| Property | Value |
|---|---|
| Developer | DeepMind |
| Architecture | Perceiver IO |
| Training Data | English Wikipedia (30%) + C4 (70%) |
| Paper | Perceiver IO: A General Architecture for Structured Inputs & Outputs |
| GLUE Score | 81.8 |
What is language-perceiver?
Language Perceiver is an innovative transformer-based model that revolutionizes how we process language data. Unlike traditional transformers, it uses a fixed number of latent vectors to process input through cross-attention, making computational requirements independent of input size. The model works directly with raw UTF-8 bytes instead of tokenized text, eliminating the need for pre-trained tokenizers or fixed vocabularies.
Implementation Details
The model employs a unique architecture where self-attention is performed on a small set of latent vectors (256 or 512), with inputs only participating in cross-attention. This design allows for efficient processing of arbitrary-length inputs while maintaining consistent computational costs. The model uses decoder queries for flexible output generation, capable of producing predictions for masked language modeling tasks.
- Direct processing of UTF-8 bytes without tokenization
- Cross-attention mechanism with latent vectors
- Flexible decoder queries for output generation
- Maximum sequence length of 2048 bytes
Core Capabilities
- Masked Language Modeling (MLM)
- Feature extraction for downstream tasks
- Flexible input processing across modalities
- Efficient handling of long sequences
Frequently Asked Questions
Q: What makes this model unique?
The model's ability to process raw bytes instead of tokens, combined with its fixed computational complexity regardless of input size, makes it highly efficient and flexible. It can handle multiple modalities and doesn't require a pre-trained tokenizer.
Q: What are the recommended use cases?
While the base model excels at masked language modeling, it's primarily designed for fine-tuning on specific downstream tasks. It's particularly useful for applications requiring efficient processing of long sequences or working with multiple modalities.