Recent works have demonstrated great success in pre-training large-scale autoregressive language models (e.g., GPT-3) on massive GPUs. To reduce the wall-clock training time, a common practice is to increase the batch size and learning rate. However, such practice is often brittle and leads to a so-called stability-efficiency dilemma: increasing the batch sizes and learning rates leads to better training efficiency but can also result in training instability, leading to poor generalization accuracy or failed runs. To better understand this phenomenon, we conduct an in-depth analysis on large-scale pre-training experiments replicating the GPT-2 model with public dataset. We find that there is a strong correlation between training instability and extreme values of gradient variance.

Recent works have demonstrated great success in pre-training large-scale autoregressive language models (e.g., GPT-3) on massive GPUs. To reduce the wall-clock training time, a common practice is to increase the batch size and learning rate. However, such practice is often brittle and leads to a so-called stability-efficiency dilemma: increasing the batch sizes and learning rates leads to better training efficiency but can also result in training instability, leading to poor generalization accuracy or failed runs. To better understand this phenomenon, we conduct an in-depth analysis on large-scale pre-training experiments replicating the GPT-2 model with public dataset. We find that there is a strong correlation between training instability and extreme values of gradient variance.

For the task of speech recognition, the use of more than 30 seconds of acoustic context during training is uncommon, and under-investigated in literature. In this work, we examine the effect of scaling the sequence length used to train/evaluate (dense-attention based) acoustic and language models on speech recognition performance. For these experiments a dataset of roughly 100,000 pseudo-labelled Spotify podcasts is used, with context lengths of 5 seconds to 1 hour being explored. Zero-shot evaluations on long-format datasets Earnings-22 and Tedlium demonstrate a benefit from training with around 80 seconds of acoustic context, showing up to a 14.9% relative improvement from a limited context baseline. Furthermore, we perform a system combination with long-context transformer language models via beam search for a fully long-context ASR system, with results that are competitive with the current state-of-the-art.

In the past decade, we have seen Deep learning (DL) science adopted at a tremendous pace by AWS customers. The plentiful and jointly trained parameters of DL models have a large representational capacity that brought improvements in numerous customer use cases, including image and speech analysis, natural language processing (NLP), time series processing, and more. In this post, we highlight challenges commonly reported specifically in DL training, and how the open-source library MosaicML Composer helps solve them. DL models are trained iteratively, in a nested for loop. A loop iterates through the training dataset chunk by chunk and, if necessary, this loop is repeated several times over the whole dataset.