High-Accuracy Low-Precision Training

Abstract

Low-precision computation is often used to lower the time and energy cost of machine learning, and recently hardware accelerators have been developed to support it. Still, it has been used primarily for inference - not training. Previous low-precision training algorithms suffered from a fundamental tradeoff: as the number of bits of precision is lowered, quantization noise is added to the model, which limits statistical accuracy. To address this issue, we describe a simple low-precision stochastic gradient descent variant called HALP. HALP converges at the same theoretical rate as full-precision algorithms despite the noise introduced by using low precision throughout execution. The key idea is to use SVRG to reduce gradient variance, and to combine this with a novel technique called bit centering to reduce quantization error. We show that on the CPU, HALP can run up to $4\times$ faster than full-precision SVRG and can match its convergence trajectory. We implemented HALP in TensorQuant, and show that it exceeds the validation performance of plain low-precision SGD on two deep learning tasks.

References in corpus (6)

• Scikit-learn: Machine Learning in Python
• Mixed Precision Training
• Visualizing and Understanding Recurrent Networks
• A Stochastic Gradient Method with an Exponential Convergence Rate for Finite Training Sets
• Stochastic Variance Reduction for Nonconvex Optimization
• The ZipML Framework for Training Models with End-to-End Low Precision: The Cans, the Cannots, and a Little Bit of Deep Learning

Cited by in corpus (34)

• Highly Scalable Deep Learning Training System with Mixed-Precision: Training ImageNet in Four Minutes
• ATOMO: Communication-efficient Learning via Atomic Sparsification
• Expanding the Reach of Federated Learning by Reducing Client Resource Requirements
• Taurus: A Data Plane Architecture for Per-Packet ML
• TensorDash: Exploiting Sparsity to Accelerate Deep Neural Network Training and Inference
• On Maintaining Linear Convergence of Distributed Learning and Optimization under Limited Communication
• Federated Learning for Healthcare Informatics
• On-Device Machine Learning: An Algorithms and Learning Theory Perspective
• Scalable Deep Learning on Distributed Infrastructures: Challenges, Techniques and Tools
• Discovering Low-Precision Networks Close to Full-Precision Networks for Efficient Embedded Inference
• Mixed Precision Training With 8-bit Floating Point
• Moniqua: Modulo Quantized Communication in Decentralized SGD
• QUOTIENT: Two-Party Secure Neural Network Training and Prediction
• Word2Bits - Quantized Word Vectors
• Accelerating Generalized Linear Models with MLWeaving: A One-Size-Fits-All System for Any-precision Learning (Technical Report)
• TF-Replicator: Distributed Machine Learning for Researchers
• Accuracy-Efficiency Trade-Offs and Accountability in Distributed ML Systems
• Post-Training 4-bit Quantization on Embedding Tables
• Mixed-Precision Embedding Using a Cache
• Sparse Weight Activation Training
• High-Performance Deep Learning via a Single Building Block
• Rethinking Floating Point Overheads for Mixed Precision DNN Accelerators
• word2ket: Space-efficient Word Embeddings inspired by Quantum Entanglement
• Pufferfish: Communication-efficient Models At No Extra Cost
• Distributed Low Precision Training Without Mixed Precision
• An FPGA Accelerated Method for Training Feed-forward Neural Networks Using Alternating Direction Method of Multipliers and LSMR
• How Low Can We Go: Trading Memory for Error in Low-Precision Training
• Quantized Epoch-SGD for Communication-Efficient Distributed Learning
• EDEN: Enabling Energy-Efficient, High-Performance Deep Neural Network Inference Using Approximate DRAM
• OD-SGD: One-step Delay Stochastic Gradient Descent for Distributed Training
• ResIST: Layer-Wise Decomposition of ResNets for Distributed Training
• On the Downstream Performance of Compressed Word Embeddings
• Estimating Principal Components under Adversarial Perturbations
• Low-Precision Hardware Architectures Meet Recommendation Model Inference at Scale