Current methods for compressing neural network weights, such as decomposition, pruning, quantization, and channel simulation, often overlook the inherent symmetries within these networks and thus waste bits on encoding redundant information. In this paper, we propose a format based on bits-back coding for storing rotationally symmetric Transformer weights more efficiently than the usual array layout at the same floating-point precision. We evaluate our method on Large Language Models (LLMs) pruned by SliceGPT (Ashkboos et al., 2024) and achieve a 3-5% reduction in total bit usage for free across different model sizes and architectures without impacting model performance within a certain numerical precision. Modern neural networks, particularly Large Language Models (LLMs), typically contain billions of parameters. Therefore, encoding and transmitting these models efficiently is gaining widespread interest. However, these techniques ignore the fact that neural networks typically exhibit symmetries in their weight space. For example, in feedforward networks, applying a random permutation to the neurons in one layer and its inverse to the weights in the subsequent layer leaves the output unchanged. Encoding weights without accounting for these symmetries will lead to suboptimal codelength.

A new approach to the sparse Canonical Correlation Analysis (sCCA)is proposed with the aim of discovering interpretable associations in very high-dimensional multi-view, i.e.observations of multiple sets of variables on the same subjects, problems. Inspired by the sparse PCA approach of Journee et al. (2010), we also show that the sparse CCA formulation, while non-convex, is equivalent to a maximization program of a convex objective over a compact set for which we propose a first-order gradient method. This result helps us reduce the search space drastically to the boundaries of the set. Consequently, we propose a two-step algorithm, where we first infer the sparsity pattern of the canonical directions using our fast algorithm, then we shrink each view, i.e. observations of a set of covariates, to contain observations on the sets of covariates selected in the previous step, and compute their canonical directions via any CCA algorithm. We also introduceDirected Sparse CCA, which is able to find associations which are aligned with a specified experiment design, andMulti-View sCCA which is used to discover associations between multiple sets of covariates. Our simulations establish the superior convergence properties and computational efficiency of our algorithm as well as accuracy in terms of the canonical correlation and its ability to recover the supports of the canonical directions. We study the associations between metabolomics, trasncriptomics and microbiomics in a multi-omic study usingMuLe, which is an R-package that implements our approach, in order to form hypotheses on mechanisms of adaptations of Drosophila Melanogaster to high doses of environmental toxicants, specifically Atrazine, which is a commonly used chemical fertilizer.

We introduce Block Sparse Canonical Correlation Analysis which estimates multiple pairs of canonical directions (together a "block") at once, resulting in significantly improved orthogonality of the sparse directions which, we demonstrate, translates to more interpretable solutions. Our approach builds on the sparse CCA method of (Solari, Brown, and Bickel 2019) in that we also express the bi-convex objective of our block formulation as a concave minimization problem over an orthogonal k-frame in a unit Euclidean ball, which in turn, due to concavity of the objective, is shrunk to a Stiefel manifold, which is optimized via gradient descent algorithm. Our simulations show that our method outperforms existing sCCA algorithms and implementations in terms of computational cost and stability, mainly due to the drastic shrinkage of our search space, and the correlation within and orthogonality between pairs of estimated canonical covariates. Finally, we apply our method, available as an R-package called BLOCCS, to multi-omic data on Lung Squamous Cell Carcinoma(LUSC) obtained via The Cancer Genome Atlas, and demonstrate its capability in capturing meaningful biological associations relevant to the hypothesis under study rather than spurious dominant variations.

Sparse Canonical Correlation Analysis (CCA) has received considerable attention in high-dimensional data analysis to study the relationship between two sets of random variables. However, there has been remarkably little theoretical statistical foundation on sparse CCA in high-dimensional settings despite active methodological and applied research activities. In this paper, we introduce an elementary sufficient and necessary characterization such that the solution of CCA is indeed sparse, propose a computationally efficient procedure, called CAPIT, to estimate the canonical directions, and show that the procedure is rate-optimal under various assumptions on nuisance parameters. The procedure is applied to a breast cancer dataset from The Cancer Genome Atlas project. We identify methylation probes that are associated with genes, which have been previously characterized as prognosis signatures of the metastasis of breast cancer.