Skip to content

Paper 019 / DeepMind

Perceiver: General Perception with Iterative Attention

Perceiver uses a small latent array to repeatedly attend to very large inputs, decoupling computational cost from raw input size.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / multimodal_models

Perceiver uses a small latent array to repeatedly attend to very large inputs, decoupling computational cost from raw input size.

CONTRIBUTION / 01

Contribution 1

Introduced latent bottleneck cross-attention for high-dimensional inputs.

CONTRIBUTION / 02

Contribution 2

Applied one architecture to images, audio, video, and point clouds.

CONTRIBUTION / 03

Contribution 3

Reduced the quadratic burden of direct input self-attention.

02

Research context

multimodal_models / 2021

Perceiver: General Perception with Iterative Attention places perceiver inside the broader multimodal models discussion at DeepMind, with multimodal supplying a second analytical lens. The editorial sequence connects three claims: Introduced latent bottleneck cross-attention for high-dimensional inputs; Applied one architecture to images, audio, video, and point clouds; and Reduced the quadratic burden of direct input self-attention. The combination matters because latent bottleneck only has meaning under the paper's stated setup. Operationally, the record points to one consequence: latent bottlenecks offer a path to handling large multimodal inputs while keeping the core network size controlled.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in Perceiver: General Perception with Iterative Attention tests introduced latent bottleneck cross-attention for high-dimensional inputs and applied one architecture to images, audio, video, and point clouds against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

For Perceiver: General Perception with Iterative Attention, the useful question is what the documented setup supports and where transfer remains untested. Its evidence chain begins with two documented moves: Introduced latent bottleneck cross-attention for high-dimensional inputs; and Applied one architecture to images, audio, video, and point clouds. Reported evidence then addresses: Reduced the quadratic burden of direct input self-attention. The resulting interpretation is practical but conditional: latent bottlenecks offer a path to handling large multimodal inputs while keeping the core network size controlled. Its boundary is that the perceiver comparison in Perceiver: General Perception with Iterative Attention is interpretable only alongside domain coverage, input modalities, prompt protocol, selected metrics, reported datasets, and image resolution, which limits claims about unseen deployments. Any extension should report how altered multimodal conditions affect the original perceiver result. For a follow-on study of Perceiver: General Perception with Iterative Attention, pair perceiver measurements with multimodal slices and preserve negative examples around Applied one architecture to images, audio, video, and point clouds as first-class evidence.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in Perceiver: General Perception with Iterative Attention supports reduced the quadratic burden of direct input self-attention.
05

Practical implication for AI builders

DeepMind / 2021

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / perceiver

For a proposed BrokenGPT experiment based on Perceiver: General Perception with Iterative Attention, document how multimodal endpoints compress long inputs and test whether small but important details survive preprocessing before routing user files. Keep the perceiver path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / multimodal

Validate the proposed perceiver route against the paper's reported outcome: Reduced the quadratic burden of direct input self-attention. A proposed Perceiver: General Perception with Iterative Attention gate needs grounding accuracy, modality-specific error, and cross-modal consistency; its multimodal cases should remain disaggregated from the overall perceiver score.

TRADEOFF / 03

Proposed decision boundary / latent bottleneck

Balance coverage, compute, and provenance before promoting the proposed latent bottleneck design. Because A production test of perceiver must also examine accessibility, memorization, misuse, representation bias, source provenance, and media rights, none of which the reported multimodal result resolves automatically, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • The perceiver comparison in Perceiver: General Perception with Iterative Attention is interpretable only alongside domain coverage, input modalities, prompt protocol, selected metrics, reported datasets, and image resolution, which limits claims about unseen deployments.
  • A production test of perceiver must also examine accessibility, memorization, misuse, representation bias, source provenance, and media rights, none of which the reported multimodal result resolves automatically.

PRIMARY SOURCES

  1. 01
    Perceiver: General Perception with Iterative Attention

    DeepMind — Primary primary arXiv paper / 4 March 2021 / Andrew Jaegle, Felix Gimeno, Andrew Brock, and 3 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does Perceiver: General Perception with Iterative Attention study?

Perceiver uses a small latent array to repeatedly attend to very large inputs, decoupling computational cost from raw input size.

02Which methods does Perceiver: General Perception with Iterative Attention use?

The experimental design in Perceiver: General Perception with Iterative Attention tests introduced latent bottleneck cross-attention for high-dimensional inputs and applied one architecture to images, audio, video, and point clouds against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does Perceiver: General Perception with Iterative Attention report?

The reported evidence in Perceiver: General Perception with Iterative Attention supports reduced the quadratic burden of direct input self-attention.

04What is the proposed BrokenGPT application for Perceiver: General Perception with Iterative Attention?

Proposed: document how multimodal endpoints compress long inputs and test whether small but important details survive preprocessing before routing user files.

MAJOR LAB RESEARCH / PAPER 019

Continue after Perceiver: General Perception with Iterative Attention

After Perceiver: General Perception with Iterative Attention, browse the full index for adjacent multimodal_models research and work from DeepMind.

Open the paper index