|
|
|
|
|
|
by colemannugent
1599 days ago
|
|
|
I'd be interested in seeing if there were any "attributes" that they could not easily explain. I suppose that a well trained model shouldn't have too many of these, but it would be good for things like detecting intentional manipulation of training data. This also makes it way easier to fool classifiers by exploiting known high-impact attributes. |
|
|
I am sure there are many, but the more obscure the less they probably contribute to class probability.
Sitting down is probably an indicator of old age because 90 year olds are sitting for more hours than a 20 year old. It probably does not contribute a lot but may contribute something.
More obscure might be a color of shirt that is more popular with 90 year old men in the USA than men under 50.