With the tremendous success of large transformer models in natural language understanding, down-sizing them for cost-effective deployments has become critical. Recent studies have explored the low-rank weight factorization techniques which are efficient to train, and apply out-of-the-box to any transformer architecture. Unfortunately, the low-rank assumption tends to be over-restrictive and hinders the expressiveness of the compressed model. This paper proposes, DSFormer, a simple alternative factorization scheme which expresses a target weight matrix as the product of a small dense and a semi-structured sparse matrix. The resulting approximation is more faithful to the weight distribution in transformers and therefore achieves a stronger efficiency-accuracy trade-off. Another concern with existing factorizers is their dependence on a task-unaware initialization step which degrades the accuracy of the resulting model. DSFormer addresses this issue through a novel Straight-Through Factorizer (STF) algorithm that jointly learns all the weight factorizations to directly maximize the final task accuracy. Extensive experiments on multiple natural language understanding benchmarks demonstrate that DSFormer obtains up to 40% better compression than the state-of-the-art low-rank factorizers, leading semi-structured sparsity baselines and popular knowledge distillation approaches. Our approach is also orthogonal to mainstream compressors and offers up to 50% additional compression when added to popular distilled, layer-shared and quantized transformers. We empirically evaluate the benefits of STF over conventional optimization practices.
This paper introduces a new learning paradigm called eXtreme Regression (XR) whose objective is to accurately predict the numerical degrees of relevance of an extremely large number of labels to a data point. XR can provide elegant solutions to many large-scale ranking and recommendation applications including Dynamic Search Advertising (DSA). XR can learn more accurate models than the recently popular extreme classifiers which incorrectly assume strictly binary-valued label relevances. Traditional regression metrics which sum the errors over all the labels are unsuitable for XR problems since they could give extremely loose bounds for the label ranking quality. Also, the existing regression algorithms won't efficiently scale to millions of labels. This paper addresses these limitations through: (1) new evaluation metrics for XR which sum only the k largest regression errors; (2) a new algorithm called XReg which decomposes XR task into a hierarchy of much smaller regression problems thus leading to highly efficient training and prediction. This paper also introduces a (3) new labelwise prediction algorithm in XReg useful for DSA and other recommendation tasks. Experiments on benchmark datasets demonstrated that XReg can outperform the state-of-the-art extreme classifiers as well as large-scale regressors and rankers by up to 50% reduction in the new XR error metric, and up to 2% and 2.4% improvements in terms of the propensity-scored precision metric used in extreme classification and the click-through rate metric used in DSA respectively. Deployment of XReg on DSA in Bing resulted in a relative gain of 27% in query coverage. XReg's source code can be downloaded from http://manikvarma.org/code/XReg/download.html.