Search

We introduce an efficient and robust auto-tuning framework for hyperparameter selection in dimension reduction (DR) algorithms, focusing on large-scale datasets and arbitrary performance metrics. By leveraging Bayesian optimization (BO) with a surrogate model, our approach enables efficient hyperparameter selection with multi-objective trade-offs and allows us to perform data-driven sensitivity analysis. By incorporating normalization and subsampling, the proposed framework demonstrates versatility and efficiency, as shown in applications to visualization techniques such as t-SNE and UMAP. We evaluate our results on various synthetic and real-world datasets using multiple quality metrics, providing a robust and efficient solution for hyperparameter selection in DR algorithms., Comment: 20 pages, 16 figures

In this paper we present the multilingual language model BLOOM-zh that features enhanced support for Traditional Chinese. BLOOM-zh has its origins in the open-source BLOOM models presented by BigScience in 2022. Starting from released models, we extended the pre-training of BLOOM by additional 7.4 billion tokens in Traditional Chinese and English covering a variety of domains such as news articles, books, encyclopedias, educational materials as well as spoken language. In order to show the properties of BLOOM-zh, both existing and newly created benchmark scenarios are used for evaluating the performance. BLOOM-zh outperforms its predecessor on most Traditional Chinese benchmarks while maintaining its English capability. We release all our models to the research community.

Accurate estimation of tail probabilities of projections of high-dimensional probability measures is of relevance in high-dimensional statistics and asymptotic geometric analysis. Whereas large deviation principles identify the asymptotic exponential decay rate of probabilities, sharp large deviation estimates also provide the "prefactor" in front of the exponentially decaying term. For fixed $p \in (1,\infty)$, consider independent sequences $(X^{(n,p)})_{n \in \mathbb{N}}$ and $(\Theta^n)_{n \in \mathbb{N}}$ of random vectors with $\Theta^n$ distributed according to the normalized cone measure on the unit $\ell_2^n$ sphere, and $X^{(n,p)}$ distributed according to the normalized cone measure on the unit $\ell_p^n$ sphere. For almost every realization $(\theta^n)_{n\in\mathbb{N}}$ of $(\Theta^n)_{n\in\mathbb{N}}$, (quenched) sharp large deviation estimates are established for suitably normalized (scalar) projections of $X^{(n,p)}$ onto $\theta^n$, that are asymptotically exact (as the dimension $n$ tends to infinity). Furthermore, the case when $(X^{(n,p)})_{n \in \mathbb{N}}$ is replaced with $(\mathscr{X}^{(n,p)})_{n \in \mathbb{N}}$, where $\mathscr{X}^{(n,p)}$ is distributed according to the uniform (or normalized volume) measure on the unit $\ell_p^n$ ball, is also considered. In both cases, in contrast to the (quenched) large deviation rate function, the prefactor exhibits a dependence on the projection directions $(\theta^n)_{n \in\mathbb{N}}$ that encodes additional geometric information that enables one to distinguish between projections of balls and spheres. Moreover, comparison with numerical estimates obtained by direct computation and importance sampling shows that the obtained analytical expressions for tail probabilities provide good approximations even for moderate values of $n$., Comment: 55 pages. This version of the paper is a slightly modified version of v1, with more details added to the proofs of Proposition 5.4 for greater clarity. The analysis of $\ell^n_p$ balls, which was removed in v2, is now added to this version

Consider the projection of an $n$-dimensional random vector onto a random $k_n$-dimensional basis, $k_n \leq n$, drawn uniformly from the Haar measure on the Stiefel manifold of orthonormal $k_n$-frames in $\mathbb{R}^n$, in three different asymptotic regimes as $n \rightarrow \infty$: "constant" ($k_n=k$), "sublinear" ($k_n \rightarrow \infty$ but $k_n/n \rightarrow 0$) and "linear" $k_n/n \rightarrow \lambda$ with $0 < \lambda \le 1$). When the sequence of random vectors satisfies a certain "asymptotic thin shell condition", we establish annealed large deviation principles (LDPs) for the corresponding sequence of random projections in the constant regime, and for the sequence of empirical measures of the coordinates of the random projections in the sublinear and linear regimes. We also establish LDPs for certain scaled $\ell_q$ norms of the random projections in these different regimes. Moreover, we verify our assumptions for various sequences of random vectors of interest, including those distributed according to Gibbs measures with superquadratic interaction potential, or the uniform measure on suitably scaled $\ell_p^n$ balls, for $p \in [1,\infty)$, and generalized Orlicz balls defined via a superquadratic function. Our results complement the central limit theorem for convex sets and related results which are known to hold under a "thin shell" condition. These results also substantially extend existing large deviation results for random projections, which are first, restricted to the setting of measures on $\ell_p^n$ balls, and secondly, limited to univariate LDPs (i.e., in $\mathbb{R}$) involving either the norm of a $k_n$-dimensional projection or the projection of $X^{(n)}$ onto a random one-dimensional subspace. Random projections of high-dimensional random vectors are of interest in a range of fields including asymptotic convex geometry and high-dimensional statistics., Comment: 51 pages. The original version introduces LDPs under Assumption A* (verified for product measures, $\ell_p^n$ balls with $p > 2$, and Gibbs measures). This update verifies Assumption A* for a class of Orlicz balls, establishes general conditions Assumptions A, B, and C (which are verified for $\ell_p^n$ balls with $p \in [1, 2)$), and introduces LDPs for norms in the sublinear regime

In this study, we present a novel ranking model based on learning neighborhood relationships embedded in the index space. Given a query point, conventional approximate nearest neighbor search calculates the distances to the cluster centroids, before ranking the clusters from near to far based on the distances. The data indexed in the top-ranked clusters are retrieved and treated as the nearest neighbor candidates for the query. However, the loss of quantization between the data and cluster centroids will inevitably harm the search accuracy. To address this problem, the proposed model ranks clusters based on their nearest neighbor probabilities rather than the query-centroid distances. The nearest neighbor probabilities are estimated by employing neural networks to characterize the neighborhood relationships, i.e., the density function of nearest neighbors with respect to the query. The proposed probability-based ranking can replace the conventional distance-based ranking for finding candidate clusters, and the predicted probability can be used to determine the data quantity to be retrieved from the candidate cluster. Our experimental results demonstrated that the proposed ranking model could boost the search performance effectively in billion-scale datasets., Comment: This paper was accepted by IEEE Transcations on Pattern Analysis and Machine Intelligence in March 2019