A filtered bucket-clustering method for projection onto the simplex and the $$\ell _1$$ ball

Abstract

We propose in this paper a new method processing the projection of an arbitrary size vector onto the probabilistic simplex or the 1 ball. Our method merges two principles. The first one is an original search of the projection using a bucket algorithm. The second one is a filtering, on the fly, of the values that cannot be part of the projection. The combination of these two principles offers a simple and efficient algorithm whose worst-case complexity is linear with respect to the vector size. Furthermore, the proposed algorithm exploits the representation of numeric values in digital computers to define the number of buckets and to accelerate the filtering. Mathematics Subject Classification 49M30 • 65C60 • 65K05 • 90C25 Part of this work has been already published [12].

Mitotic Index Determination on Live Cells From Label-Free Acquired Quantitative Phase Images Using a Supervised Autoencoder

“…Where φ(Z, Y ) is the classification loss in the latent space and ψ( X − X) is the reconstructed loss. We also introduce a constrained regularization loss W 1 ≤ η for features selection [30], [31]. The parameter λ weights the classification loss and the reconstruction loss.…”

Learning sparse deep neural networks using efficient structured projections on convex constraints for green AI

“…Classical Learning structured sparse DNNs are based on proximal regularization methods. In this paper, we propose an alternative constrained approach that takes advantage of available efficient projections on the 1 -ball [47], [48], [49] and on the 2,1 ball [50], [51]. We further propose a new 1,1 projection.…”

“…In this work, we propose a constrained approach in which the constraint is directly related to the number of zero-weights. Moreover it takes advantage of an available efficient projection on the 1 -ball [47], [49]. Let L(W ) be a gradient Lipschitz loss, R(w) be a convex constraint, and C its convex set.…”

“…In the case of the constraint R(w) = W 1 , efficient algorithms have been proposed [47], [49]. Unfortunately this 1 constraint does not induce a sparse structure.…”

From the simplex to the sphere: Faster constrained optimization using the Hadamard parametrization

“…Theoretically, P ∆n (x) can be computed with worst case complexity O(n). However, this involves either (i) using a linear time median selection algorithm Blum et al (1973) known to be slow in practice; or (ii) using the recently proposed algorithm of Perez et al (2020), which requires a priori knowledge (i.e. a bound on the size of the entries of x) as well as some non-standard floating point operations.…”