Question about Features versus Locations


#1

This is kind of a suggestion, but it’s mostly a question about the right way to think about features and locations.

As I understand it, the input layer represents a feature in context of its location, in the same way that temporal memory represents a feedforward input in context of the sequence. This means it represents the feature-location pair, as opposed to any cells activating for the feature or location regardless of context.

My question is, why are the features so critical for an object’s identity? For example, two pens with the same shape but different textures or colors are very similar objects. So it seems to me like locations are the primary indicators of objects.

To be clear, I consider each point on an object a different location, so features wouldn’t have any shapes. Are shapes actually features?


#2

It’s possible that there is a “manifold” representation that represents just a set of locations corresponding to an object in the 3D world. This would be really useful as you point out. But clearly we are also able to distinguish by the sensory stimuli, and sensory input will be critical for determining locations.

And when you look at sensory types like vision, the sensory input seems even more essential. It is possible that you still end up with a representation of spatial manifolds lacking sensory information but that isn’t all you need.

One other note, if the locations are object specific then they would represent the sensory input in that location even if the representation was derived as a location. I.e. each object-specific location has one possible sensory representation at that location on the object.

But I agree that the locations that make up the shape (what I called manifold) of the object is a really powerful intermediate representation for object inference.


#3

A simple exercise to clarify:

  1. Close your eyes and hold a metal pen in your hand: you’ll recognize it as a pen.
  2. Close your eyes and hold a plastic pen in your hand: you’ll recognize it as a pen.
  3. Close your eyes and hold a grissino (or bread stick) in your hand: you’ll not recognize it as a pen.

Features are as important as locations. In your example, two pens with different textures are similar objects only if such textures are both textures you recognize as possible ones for a pen; i.e. only if you already experienced a pen with such texture.

Similarly, you wouldn’t care if a dog is white or black to recognize it as a dog; but if it were purple, you would ask yourself: “Is it a dog or an alien?”