Speaker
Description
Identity management (IM) for multi-object tracking is the problem of evolving a belief state over track–object associations, accounting for mixing events and measurement uncertainties.
In the most general setting the problem state is described by a probability distribution over the $n!$ permutations of $n$ objects, whose exact representation and update are inefficient in classical computation. This forces machine learning methods to rely on approximations like the evolution of low-order marginal probabilities.
We investigate an efficient, unconstrained quantum machine learning approach based on non-abelian Fourier analysis over the symmetric group $S_n$.
Exploiting the efficient scaling of the quantum Fourier transform for $S_n$, we propose an iterative two-step quantum pipeline. The algorithm models identity mixing events by a diffusion step that acts in the spectral domain, smoothing the probability distribution via its spectral decomposition. The resulting state is then conditioned on identity observations through a Bayes update in the anti-transformed space.
We present the group theoretical formalism underlying the algorithm, and provide a first blueprint for the two main sub-routines, including scalability studies. Finally, we discuss the potential of this framework as a novel quantum machine learning approach to scalable multi-object tracking and related data-association tasks.
| Sessions | Quantum Machine Learning: |
|---|---|
| Invited | No |