BERT Compression-Data Quantization

A survey of BERT compressin using data quantization.


Since 2018, BERT and its variants (RoBERT and XLNet) show significant improvements in many Natural Language Processing tasks. However, all these models contain a large number of parameters, which results in prohibitive memory footprint and slow inference. In order to overcome such problems, many methods have been proposed to compress these large models. In this post, I will survey papers that compress these large models using Data Quantization.

Data quantization refers to representing each model weight using fewer bits, which reduces the memory footprint of the model and lowers the precision of its numerical calculations. There are a few characteristics of data quantization:

A straightforward idea is directly to truncate the weight bits to the target bandwidth. However, this operation usually leads to a large drop in accuracy Ganesh et al.. In order to overcome this problem, the Quantization Aware Training(QAT) approach is proposed. Unlike simply truncating the weights, QAT quantizes the weights and retrain the model to adjust the new weights.

Data Quantization Papers

There are not many papers that directly quantizes the models of the BERT model family.

Zafrir et al. quantizes the all general matrix multiply operations in BERT fully connected and embedding layers. All these weights are quantized to 8-bit integers. In the fine-tuning process, they manually introduce some quantization error to let the model learn the error gap. They use Straight-Through Estimator(STE) to estimate the gradient. According to their experiments, they claim they can compress the model by \(4\) times with minimal accuracy loss.

Similarly, Fan et al. also tries to let the model adjust to quantization during the training process. In this paper, the authors made a simple modification to the QAT approach. The idea is, in the forward pass, they randomly select a subset of weights instead of the full network as in QAT. Other parts of the weights remain the same. They call this approach Quant-Noise. I think Quant-Noise is very similar to the idea of dropout.

Instead of quantizing all the weights into the same bandwidth, Shen et al. tries to quantizes models dynamically. The motivation is different components (or layers) of a model that have different sensitivity. We should assign more bits to more sensitive components. In this paper, they proposed a new hessian matrix-based metric to evaluate the sensitivity of different layers. Interestingly, in their experiments, they find that the embedding layer is more sensitive than other weights for quantization, which means the embedding layer needs more bits to maintain accuracy.


The comparison between data quantization papers:

Paper Memory (BERT:100%) Inference Component Pretrain Performace
Zafrir et al. 25% N/A Float Bits False very close
Fan et al. 5% N/A Float Bits False worse
Shen et al. 7.8% N/A Float Bits False slightly worse