Martinetz et al. (1993) — DOI:10.1109/72.238311
Key takeaways
Introduced the Neural Gas algorithm: an unsupervised competitive learning method for vector quantization that adapts reference vectors by ranking all neurons by distance to each input and applying exponentially decaying updates by rank, with no fixed grid topology. Proved convergence and showed superiority over k-means and LVQ on time-series prediction benchmarks.
- Convergence theorem. Neural gas converges to a local minimum of the expected distortion E[||x - w_{s(x)}||^2] under standard stochastic approximation conditions on the learning rate schedule Holds when: Requires epsilon(t) -> 0, sum epsilon(t) = inf, sum epsilon(t)^2 < inf; lambda(t) -> 0
Standing in the anthology
Read — the reading is NOTE-125. Arrived in the imported batch, which brought in the competitive-learning and vector-quantization line.