| Challenge: | Asynchronous stochastic gradient descent (SGD) converges poorly for Transformer models . synchronous SGD is faster at raw training speed since it avoids waiting for synchronization . |
| Approach: | They propose a method to restore convergence by summing several asynchronous updates instead of applying them immediately. |
| Outcome: | The proposed method achieves the same BLEU score 1.36 times faster than asynchronous SGD. |
Similar Papers
Accelerating Asynchronous Stochastic Gradient Descent for Neural Machine Translation (D18-1)
Copied to clipboard
| Challenge: | In order to achieve faster training we increase the mini-batch size and scale the learning rate accordingly. |
| Approach: | They propose a technique that delays gradient updates by increasing the mini-batch size to improve the model's convergence. |
| Outcome: | The proposed technique can train a shallow machine translation system 27% faster than an optimized baseline with negligible penalty in BLEU. |
Understanding the Difficulty of Training Transformers (2020.emnlp-main)
Copied to clipboard
| Challenge: | Admin (Adaptive model initialization) is more stable, converges faster, and leads to better performance. |
| Approach: | They propose a model initialization algorithm to stabilize early training and unleash its full potential in the late stage. |
| Outcome: | The proposed model initialization method stabilizes early training and unleashes full potential in late stage. |
Combining Global Sparse Gradients with Local Gradients in Distributed Neural Network Training (D19-1)
Copied to clipboard
| Challenge: | In recent years, neural network models have grown dramatically in terms of number of parameters, so exchanging gradients during data-parallel training is costly in terms both of bandwidth and time. |
| Approach: | They propose to combine the compressed global gradient with the local gradient to restore Transformer convergence while RNNs converge faster. |
| Outcome: | The proposed method restores transformer convergence while RNNs converge faster. |
Should I try multiple optimizers when fine-tuning a pre-trained Transformer for NLP tasks? Should I tune their hyperparameters? (2024.eacl-long)
Copied to clipboard
| Challenge: | NLP research has explored different neural model architectures and sizes, datasets, training objectives, and transfer-learning techniques. |
| Approach: | They propose to use a variant of Stochastic Gradient Descent (SGD) to select among numerous variants, often with minimal or no tuning of the optimizer’s hyperparameters. |
| Outcome: | Experiments with five GLUE datasets, two models and seven popular optimizers show that tuning just the learning rate is as good as tuning all the hyperparameters. |
Towards Incremental Transformers: An Empirical Analysis of Transformer Models for Incremental NLU (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent work attempts to apply incremental processing to NLUs but this is computationally expensive and does not scale efficiently for long sequences. |
| Approach: | They propose to apply Transformers incrementally via restart-incrementality by repeatedly feeding, to an unchanged model, increasingly longer input prefixes to produce partial outputs. |
| Outcome: | The proposed model has better incremental performance and faster inference speed compared to the standard Transformer and LT with restart-incrementality, at the cost of part of the non-incremental quality. |
Effects of Parameter Norm Growth During Transformer Training: Inductive Bias from Gradient Descent (2021.emnlp-main)
Copied to clipboard
| Challenge: | Evidence is emerging that neural networks learn due to inductive bias in the training routine, typically a variant of gradient descent (GD). |
| Approach: | They propose to characterize GD as an inductive bias in transformer training . they document norm growth in transformer language models and show they are saturated . |
| Outcome: | Empirically, we document norm growth in the training of transformer language models . the results suggest saturation is a new characterization of an inductive bias implicit in GD . |
Reservoir Transformers (2021.acl-long)
Copied to clipboard
| Challenge: | Using random initialization, we show that some transformers obtain impressive performance even when some of the layers are frozen. |
| Approach: | They propose to freeze transformer layers and use them to improve performance . they find that the transformers obtain impressive performance even when some of the layers are randomly initialized and never updated. |
| Outcome: | The proposed model improves on translation and language modelling tasks even when some layers are frozen. |
Revisiting the Weaknesses of Reinforcement Learning for Neural Machine Translation (2021.naacl-main)
Copied to clipboard
| Challenge: | In neural sequence-to-sequence learning, Reinforcement Learning (RL) has gained popularity due to the suitability of Policy Gradient (PG) methods for the end-to end training paradigm. |
| Approach: | They propose to let the model explore the output space beyond the reference output that is used for standard cross-entropy minimization by reinforcing model outputs according to their quality, effectively increasing the likelihood of higher-quality samples. |
| Outcome: | The proposed model explores the output space beyond the reference output that is used for cross-entropy minimization, increasing the likelihood of higher-quality samples. |
How Effective is Task-Agnostic Data Augmentation for Pretrained Transformers? (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Task-agnostic data augmentations have proven widely effective in computer vision, even on pretrained models. |
| Approach: | They examine the effects of two types of task-agnostic data augmentation on pretrained transformers using 5 classification tasks and 6 datasets. |
| Outcome: | The proposed techniques improve performance on 5 classification tasks, 6 datasets, and 3 variants of modern pretrained transformers. |
Fully Non-autoregressive Neural Machine Translation: Tricks of the Trade (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing non-autoregressive neural machine translation models are slow to learn the dependency between output tokens. |
| Approach: | They propose to use fully non-autoregressive neural machine translation (NAT) to predict tokens with single forward of neural networks. |
| Outcome: | The proposed model achieves state-of-the-art results on three translation benchmarks with comparable performance to autoregressive and iterative NAT systems. |