Do i really look like a monkey: Unexpected image classification results?

I have a question on vision prediction. I trained a vision model to classify 10 types of monkeys

https://www.kaggle.com/tauvic/fast-monkeys-know-your-customer

learn = cnn_learner(dls, resnet34, metrics=error_rate)

When i tested the model i also used a picture of myself and a 100% black image.
The results:
Myself: Prediction: white_headed_capuchin; Probability: 0.9944
100% black image: Prediction: bald_uakari; Probability: 0.4602

So its good i did not end up as a bald uakari. But im not amused.
Im sure that i don’t look like a monkey and expected a very low probability.

So is this to be expected? Does this type of model predicts that everything it has not seen before must be a monkey? What can i do to get a better result, (low probability prediction of being monkey)

Hi @TauvicR, there is a whole topic about multilabel classification. You might want to have a look at it :

In general, one problem is probably that your model does not know what to do with such different image if it has always seen monkeys.

2 Likes

There is also this thread that discusses how easy it can be in the API:

2 Likes

And this thread also discussing how to handle data from classes not previously encountered in training.

1 Like

Thanks for the super fast reply. I see its a known problem with many solutions. Thank i have to read and im sure it will show me where to go. I sort of expected the fast.ai library to help me out and make probability 0.0 for non-trained objects. But now i see that is not the case.

Might be good to put this in the course, or i have read over it.

Happy that im not a monkey afterall.

Hi TauvicR Hope all is well!

The fastai book covers this subject in some detail.

In addition to the posts above I would look at the ethics notebook https://github.com/fastai/fastbook/blob/master/03_ethics.ipynb where you can see plenty of examples of your precise experience from major companies, and https://github.com/fastai/fastbook/blob/master/02_production.ipynb

Jeremy gives an example of a skin classifier with unexpected consequences.

Also in above notebook Jeremy talks about Avoiding disaster by having a deployment process that helps avoid deploying a model such as your bear classifier, that if deployed on an app might possibly upset a few people :cry: :cry:.

27. What are the three steps in the deployment process?

Out-of-domain data and domain shift are examples of a larger problem: that you can never fully understand the entire behaviour of your neural network. They have far too many parameters to be able to analytically understand all of their possible behaviors. This is the natural downside of their best feature—their flexibility, which enables them to solve complex problems where we may not even be able to fully specify our preferred solution approaches. The good news, however, is that there are ways to mitigate these risks using a carefully thought-out process. The details of this will vary depending on the details of the problem you are solving, but we will attempt to lay out here a high-level approach, summarized in

As part of my pipeline | go through the concepts laid out in these two chapters for every model I create.

Because if we the engineers, developers and scientists get it wrong, what chance have the general public.

Cheers mrfabulous1 :smiley: :smiley: