. Serotonin is one of dozens of chemical signaling compounds that operate regionally in this way.

Emotions, being a largely social set of phenomena, should not be characterized as purely Darwinian. Although related to survival, emotional processing and communications impact mate selection and, more generally, social patterns within a community, including altruistic and collaborative activity.

Emotions don't always lead to survival. In some cases, emotional states may lead to death prior to reproduction. One could say that emotional balance and the ability to interact on emotional levels may improve odds of having offspring. Imbalance to the degree of any of hundreds of emotional extremes can lead to childlessness.

Emotional intelligence is different than using emotions within an algorithm, but not extremely so.

Discussion of emotional intelligence is one of many advancements in the concept of intelligence since the formation of one-dimensional conceptions of intelligence. Those nineteenth century conceptions, such as IQ and G-factor are poorly supported by genetic evidence and anthropological theory. Mathematically unproven and naive concepts like general intelligence rest on those one-dimensional concepts.

Emotional intelligence is a form of mental capability related to emotional balance. If a person's cognitive skills are honed with respect to their emotions and the assessment of the emotional states of others, then they have greater emotional intelligence than someone who cannot read the affect and linguistic clues of another and cannot integrate cognitive and emotional skill to balance of their own emotions.

Cybernetic Analysis

The interface between natural emotion and artificial emotion fits within the realm of cybernetics, the conceptual study of the interface between humans and machines. Such interaction is clearly related to both algorithms and topology, two important concepts in AI research and development.

Emotion has an algorithmic context because there is clearly some combination of neurons and chemistry that produce this algorithmic difference between a reactive person and one who has developed emotional intelligence.

    emotion[person] = recognize_emotion[person]
if emotion[person] = anger
be_in_responses(angry)

emotion[person] = recognize_emotion[person]
if emotion[person] = anger
be_in_responses(extra_calm)


The former is reactive and the later exhibits emotional intelligence. The acquisition of the later skill may be cognitive and conscious or it may be intuitive and unconscious. In either case, the actual algorithms at a lower level may be entirely different than those shown above, yet the external behavior of the person as marshaled by the brain is essentially one of those shown.

The plural, algorithms, is used rather than the singular, algorithm, because it is unlikely that a single synchronous algorithm is involved. The brain is a massively parallel processor. Emotional processing is likely best expressed in artificial form as hundreds of thousands of algorithms operating in parallel and forming millions of balances within the system — multidimensional and highly parallel stasis.

This is why emotional recognition and emotional responses are not very sophisticated in computer systems as of this writing. The balances have much social nuance. It may be easier to simulate rational thought than emotional thought.

Desire as a Systemic Behavior

Hunger and thirst may sometimes be called feelings, but they are not strictly emotional. The detection of the need for air, energy, nutrients, and water may stimulate emotional states if the needs are unmet and other emotional states if met. A person may become frustrated and irritable when lacking something essential and confronted with another person's less important agenda. A robot may someday do the same. A person may become elated and generous when all such essentials have recently been made available in surplus. A robot may someday do the same. These relationships are expressed in the question this way.

Emotions also sometimes don't just focus on objectives/priorities — but categorize how much certain classifications are "projected" into a particular emotions.

That statement in the question and its explanation is true in some respects. If a robot that needs fuel but is afraid of passing in front of a moving vehicle because it has been hit in the past can be seen in more than one way.

• Probabilistic risk management based on past experiences
• Fuzzy logic that produces control behavior
• Feeling fear because of past experience

In AI design, these three would be handled in different ways.

• Development of a function that ties visual and auditory information to a model of collision and produces a projected likelihood of injury based each of a number of paths to obtaining fuel
• Rules that relates to travel risk with learned probabilities for each along with rules that relate to energy depletion risk with learned probabilities for each and a fuzzy logic rules engine
• Artificial networks that have no audit trail but simulate reptilian emotional circuitry and simulate base instinct

Maintaining Scientific Perspective

Emotion is not hard coded into the brain circuitry or DNA. The reality is significantly more complex.

The DNA provides parameters to a genetic expression system that leads to protein synthesis that leads to brain structure and function that leads to the ability to learn emotional responses that lead to improved social behavior that may lead to higher probabilities of gene pool survival.

Applying digital system traditions to biological process can be counter productive, like anthropomorphic views of programs. Artificial networks don't actually learn; they converge. Nothing is hard coded into biology because the term code applied to DNA isn't anything like a page of Java or Python code.

It is true that some behavioral predispositions are strong forms of stasis within the course of a species. An organism will normally exhibit a strong desire to acquire resources from the biosphere, such as oxygen, proteins, nutrients, carbs, fats, and water. A robot might replace those with a voltage to use for a charge and lubricants for moving parts. An organism will normally exhibit a string desire to reproduce. A robot might be given a simulation of that recursive process and wish to build another like itself.

These are not hard coded in biology. They form a kind of stasis within a population. Some humans don't want children. Some are hospitalized for anorexia nervosa. Some commit suicide by asphyxiation. The statistical mean produces the behavior of the species, not a fixed behavior identical across individuals within the species.

Nature and Nurture

Nature and nurture are useful umbrella terms for general categories of causality in biology and may have equivalents in future robotic products, but they are broad generalities. There are no nature algorithms or nurture algorithms or algorithms that balance nature and nurture. That is where topology is of paramount conceptual importance.

Topology of Algorithmic Components

There is massive interaction between many systems operating independently in multiple dimensions. The visualization of such interactive structure would look more like the topology of all the web sites in a country than a machine learning block diagram. If somehow coded into one algorithm it is possible that all the silicon from all the sand on earth converted to random access memory (RAM) might be insufficient to hold the code expressing the algorithm. Perhaps not. Perhaps a simplicity underlies the interactive system design of life. Perhaps we'll someday know. Perhaps not.

The elegance in the design of life on earth is that multiple independent processes are tuned by billions of years of trial and error to inter-operate and support complex organic processes with billions of moving parts at a molecular level.

Veins of Interdisciplinary Research

Study of these are important for biology, for bioinformatics, for cognitive science, and for artificial intelligence. Emotional recognition and integration of emotional reaction and control into natural communications is part of this research and development.

• Interesting, thank you. Fuzzy logic that produces "control behavior" seems like it probably most accurately models what I was thinking about when speaking about "emotion" used in the algorithm. But I thought (maybe incorrectly) fuzzy logic was mostly used in sort of "pre-programmed" control systems, as opposed to AI or "learning" systems that regress on some datasets. – Steven Sagona Jan 10 at 22:06
• @StevenSagona — I addressed the last sentence in more detail here: ai.stackexchange.com/questions/9930/… – Douglas Daseeco Jan 12 at 3:26

Not a bad question but we can solve this with a little thought experiment. Consider what it means to be "afraid", or to even "feel". It's a DESIRE for something. That something is what pushes us towards general survival. It forces us to focus on what is important right now. And it's relative to our immediate environment & generalized to our abstract conceptualization.

The difference with modern ai paradigms is that they are very structured/rigid in their objectives. There's no general sense of "okayness" or generalized sense of guidance on what it should do. This would require a radically different approach to AI design & infrastructure.

Being that most companies are trying to make money, there's not a lot to be gained by experimenting with "feeling" machines.