Machine Loving
Searching for a Definition of Artificial Emotional Intelligence
With Valentine’s day approaching, I felt an urge to write something in the name of romance. A love poem, you say? A flashback narrative of a romantic relationship, perhaps? Staying true to my nature as a computer scientist, I nevertheless decided to do some research into the origins of machine learning and answer this: how would a computer scientist go about defining the ability of machines to love?
Computer scientists are people of precision and coherence. They know that, when confronted with abstract concepts, the first thing to do is to come up with practical definitions.
Take the story of Artificial Intelligence (AI) for example. The Wikipedia entry for AI states “ […] AI is […] the capacity for logic, understanding, self-awareness, learning, emotional knowledge, reasoning, planning, creativity, critical thinking, and problem solving.” To define this abstract concept, we resorted to using more abstract concepts.
For computer scientists, AI was born as a term and a field at the Dartmouth workshop in 1956, where participants agreed that any aspect of learning can be so precisely described that a machine can be made to simulate it.
Even that definition was too vague however. Machine learning, the down-to-earth offspring of an all-over-the-place AI, was soon coined as a term in 1959. Depending on the application, machine learning can refer to the ability of algorithms to find patterns in large amounts of data, predict future behaviour based on the past or adapt to their environment by interacting with it.
Its mostly cited definition was given by Tom Mitchell:
A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.
Computer scientists have often been criticised for this mechanistic attitude towards the learning process. Lately, the community has embraced the idea that AI may be an unattainable objective. But is that even bad? AI has admittedly made for a great carrot stick that motivates researchers and practitioners to create many useful applications.
Could the same approach work in our attempt to define artificial emotional intelligence?
When talking about machines and love, people often focus on the question:
Can we, humans, fall in love with a machine?
Movies such as Her and Blade Runner can fuel this discussion, but the answer here is, perhaps surprisingly, easy. YES. (Most) humans have the capacity to develop empathy and love for others. Experiments of robots being tortured helped us realise that it only takes for some remotely humanlike traits, to make people develop feelings for robots. The question we are flirting with here, is whether robots can (appear to) have emotional intelligence, not whether they can be the recipients of it.
According to the Wikipedia entry again, Emotional Intelligence (EI) is the “capability of individuals to recognise their own emotions and those of others, discern between different feelings and label them appropriately, use emotional information to guide thinking and behavior, and manage and/or adjust emotions to adapt to environments or achieve one’s goal(s).”
EI seems to include two abilities that lie at the core of current machine learning techniques: recognition and adaptation. The former can be achieved using classification algorithms, which, given data about emotions and feelings, can learn to recognise them. Adaptation is part of reinforcement learning, where agents set goals and are rewarded when they satisfy them. Could it be that EI is not as vague as we initially thought?
A core concept in EI is empathy, which can be broken down into two very distinct types. Cognitive empathy is the ability of imagining someone’s feelings but not actually feeling them yourself, while emotional empathy is the ability of actually feeling others’ feelings. Cognitive empathy seems to be the objective we were looking for.
In a contest between abstract concepts that allow for a high degree of subjective interpretation, love would probably be among the finalists. Love is believed to involve some level of empathy, but empathy doesn’t always equal love.
In our attempt to find a definition for the ability of machines to love, we observe the following: learning is defined based on the ability of computer programs to improve. Mitchell and his contemporaries did not aim for IQ tests, ex post evaluations of whether a computer program knows stuff. Measuring improvement of an ability is a more tangible objective than measuring the absolute ability.
For this reason, our definition will try to capture the ability of falling in love:
A computer program is said to fall in love with an object O based on their interaction I, if the quality of its relationship with O, as measured by a quality measure Q, improves with interaction I.
Mechanistic enough, right? Let’s see how we could make a well-defined machine loving problem using our definition:
Test whether a program is falling in love with a director
Object O: the director
Interaction I: watching the director’s movies
Quality measure Q: number of hours the program can discuss about them
It shouldn’t be surprising that Artificial EI is already a scientific field that has given us a number of applications. From advertising to mental health support, automation can help us augment current practises and perhaps decode human psyche.
As we evolve, new types of intelligence enter our field of view. A careful look into history, dictionaries and Wikipedia entries, can show us that, any type of, intelligence, is usually described according to some objective. Should we perhaps excuse computer scientists for taking philosophy and romance out of their equations?
Applied Data Science Partners is a London based consultancy that implements end-to-end data science solutions for businesses, delivering measurable value. If you’re looking to do more with your data, please get in touch via our website. Follow us on LinkedIn for more AI and data science stories!
