Thread by @enn_nafnlaus

@ReflexFunds I believe you missed a key aspect of 4D: that it's adding time you covered well, but what's left out is that it's adding time to *3d*.

Neural nets are pattern finders. Current AP is a 2D pattern finder. It has no understanding of 3D geometry; is just knows what sort of patterns

@ReflexFunds ... to look for from all angles an object may present at (to the limits of its training dataset).

There's a common assumption that "a neural net will learn anything, so you just need to throw more compute power and data at it". But this isn't true. 2.5D AP will *never*...

@ReflexFunds understand 3D transformations or physics. It's a pattern finder, not an algorithm finder.

This aversion to adding hard-coded algorithms to neural nets is unjustified. We're born with all sorts of hard-coded "algorithms" in our brains, from face recognition to motion tracking...

@ReflexFunds to counting small numbers. Our pattern-recognition wetware does not need to solve these ubiquitous problems; the algorithms are given.

I firmly believe that AP 4D will no longer be "matching patterns in 2D imaged". i believe it will be creating a best-fit match to 3-space,...

@ReflexFunds pattern matching 3d models, textures, lighting, etc. Hard-coded 3D transformations and a physics engine would be able to compare expectations (across both space and time) vs. calculations for backpropagation.

You would no longer just have a point cloud and labeled 2D blobs....

@ReflexFunds You'd have modeled, textured, moving, transforming 3D scenes of *everything* around the vehicle. This should be not just a step forward, but a tremenous leap. The world would no longer be weird patterns, but rather, coherent scenes.