We present the Compressive Transformer, an attentive sequence model which compresses past memories for long-range sequence learning. We find the Compressive Transformer obtains state-of-the-art language modelling results in the WikiText-103 and Enwik8 benchmarks, achieving 17.1 ppl and 0.97 bpc respectively. We also find it can model high-frequency speech effectively and can be used as a memory mechanism for RL, demonstrated on an object matching task. To promote the domain of long-range sequence learning, we propose a new open-vocabulary language modelling benchmark derived from books, PG-19.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Language Modelling | WikiText-103 | Test perplexity | 17.1 | Compressive Transformer (18L, M=1024) |
| Language Modelling | WikiText-103 | Validation perplexity | 16 | Compressive Transformer (18L, M=1024) |
| Language Modelling | Hutter Prize | Bit per Character (BPC) | 0.97 | Compressive Transformer |
| Language Modelling | enwik8 | Bit per Character (BPC) | 0.97 | Compressive Transformer (24 layers) |