What bees can learn

Why bees can do any of this

A forager has a hard job. She has to find flowers among a great many things that are not flowers, learn which ones pay, handle each species' particular mechanics, remember where the patch is, get home, and communicate it. Every one of those is a learning problem, and natural selection has been solving them for a hundred million years. The cognition is not a curiosity bolted onto a simple animal; it is what the animal is for.

That is also why the experiments work. A bee will do almost anything for sugar water, and it learns fast enough that a whole discrimination task fits inside an afternoon. Bees are a model organism for learning research because they are convenient, not because they are exceptional.

What is well established

• Associative learning of colour, shape, pattern and scent, often within a handful of rewarded visits, and retained for days.
• Generalisation to novel stimuli that share the trained property, which is what separates learning a rule from memorising an image.
• Relational concepts: bees trained on above-and-below, or on same-and-different, transfer the rule to stimuli they have never seen, and in some designs across sensory modalities.
• Configural processing: they can learn an arrangement of features rather than the features alone.
• Route and landmark memory good enough to navigate several kilometres and return to one entrance among many.
• Social learning: a bee that watches another bee solve a task solves it faster, and the solution spreads through a group.

The relational concept work is the strongest result in the field and the least famous. Learning that two stimuli are the same regardless of what they are made of is an abstraction, and it was not expected in an insect.

What is real but contested

• Quantity discrimination. Bees reliably choose between small numbers of elements. Whether the cue is number, or area, edge length or density that happen to correlate with it, is exactly what the controls are arguing about.
• Zero-like ordering. One influential study reports bees placing an empty stimulus at the low end of a learned quantity series. A published critique argues the effect follows from continuous visual properties, a published reply disputes how the critique measured them, and nobody outside the two groups has replicated it.
• Arithmetic. A single study by one group reports symbolic addition and subtraction, with no independent replication and no formal rebuttal aimed at it. This site does not report its conclusion as a fact.
• Emotion-like states. Agitated bees judge ambiguous stimuli more pessimistically and rewarded bumblebees more optimistically, with a demonstrated role for dopamine. These are judgement-bias results, which is a genuine internal-state finding and not a finding about feelings.
• Play. Bumblebees roll balls with no reward, more when young. It meets several published criteria for play, and whether it is enjoyed is not a question the experiment can address.

The verbs that do the damage

How big a bee brain is

Around a million neurons, in a structure a cubic millimetre or so in volume. The mushroom bodies — paired centres receiving visual and olfactory input — are where learning and memory are concentrated, and they change measurably in structure as a bee moves from in-hive work to foraging.

The interesting comparison is not with a human brain but with what a million neurons is enough for. The answer, on the current evidence, is a great deal more than anyone assumed, and considerably less than the popular framing claims.