The problem of efficient approximation of a linear operator induced by the Gaussian or softmax kernel is often addressed using random features (RFs) which yield an unbiased approximation of the operator's result. Such operators emerge in important applications ranging from kernel methods to efficient Transformers. We propose parameterized, positive, non-trigonometric RFs which approximate Gaussian and softmax-kernels. In contrast to traditional RF approximations, parameters of these new methods can be optimized to reduce the variance of the approximation, and the optimum can be expressed in closed form. We show that our methods lead to variance reduction in practice ($e^{10}$-times smaller variance and beyond) and outperform previous methods in a kernel regression task. Using our proposed mechanism, we also present FAVOR#, a method for self-attention approximation in Transformers. We show that FAVOR# outperforms other random feature methods in speech modelling and natural language processing.
Chromosome analysis is essential for diagnosing genetic disorders. For hematologic malignancies, identification of somatic clonal aberrations by karyotype analysis remains the standard of care. However, karyotyping is costly and time-consuming because of the largely manual process and the expertise required in identifying and annotating aberrations. Efforts to automate karyotype analysis to date fell short in aberration detection. Using a training set of ~10k patient specimens and ~50k karyograms from over 5 years from the Fred Hutchinson Cancer Center, we created a labeled set of images representing individual chromosomes. These individual chromosomes were used to train and assess deep learning models for classifying the 24 human chromosomes and identifying chromosomal aberrations. The top-accuracy models utilized the recently introduced Topological Vision Transformers (TopViTs) with 2-level-block-Toeplitz masking, to incorporate structural inductive bias. TopViT outperformed CNN (Inception) models with >99.3% accuracy for chromosome identification, and exhibited accuracies >99% for aberration detection in most aberrations. Notably, we were able to show high-quality performance even in "few shot" learning scenarios. Incorporating the definition of clonality substantially improved both precision and recall (sensitivity). When applied to "zero shot" scenarios, the model captured aberrations without training, with perfect precision at >50% recall. Together these results show that modern deep learning models can approach expert-level performance for chromosome aberration detection. To our knowledge, this is the first study demonstrating the downstream effectiveness of TopViTs. These results open up exciting opportunities for not only expediting patient results but providing a scalable technology for early screening of low-abundance chromosomal lesions.
We revisit the problem of learning mixtures of spherical Gaussians. Given samples from mixture $\frac{1}{k}\sum_{j=1}^{k}\mathcal{N}(\mu_j, I_d)$, the goal is to estimate the means $\mu_1, \mu_2, \ldots, \mu_k \in \mathbb{R}^d$ up to a small error. The hardness of this learning problem can be measured by the separation $\Delta$ defined as the minimum distance between all pairs of means. Regev and Vijayaraghavan (2017) showed that with $\Delta = \Omega(\sqrt{\log k})$ separation, the means can be learned using $\mathrm{poly}(k, d)$ samples, whereas super-polynomially many samples are required if $\Delta = o(\sqrt{\log k})$ and $d = \Omega(\log k)$. This leaves open the low-dimensional regime where $d = o(\log k)$. In this work, we give an algorithm that efficiently learns the means in $d = O(\log k/\log\log k)$ dimensions under separation $d/\sqrt{\log k}$ (modulo doubly logarithmic factors). This separation is strictly smaller than $\sqrt{\log k}$, and is also shown to be necessary. Along with the results of Regev and Vijayaraghavan (2017), our work almost pins down the critical separation threshold at which efficient parameter learning becomes possible for spherical Gaussian mixtures. More generally, our algorithm runs in time $\mathrm{poly}(k)\cdot f(d, \Delta, \epsilon)$, and is thus fixed-parameter tractable in parameters $d$, $\Delta$ and $\epsilon$. Our approach is based on estimating the Fourier transform of the mixture at carefully chosen frequencies, and both the algorithm and its analysis are simple and elementary. Our positive results can be easily extended to learning mixtures of non-Gaussian distributions, under a mild condition on the Fourier spectrum of the distribution.
Opinion summarization is the task of creating summaries capturing popular opinions from user reviews. In this paper, we introduce Geodesic Summarizer (GeoSumm), a novel system to perform unsupervised extractive opinion summarization. GeoSumm involves an encoder-decoder based representation learning model, that generates representations of text as a distribution over latent semantic units. GeoSumm generates these representations by performing dictionary learning over pre-trained text representations at multiple decoder layers. We then use these representations to quantify the relevance of review sentences using a novel approximate geodesic distance based scoring mechanism. We use the relevance scores to identify popular opinions in order to compose general and aspect-specific summaries. Our proposed model, GeoSumm, achieves state-of-the-art performance on three opinion summarization datasets. We perform additional experiments to analyze the functioning of our model and showcase the generalization ability of {\X} across different domains.
We introduce chefs' random tables (CRTs), a new class of non-trigonometric random features (RFs) to approximate Gaussian and softmax kernels. CRTs are an alternative to standard random kitchen sink (RKS) methods, which inherently rely on the trigonometric maps. We present variants of CRTs where RFs are positive, a key requirement for applications in recent low-rank Transformers. Further variance reduction is possible by leveraging statistics which are simple to compute. One instantiation of CRTs, the optimal positive random features (OPRFs), is to our knowledge the first RF method for unbiased softmax kernel estimation with positive and bounded RFs, resulting in exponentially small tails and much lower variance than its counterparts. As we show, orthogonal random features applied in OPRFs provide additional variance reduction for any dimensionality $d$ (not only asymptotically for sufficiently large $d$, as for RKS). We test CRTs on many tasks ranging from non-parametric classification to training Transformers for text, speech and image data, obtaining new state-of-the-art results for low-rank text Transformers, while providing linear space and time complexity.
In this work we introduce KERNELIZED TRANSFORMER, a generic, scalable, data driven framework for learning the kernel function in Transformers. Our framework approximates the Transformer kernel as a dot product between spectral feature maps and learns the kernel by learning the spectral distribution. This not only helps in learning a generic kernel end-to-end, but also reduces the time and space complexity of Transformers from quadratic to linear. We show that KERNELIZED TRANSFORMERS achieve performance comparable to existing efficient Transformer architectures, both in terms of accuracy as well as computational efficiency. Our study also demonstrates that the choice of the kernel has a substantial impact on performance, and kernel learning variants are competitive alternatives to fixed kernel Transformers, both in long as well as short sequence tasks.
Hierarchical clustering is a critical task in numerous domains. Many approaches are based on heuristics and the properties of the resulting clusterings are studied post hoc. However, in several applications, there is a natural cost function that can be used to characterize the quality of the clustering. In those cases, hierarchical clustering can be seen as a combinatorial optimization problem. To that end, we introduce a new approach based on A* search. We overcome the prohibitively large search space by combining A* with a novel \emph{trellis} data structure. This combination results in an exact algorithm that scales beyond previous state of the art, from a search space with $10^{12}$ trees to $10^{15}$ trees, and an approximate algorithm that improves over baselines, even in enormous search spaces that contain more than $10^{1000}$ trees. We empirically demonstrate that our method achieves substantially higher quality results than baselines for a particle physics use case and other clustering benchmarks. We describe how our method provides significantly improved theoretical bounds on the time and space complexity of A* for clustering.
Bottom-up algorithms such as the classic hierarchical agglomerative clustering, are highly effective for hierarchical as well as flat clustering. However, the large number of rounds and their sequential nature limit the scalability of agglomerative clustering. In this paper, we present an alternative round-based bottom-up hierarchical clustering, the Sub-Cluster Component Algorithm (SCC), that scales gracefully to massive datasets. Our method builds many sub-clusters in parallel in a given round and requires many fewer rounds -- usually an order of magnitude smaller than classic agglomerative clustering. Our theoretical analysis shows that, under a modest separability assumption, SCC will contain the optimal flat clustering. SCC also provides a 2-approx solution to the DP-means objective, thereby introducing a novel application of hierarchical clustering methods. Empirically, SCC finds better hierarchies and flat clusterings even when the data does not satisfy the separability assumption. We demonstrate the scalability of our method by applying it to a dataset of 30 billion points and showing that SCC produces higher quality clusterings than the state-of-the-art.
Transformers-based models, such as BERT, have been one of the most successful deep learning models for NLP. Unfortunately, one of their core limitations is the quadratic dependency (mainly in terms of memory) on the sequence length due to their full attention mechanism. To remedy this, we propose, BigBird, a sparse attention mechanism that reduces this quadratic dependency to linear. We show that BigBird is a universal approximator of sequence functions and is Turing complete, thereby preserving these properties of the quadratic, full attention model. Along the way, our theoretical analysis reveals some of the benefits of having $O(1)$ global tokens (such as CLS), that attend to the entire sequence as part of the sparse attention mechanism. The proposed sparse attention can handle sequences of length up to 8x of what was previously possible using similar hardware. As a consequence of the capability to handle longer context, BigBird drastically improves performance on various NLP tasks such as question answering and summarization. We also propose novel applications to genomics data.