Hacker News new | past | comments | ask | show | jobs | submit login

I start to like this way of looking for false positives or false negatives more and more.

It would be interesting to introduce some kind of aspects known from the human brain and see if the misclassified items "move" in some conceptually understandable direction.

* Introduce time. Humans are not just image classifiers; humans are able to recognize objects in visual streams of images. Such streams can be seen as latent variables that introduce correlations over time as well as space. What constitutes spatial noise might very well be influenced in our brains by the temporal correlations we see as well.

* Introduce saccades. A computer is only able to see a picture from one viewpoint. Our eyes undergo saccades and microsaccades. That's an unfair advantage for us, being able to see a picture multiple times from different directions!

* Introduce the body. We can move around an object. This again introduces correlations that 1.) are available to us, and 2.) might define priors even when we are not able to move around the picture. In other words, we can (unconsciously) rotate things in our head.




You might be interested in this paper, if you haven't seen it already: http://arxiv.org/abs/1406.6247




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: