Transformers-based models, such as BERT, have been one of the most successful deep learning models for NLP. Unfortunately, one of their core limitations is the quadratic dependency (mainly in terms of memory) on the sequence length due to their full attention mechanism.To remedy this, we propose, BigBird, a sparse attention mechanism that reduces this quadratic dependency to linear. We show that BigBird is a universal approximator of sequence functions and is Turing complete, thereby preserving these properties of the quadratic, full attention model. Along the way, our theoretical analysis reveals some of the benefits of having O(1) global tokens (such as CLS), that attend to the entire sequence as part of the sparse attention mechanism. The proposed sparse attention can handle sequences of length up to 8x of what was previously possible using similar hardware. As a consequence of the capability to handle longer context, BigBird drastically improves performance on various NLP tasks such as question answering and summarization. We also propose novel applications to genomics data.

Transformer models have advanced the state of the art in many Natural Language Processing (NLP) tasks.In this paper, we present a new Transformer architecture, Extended Transformer Construction (ETC), that addresses two key challenges of standard Transformer architectures, namely scaling input length and encoding structured inputs. To scale attention to longer inputs, we introduce a novel global-local attention mechanism between global tokens and regular input tokens. We also show that combining global-local attention with relative position encodings and a Contrastive Predictive Coding (CPC) pretraining objective allows ETC to encode structured inputs. We achieve state-of-the-art results on four natural language datasets requiring long and/or structured inputs.

The Muller F element (4.2 Mb, ~80 protein-coding genes) is an unusual autosome of Drosophila melanogaster; it is mostly heterochromatic with a low recombination rate. To investigate how these properties impact the evolution of repeats and genes, we manually improved the sequence and annotated the genes on the D. erecta, D. mojavensis, and D. grimshawi F elements and euchromatic domains from the Muller D element. We find that F elements have greater transposon density (25–50%) than euchromatic reference regions (3–11%). Among the F elements, D. grimshawi has the lowest transposon density (particularly DINE-1: 2% vs. 11–27%). F element genes have larger coding spans, more coding exons, larger introns, and lower codon bias. Comparison of the Effective Number of Codons with the Codon Adaptation Index shows that, in contrast to the other species, codon bias in D. grimshawi F element genes can be attributed primarily to selection instead of mutational biases, suggesting that density and types of transposons affect the degree of local heterochromatin formation. F element genes have lower estimated DNA melting temperatures than D element genes, potentially facilitating transcription through heterochromatin. Most F element genes (~90%) have remained on that element, but the F element has smaller syntenic blocks than genome averages (3.4–3.6 vs. 8.4–8.8 genes per block), indicating greater rates of inversion despite lower rates of recombination. Overall, the F element has maintained characteristics that are distinct from other autosomes in the Drosophila lineage, illuminating the constraints imposed by a heterochromatic milieu.

No abstract

No abstract