Inattentional Blindness is common in humans (see: https://en.wikipedia.org/wiki/Inattentional_blindness ). Could this also be common with machines built with artificial vision?
Presumably what happens to people in the famous Invisible Gorilla experiment, is that an incongruous object is simply filtered out of human perception.
If we wish to interpret this mechanistically, we could hypothesize that a 'gorilla object' is simply not presented by low levels of perception to our higher level pattern recognizers because the lower levels are not biased towards the construction of 'gorilla-like' features in such a context.
The recent Tesla fatality (arising from a failure to distinguish between the sky and a high-sided white truck) could conceivably be considered to be an example of this.
See this AI SE question.
Although there might not conceptually be any sort of inattentional blindness associated with an AI system, there might be cases of partial blindness.
Inattentional blindness could occur to a person due to either over-exhaustion limiting cognitive abilities or overuse of frequent cognitive patterns. Our mind takes short-cuts to prevent processing of too much information -- more than what the mind thinks is necessary. But this sometimes backfires when the minor anomalies are not seen (or rather, perceived). Another form of this could also occur when events occur as part of the peripheral vision while the person concentrates only on the foveal vision.
This doesn't happen to a AI system because:
- Machines are not designed to accidentally break defined rule sets by taking mental short-cuts like humans do.
- Computers, in general, do not have peripheral and foveal visual distinctions.
There may be, however, cases where it cannot be able to capture detail as much as humans can and hence could not perceive what it is actually intended -- partial blindness.
An AI agent is constantly processing its input percept sequence and validating it with its knowledge base and forming action sequence based on the its rule set. It does not make mental shortcuts in terms of perception as humans do (at least as part of its standard definition). So whatever it is good at perceiving, it would be good all throughout the vision it captures.
Yes, it is possible. For instance, if your vision system can only track one object at a time and is currently tracking one, any other object in the scene cannot be tracked. So there is inattentional blindness.
A feature like this could be used in artificial vision system as a means of "graceful degradation" when the available computation power is not enough to allow for the tracking/labelling of all elements of a scene.