Output details

Sparse multinomial kernel discriminant analysis (sMKDA)

Sparse representations underpin the success of kernel machines in machine learning. This paper was the first truly sparse, kernel-based generalization of classical multinomial discriminant analysis. Previous attempts relied on the inherently dense "centring" operation, a deficiency overlooked elsewhere in the community. It is important because the leading kernel machine (SVM), does not generalize to k(>2)-class problems. Compared to the then state-of-art, this algorithm matched or exceeded, on numerous benchmarks, both performance and sparsity of leading algorithms, requiring vastly reduced computational resource.