Analogy to Travelling Salesperson Problem:
-
Hopfield-Tank use h(x) with x = (x,t) with t labelling tour and x labelling city
-
Two critical differences
-
TSP: Only 1 tour so Hopfield-Tank very redundant and must satisfy difficult constraint
-
Vision: Many tours and indeed unknown number of tours, so less redundancy and no constraints!
|
|
TRW have implemented this approach to tracking
|
|
Neural network degrees of freedom independent of number of tracks! (Kalman filter gets difficult as number of tracks becomes large and unknown)
|
|
Neural network has essentially unlimited parallelism
|
|
Again neural networks "work" when these are natural degrees of freedom
|
