arXiv subjects

Connor Holmes

Publications and source records attributed to Connor Holmes.

23 records · Page 2Linked to original sources

arXiv2022

Compressing Pre-trained Transformers via Low-Bit NxM Sparsity for Natural Language Understanding

In recent years, large pre-trained Transformer networks have demonstrated dramatic improvements in many natural language understanding tasks. However, the huge size of these models brings significant challenges to their fine-tuning and online deployment due to latency and cost constraints. New hardware supporting both N:M semi-structured sparsity and low-precision integer computation is a promising solution to boost DNN model serving efficiency. However, there have been very few studies that systematically investigate to what extent pre-trained Transformer networks benefit from the combination of these techniques, as well as how to best compress each component of the Transformer. We propose a flexible compression framework NxMiFormer that performs simultaneous sparsification and quantization using ADMM and STE-based QAT. Furthermore, we present and inexpensive, heuristic-driven search algorithm that identifies promising heterogeneous compression configurations that meet a compression ratio constraint. When evaluated across the GLUE suite of NLU benchmarks, our approach can achieve up to 93% compression of the encoders of a BERT model while retaining 98.2% of the original model accuracy and taking full advantage of the hardware's capabilities. Heterogeneous configurations found the by the search heuristic maintain 99.5% of the baseline accuracy while still compressing the model by 87.5%.

cs.CL↗

arXiv2022

An Efficient Global Optimality Certificate for Landmark-Based SLAM

Modern state estimation is often formulated as an optimization problem and solved using efficient local search methods. These methods at best guarantee convergence to local minima, but, in some cases, global optimality can also be certified. Although such global optimality certificates have been well established for 3D \textit{pose-graph optimization}, the details have yet to be worked out for the 3D landmark-based SLAM problem, in which estimated states include both robot poses and map landmarks. In this paper, we address this gap by using a graph-theoretic approach to cast the subproblems of landmark-based SLAM into a form that yields a sufficient condition for global optimality. Efficient methods of computing the optimality certificates for these subproblems exist, but first require the construction of a large data matrix. We show that this matrix can be constructed with complexity that remains linear in the number of landmarks and does not exceed the state-of-the-art computational complexity of one local solver iteration. We demonstrate the efficacy of the certificate on simulated and real-world landmark-based SLAM problems. We also integrate our method into the state-of-the-art SE-Sync pipeline to efficiently solve landmark-based SLAM problems to global optimality. Finally, we study the robustness of the global optimality certificate to measurement noise, taking into consideration the effect of the underlying measurement graph.

cs.RO↗

arXiv2022

A Fine Line: Total Least-Squares Line Fitting as QCQP Optimization

This note uses the Total Least-Squares (TLS) line-fitting problem as a canvas to explore some modern optimization tools. The contribution is meant to be tutorial in nature. The TLS problem has a lot of mathematical similarities to important problems in robotics and computer vision but is easier to visualize and understand. We demonstrate how to turn this problem into a Quadratically Constrained Quadratic Program (QCQP) so that it can be cast either as an eigenproblem or a Semi-Definite Program (SDP). We then turn to the more challenging situation where a Geman-McClure cost function and M-estimation are used to reject outlier datapoints. Using Black-Rangarajan duality, we show this can also be cast as a QCQP and solved as an SDP; however, with a lot of data the SDP can be slow and as such we show how we can construct a certificate of optimality for a faster method such as Iteratively Reweighted Least-Squares (IRLS).

cs.RO↗

arXiv2021

NxMTransformer: Semi-Structured Sparsification for Natural Language Understanding via ADMM

Natural Language Processing (NLP) has recently achieved success by using huge pre-trained Transformer networks. However, these models often contain hundreds of millions or even billions of parameters, bringing challenges to online deployment due to latency constraints. Recently, hardware manufacturers have introduced dedicated hardware for NxM sparsity to provide the flexibility of unstructured pruning with the runtime efficiency of structured approaches. NxM sparsity permits arbitrarily selecting M parameters to retain from a contiguous group of N in the dense representation. However, due to the extremely high complexity of pre-trained models, the standard sparse fine-tuning techniques often fail to generalize well on downstream tasks, which have limited data resources. To address such an issue in a principled manner, we introduce a new learning framework, called NxMTransformer, to induce NxM semi-structured sparsity on pretrained language models for natural language understanding to obtain better performance. In particular, we propose to formulate the NxM sparsity as a constrained optimization problem and use Alternating Direction Method of Multipliers (ADMM) to optimize the downstream tasks while taking the underlying hardware constraints into consideration. ADMM decomposes the NxM sparsification problem into two sub-problems that can be solved sequentially, generating sparsified Transformer networks that achieve high accuracy while being able to effectively execute on newly released hardware. We apply our approach to a wide range of NLP tasks, and our proposed method is able to achieve 1.7 points higher accuracy in GLUE score than current practices. Moreover, we perform detailed analysis on our approach and shed light on how ADMM affects fine-tuning accuracy for downstream tasks. Finally, we illustrate how NxMTransformer achieves performance improvement with knowledge distillation.

cs.CL↗

arXiv2019

GraphZero: Breaking Symmetry for Efficient Graph Mining

Graph mining for structural patterns is a fundamental task in many applications. Compilation-based graph mining systems, represented by AutoMine, generate specialized algorithms for the provided patterns and substantially outperform other systems. However, the generated code causes substantial computation redundancy and the compilation process incurs too much overhead to be used online, both due to the inherent symmetry in the structural patterns. In this paper, we propose an optimizing compiler, GraphZero, to completely address these limitations through symmetry breaking based on group theory. GraphZero implements three novel techniques. First, its schedule explorer efficiently prunes the schedule space without missing any high-performance schedule. Second, it automatically generates and enforces a set of restrictions to eliminate computation redundancy. Third, it generalizes orientation, a surprisingly effective optimization that was mainly used for clique patterns, to apply to arbitrary patterns. Evaluated on multiple graph mining applications and complex patterns with 7 real-world graph datasets, GraphZero demonstrates up to 40X performance improvement and up to 197X reduction on schedule generation overhead over AutoMine.

cs.PF↗