Kronecker-sparse (KS) matrices—whose supports are Kronecker products of identity and all-ones blocks—underpin the structure of Butterfly and Monarch matrices and offer the promise of more efficient models. However, existing GPU kernels for KS matrix multiplication suffer from high data movement costs, with up to 50% of time spent on memory-bound tensor permutations. We propose a fused, output-stationary GPU kernel that eliminates these overheads, reducing global memory traffic threefold. Across 600 KS patterns, our kernel achieves in FP32 a median speedup of x1.4 and lowers energy consumption by 15%. A simple heuristic based on KS pattern parameters predicts when our method outperforms existing ones. We release all code at github.com/PascalCarrivain/ksmm, including a PyTorch-compatible KSLinear layer, and demonstrate in FP32 end-to-end latency reductions of up to 22% in ViT-S/16 and 16% in GPT-2 medium.

In this paper, we study the asymptotic nonnegative rank of matrices, which characterizes the asymptotic growth of the nonnegative rank of fixed nonnegative matrices under the Kronecker product. This quantity is important since it governs several notions in information theory such as the so-called exact Renyi common information and the amortized communication complexity. By using the theory of asymptotic spectra of V. Strassen (J. Reine Angew. Math. 1988), we define formally the asymptotic spectrum of nonnegative matrices and give a dual characterization of the asymptotic nonnegative rank. Comple-mentary to the nonnegative rank, we introduce the notion of the sub rank of a nonnegative matrix and show that it is exactly equal to the size of the maximum induced matching of the bipartite graph defined on the support of the matrix (therefore, independent of the value of entries). Finally, we show that two matrix parameters, namely rank and fractional cover number, belong to the asymptotic spectrum of nonnegative matrices.

Given a training set, a loss function, and a neural network architecture, it is often taken for granted that optimal network parameters exist, and a common practice is to apply available optimization algorithms to search for them. In this work, we show that the existence of an optimal solution is not always guaranteed, especially in the context of sparse ReLU neural networks. In particular, we first show that optimization problems involving deep networks with certain sparsity patterns do not always have optimal parameters, and that optimization algorithms may then diverge. Via a new topological relation between sparse ReLU neural networks and their linear counterparts, we derive --using existing tools from real algebraic geometry-- an algorithm to verify that a given sparsity pattern suffers from this issue. Then, the existence of a global optimum is proved for every concrete optimization problem involving a shallow sparse ReLU neural network of output dimension one. Overall, the analysis is based on the investigation of two topological properties of the space of functions implementable as sparse ReLU neural networks: a best approximation property, and a closedness property, both in the uniform norm. This is studied both for (finite) domains corresponding to practical training on finite training sets, and for more general domains such as the unit cube. This allows us to provide conditions for the guaranteed existence of an optimum given a sparsity pattern. The results apply not only to several sparsity patterns proposed in recent works on network pruning/sparsification, but also to classical dense neural networks, including architectures not covered by existing results.

We revisit the problem of finding small ε-separation keys introduced by Motwani and Xu (2008). In this problem, the input is a data set consisting of m-dimensional tuples {x1,x2,...,xn}. The goal is to find a small subset of coordinates that separates at least (1-ε)(n2) pairs of tuples. When n is large, they provided a fast algorithm that runs on Θ(m/ε) tuples sampled uniformly at random. We show that the sample size can be improved to Θ(m/√ε). Our algorithm also enjoys a faster running time.

Approximating a matrix by a product of few sparse factors whose supports possess the butterfly structure, which is common to many fast transforms, is key to learn fast transforms and speed up algorithms for inverse problems. We introduce a hierarchical approach that recursively factorizes the considered matrix into two factors. Using recent advances on the well-posedness and tractability of the two-factor fixed- support sparse matrix factorization problem, the proposed algorithm is endowed with exact recovery guarantees. Experiments show that speed and accuracy of the factorization can be jointly improved by several orders of magnitude, compared to gradient-based optimization methods.

Matrix factorization with sparsity constraints plays an important role in many machine learning and signal processing problems such as dictionary learning, data visualization, dimension reduction. Among the most popular tools for sparse matrix factorization are proximal algorithms, a family of algorithms based on proximal operators. In this paper, we address two problems with the application of proximal algorithms to sparse matrix factorization. On the one hand, we analyze a weakness of proximal algorithms in sparse matrix factorization: the premature convergence of the support. A remedy is also proposed to address this problem. On the other hand, we describe a new tractable proximal operator called Generalized Hungarian Method, associated to so-called k-regular matrices, which are useful for the factorization of a class of matrices associated to fast linear transforms. We further illustrate the effectiveness of our proposals by numerical experiments on the Hadamard Transform and magnetoencephalography matrix factorization.