Our brains see affordances, not because we see 3D objects, but rather how we move to see our 3D worlds. This is different from inverse graphics.
Inverse graphics implies that a representation of a 3D object is created in our minds. If we assume the 'lazy brain hypothesis' then this doesn't make sense because it is wasteful computation.
When we look at a Rubik's cube, we don't instantly know where which colors are on which sides. We are aware of the layout only when we attend to it. It is an active form of perception rather than a photographic snapshot.
We see the whole first before we see the parts. We don't see it the other way around. That is, parts creating wholes. Deep Learning methods do this, but not human brains.
The adversarial inputs (i.e. visual illusions) that do fool humans differ from the kinds found in Deep Learning. The human kind fools us in how we 'feel' the image. Top-down re-construction rather than bottom-up.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
