What are concepts?

Is it possible to learn just one concept? E.g. suppose I bring a porcupine into the lab, point at it and say 'This is a porcupine'. What is Fido supposed to learn from that?

Will it regard 'porcupine' as a label for what we would describe as the concept:

• physical object
• something Aaron has just brought in
• animal
• mammal
• rodent
• thing with two eyes
• quadruped
• small bundle
• prickly-object
• thing pointed at

For any new concept to be learnt there must already be in place a framework into which it fits, which will be different for learning different sorts of concepts, e.g.

• the concept of a colour e.g. ultramarine
• the concept of a kind of fish e.g. carp
• the concept of a species
• the concept of a sort of living thing, e.g. a fungus
• the concept of a type of fungus, e.g. bracket-fungus
• the concept of a relation, e.g. between
• the concept of a type of action, e.g. pushing, giving
• the concept of a type of mental state, e.g. disappointment
• the concept of a type of social interaction e.g. promising

So there must be something about the context of learning which enables the learner to work out (not necessarily consciously) which of the many possible kinds of learning should be produced in a particular learning episode.

Of course this fact is totally obscured by learning machines that are only capable of learning one sort of thing because that is all they are built for! In such cases it is very misleading (even a misuse of ordinary language) to describe them as learning concepts: they are merely being given a new association between for instance a sensory input pattern and a verbal label (heard, uttered, or possibly typed).

That has very little to do with learning concepts in humans or animals, which typically involves some new processing capability that is relevant to actions that may or may not achieve some goal or more generally be used in some purposive decision-making or action selection process. This can also include internal processes, such as formulating a hypothesis that all porcupines are small enough to carry, or pondering a question such as how big do porcupines grow? (Which assumes that the learner already has a rich framework of biological knowledge.)

For all these reasons we need to be very careful about designing scenarios that we claim demonstrate concept formation, or concept learning, whether instigated by a tutor or (more profoundly) resulting from purposive interactions with the environment that produce something novel.

Good work in this area will presuppose a survey of different kinds of learning, how they fit into different parts of the architecture, which sorts of information-processing capabilities and behavioural capabilities they alter, what sorts of mechanisms produce those changes, whether learning can include mistakes, how the mistakes are detected and remedied, whether there are constraints regarding the order in which certain sorts of things can be learnt, how different sorts of architectures, different forms of representations, different sorts of algorithms, different sorts of sensory and motor capabilities constrain what is learnt, whether learning of some new concepts can happen in one part of the system while another part gets no benefit (e.g. unconsciously learning to distinguish kinds of perceptual patterns relevant to posture control, without being able to think about or talk about those patterns).