Over the previous couple of years, autoregressive Transformers have introduced a gradual stream of breakthroughs in generative modeling. These fashions generate every ingredient of a pattern – the pixels of a picture, the characters of a textual content (usually in “token” chunks), the samples of an audio waveform, and so forth – by predicting one ingredient after the opposite. When predicting the subsequent ingredient, the mannequin can look again at people who had been created earlier.
Nevertheless, every of a Transformer’s layers grows costlier as extra parts are used as enter, and practitioners can solely afford to coach deep Transformers on sequences not more than about 2,048 parts in size. And so, most Transformer-based fashions ignore all parts past the newest previous (round 1,500 phrases or 1/6 of a small picture) when making a prediction.
In distinction, our lately developed Perceiver models give wonderful outcomes on quite a lot of real-world duties with as much as round 100,000 parts. Perceivers use cross-attention to encode inputs right into a latent area, decoupling the enter’s compute necessities from mannequin depth. Perceivers additionally spend a set value, no matter enter dimension, at almost each layer.
Whereas latent-space encoding handles all parts in a single cross, autoregressive era assumes processing occurs one ingredient at a time. To deal with this downside, Perceiver AR proposes a easy answer: align the latents one after the other with the ultimate parts of the enter, and thoroughly masks the enter so latents see solely earlier parts.
The result’s an structure (proven above) that attends to as a lot as 50x longer inputs as commonplace Transformers, whereas deploying as broadly (and primarily as simply) as commonplace decoder-only Transformers.
Perceiver AR scales significantly higher with dimension than each commonplace Transformers and Transformer-XL fashions at a variety of sequence lengths in actual phrases. This property permits us to construct very efficient long-context fashions. For instance, we discover {that a} 60-layer Perceiver AR with context size 8192 outperforms a 42-layer Transformer-XL on a book-length era activity, whereas working sooner in actual wall-clock phrases.
On commonplace, long-context picture (ImageNet 64×64), language (PG-19), and music (MAESTRO) era benchmarks, Perceiver AR produces state-of-the-art outcomes. Rising enter context by decoupling enter dimension from compute funds results in a number of intriguing outcomes:
- Compute funds might be tailored at eval time, permitting us to spend much less and easily degrade high quality or to spend extra for improved era.
- A bigger context permits Perceiver AR to outperform Transformer-XL, even when spending the identical on compute. We discover that better context results in improved mannequin efficiency even at inexpensive scale (~1B parameters).
- Perceiver AR’s pattern high quality reveals a lot much less sensitivity to the order by which it generates parts. This makes Perceiver AR straightforward to use to settings that don’t have a pure left-to-right ordering, akin to knowledge like photos, with construction that spans multiple dimension.
Utilizing a dataset of piano music, we educated Perceiver AR to generate new items of music from scratch. As a result of every new be aware is predicted based mostly on the complete sequence of notes that got here earlier than, Perceiver AR is ready to produce items with a excessive degree of melodic, harmonic, and rhythmic coherence:
