Not all layers are equally as important: Every Layer Counts BERT
This paper introduces a novel modification of the transformer architecture, tailored for the data-efficient pretraining of language models. This aspect is evaluated by participating in the BabyLM challenge, where our solution won both the strict and strict-small tracks. Our approach allows each transformer layer to select which outputs of previous layers to process. The empirical results verify the potential of this simple modification and show that not all layers are equally as important.
PDF AbstractTask | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
---|---|---|---|---|---|---|
Linguistic Acceptability | CoLA | ELC-BERT-base 98M | Accuracy | 82.6 | # 7 | |
Linguistic Acceptability | CoLA | ELC-BERT-small 24M | Accuracy | 76.1 | # 11 | |
Linguistic Acceptability | CoLA | LTG-BERT-small 24M | Accuracy | 77.6 | # 10 | |
Linguistic Acceptability | CoLA | LTG-BERT-base 98M | Accuracy | 82.7 | # 6 | |
Natural Language Inference | MultiNLI | ELC-BERT-small 24M | Matched | 79.2 | # 41 | |
Mismatched | 79.9 | # 31 | ||||
Natural Language Inference | MultiNLI | LTG-BERT-small 24M | Matched | 78 | # 42 | |
Mismatched | 78.8 | # 32 | ||||
Natural Language Inference | MultiNLI | LTG-BERT-base 98M | Matched | 83 | # 34 | |
Mismatched | 83.4 | # 22 | ||||
Natural Language Inference | MultiNLI | ELC-BERT-base 98M (zero init) | Matched | 84.4 | # 30 | |
Mismatched | 84.5 | # 19 | ||||
Natural Language Inference | RTE | ELC-BERT-small 24M | Accuracy | 55.4 | # 82 | |
Natural Language Inference | RTE | ELC-BERT-base 98M (zero init) | Accuracy | 63 | # 67 | |
Natural Language Inference | RTE | LTG-BERT-base 98M | Accuracy | 54.7 | # 84 | |
Natural Language Inference | RTE | LTG-BERT-small 24M | Accuracy | 53.7 | # 87 |