Bees Can Solve Math Problems With Addition and Subtraction

According to a new study published in the journal Science Advances, researchers from Australia and France have shown that bees can perform simple arithmetic, adding and subtracting small numbers by studying color-coded shapes. CNET reports: To test the buzzers' ability to perform arithmetic, the team used a three-chambered maze shaped like a Y, training bees to enter through a hole into a small chamber where they would see their first stimulus: blue or yellow shapes on a plain, grey background. The number of shapes varied between 1 and 5 and the color of the shapes told the bee whether it needed to add one (blue) or subtract one (yellow) from the initial number. The bee then flew into a subsequent chamber which presented both a correct option and an incorrect option. To train the bees, the correct option rewarded the critters with a drop of tasty sugar solution -- a delightful dessert for the bee. On the other hand, selecting the incorrect solution resulted in a nasty drop of quinine -- like a slab of Brussels sprouts slathered in chocolate.

The testing procedure itself focused on 14 bees undergoing four tests of 10 choices. The tests themselves were "non-reinforced," so they didn't receive reward or punishment when selecting their "answers" during testing. Because the bees were subjected to two answers each time, the expectation is that -- purely by chance -- they would select the correct answer 50 percent of the time. But the bees performed significantly better than chance would predict, selecting the correct answer around 65 percent of the time.

Comments

Re:math not needed

[umafuckit]

Memorization of the correct and incorrect answers is all that is needed for the described (too small of a sample size to be considered an) "experiment".

How far /. has fallen...

It is true that memorisation could explain this. You point about the sample size is trickier, though. Firstly there is no magic number that constitutes a large (vs small) sample size. What is suitable depends on the size of the effect, the variance, and the degree to which you want to generalise the results to a wider population. This is often balanced against what is possible. In biology a lot of experiments have a small sample size because of the cost or difficulty in gathering the data. For instance, I just reviewed a paper where the authors have gathered data from just a single subject. However, they gather a vast amount and do a very thorough job. Their work still stands as it is (it's in a sense a methods paper) and given that they aren't targetting a big name journal or over-selling their results I'm going to let the n=1 slide.

In the particular case of this paper, what I find most annoying isn't the n=14 but that their graphs hide the underlying data by displaying them as just bars with a 95% confidence interval for the mean. I would also agree, however, that I don't see why in this case they couldn't have produced a larger sample size. That's not the main issue, IMHO, however.

Re:math not needed

[umafuckit]

This isn't like they are studying the remaining living WWII veterans or Japanese anorexics. They should be able to find some extra BEES to run the tests on. Presumably they have access to a hive, at a minimum.

I agree a larger n would be nice (say, n=30 at least) and I *think* it's likely not too hard to obtain in this case. I would caution, however, that sometimes it's a lot harder than it looks to obtain these data. It could be that n=14 is hard to do.

I used to work in insect neuroscience and I collaborated with people who did experiments of the general sort described in the paper. The issue was of course not finding insects -- we had lots of insects -- the problem we had was that running the experiments was very time consuming and could often fail for unclear reasons. You may get drift in behavioral scores over time, batches of insects that produce suspect results, etc. All sorts of really weird stuff happens with animal behavior and so to get solid results you believe in might require throwing out most of your data (e.g. because variance was weirdly high on some days). After all is said and done your sample size isn't always what you hope for. I've seen really good people work for years and still end up with sample size of less than 10 animals.