SqueezeBERT: What can computer vision teach NLP about efficient neural networks?

Abstract

Humans read and write hundreds of billions of messages every day. Further, due to the availability of large datasets, large computing systems, and better neural network models, natural language processing (NLP) technology has made significant strides in understanding, proofreading, and organizing these messages. Thus, there is a significant opportunity to deploy NLP in myriad applications to help web users, social networks, and businesses. In particular, we consider smartphones and other mobile devices as crucial platforms for deploying NLP models at scale. However, today's highly-accurate NLP neural network models such as BERT and RoBERTa are extremely computationally expensive, with BERT-base taking 1.7 seconds to classify a text snippet on a Pixel 3 smartphone. In this work, we observe that methods such as grouped convolutions have yielded significant speedups for computer vision networks, but many of these techniques have not been adopted by NLP neural network designers. We demonstrate how to replace several operations in self-attention layers with grouped convolutions, and we use this technique in a novel network architecture called SqueezeBERT, which runs 4.3x faster than BERT-base on the Pixel 3 while achieving competitive accuracy on the GLUE test set. The SqueezeBERT code will be released.

ADAHESSIAN: An Adaptive Second Order Optimizer for Machine Learning

“…Note that similar to NMT and LM, except learning rate and block size, ADAHESSIAN directly uses the same hyperparameters as AdamW. Interestingly note that these results are better than those reported in SqueezeBERT (Iandola et al 2020), even though we only change the optimizer to ADAHESSIAN instead of AdamW.…”

“…Comparison of AdamW and ADAHESSIAN for SqueezeBERT on the development set of the GLUE benchmark. The result of AdamW + is directly from(Iandola et al 2020) and the result of AdamW * is reproduced by us. AdamW 35.42 ± .09 35.66 ± .11 35.37 ± .07 35.18 ± .07 34.79 ± .15 14.41 ± 13.25 0.41 ± .32 Diverge ADAHESSIAN 35.33 ± .10 35.79 ± .06 35.21 ± .14 34.74 ± .10 34.19 ± .06 33.78 ± .14 32.70 ± .10 32.48 ± .83…”

Automatic Construction of Educational Knowledge Graphs: A Word Embedding-Based Approach

“…Lastly, we generated embeddings of these POS-tagged tokens using the SIFRank and SIFRankplus embedding techniques for short and long documents, respectively. Recognizing the performance of state-of-the-art transformer-based pre-trained language models, we replaced ELMo with the pre-trained SqueezeBERT [14] as a word embedding method in SIFRank and SIFRankplus. The decision to use SqueezeBERT is motivated by its lightweight transformer architecture with higher information flow between the layers; moreover, it is faster than the BERT model [14].…”

“…Through this research, we make the following contributions: (1) we adopt and adapt state-of-the-art word/sentence embedding techniques to automatically construct an EduKG, (2) we enhance the SIFRank keyphrase extraction method proposed in [13] by adopting SqueezeBERT [14], a transformer model for word embedding, (3) we propose an embeddingbased concept-weighting strategy using the sentence embedding technique SBERT [15], and (4) we conduct empirical studies on different datasets, demonstrating the effectiveness of the SqueezeBERT-enhanced SIFRank keyphrase extraction method as well as the efficiency of SBERT-based concept-weighting strategy against several baselines.…”

DCT-Former: Efficient Self-Attention with Discrete Cosine Transform

“…Pre-trained transformers can be then effectively fine-tuned for downstream supervised tasks, usually with little to none architectural changes. Among the considerable variety of pre-trained transformer models, BERT [2] and its derivatives [17][18][19][20][21] have become the de facto standard for deep language modeling. The strength of this model comes from the bidirectional pretraining strategy, which leverages a huge amount of unlabeled text in an unsupervised fashion.…”