[Paper Review] Perceiver: General Perception with Iterative Attention

Andrew Jaegle, Felix Gimeno+4 more|arXiv (Cornell University)|Mar 4, 2021

Neural dynamics and brain function91 references128 citations

TL;DR

Perceiver presents a Transformer-based architecture that scales to very large, multimodal inputs by using a cross-attention bottleneck to a small latent array and iterative latent self-attention, achieving competitive results across images, audio, video, and point clouds without modality-specific priors.

ABSTRACT

Biological systems perceive the world by simultaneously processing high-dimensional inputs from modalities as diverse as vision, audition, touch, proprioception, etc. The perception models used in deep learning on the other hand are designed for individual modalities, often relying on domain-specific assumptions such as the local grid structures exploited by virtually all existing vision models. These priors introduce helpful inductive biases, but also lock models to individual modalities. In this paper we introduce the Perceiver - a model that builds upon Transformers and hence makes few architectural assumptions about the relationship between its inputs, but that also scales to hundreds of thousands of inputs, like ConvNets. The model leverages an asymmetric attention mechanism to iteratively distill inputs into a tight latent bottleneck, allowing it to scale to handle very large inputs. We show that this architecture is competitive with or outperforms strong, specialized models on classification tasks across various modalities: images, point clouds, audio, video, and video+audio. The Perceiver obtains performance comparable to ResNet-50 and ViT on ImageNet without 2D convolutions by directly attending to 50,000 pixels. It is also competitive in all modalities in AudioSet.

Motivation & Objective

• Motivate a general perceptual architecture that minimizes modality-specific priors.
• Introduce the Perceiver, which scales attention by projecting high-dimensional inputs to a small latent bottleneck.
• Demonstrate competitive performance across diverse modalities without 2D convolutions or domain-specific priors.
• Show how iterative cross-attention and latent self-attention enable deep representations from large inputs.

Proposed method

• Use a cross-attention module to map a high-dimensional input byte array to a fixed-size latent array (N << M).
• Process the latent array with a deep Transformer in the latent space (complexity ~ O(N^2)).
• Iteratively alternate cross-attention and latent self-attention blocks to refine input representations.
• Share weights across cross-attention modules and latent Transformer blocks to improve efficiency and enable deep architectures.
• Attach position/modality information to each input element via scalable Fourier features or learned encodings to retain spatial/temporal structure.
• Optionally use multiple cross-attention layers to improve information extraction from inputs.

Experimental results

Research questions

• RQ1 Can a largely modality-agnostic Transformer-based architecture achieve competitive perceptual performance across vision, audio, video, and point clouds?
• RQ2 Does the asymmetric cross-attention bottleneck enable scaling to tens of thousands of inputs while maintaining accuracy?
• RQ3 How do Fourier-based position encodings influence performance and permutation robustness across modalities?
• RQ4 What are the trade-offs between cross-attention depth and latent Transformer depth, and how does weight sharing affect efficiency and accuracy?
• RQ5 How does Perceiver perform relative to specialized architectures (e.g., ResNet-50, ViT) on ImageNet, AudioSet, and ModelNet40?

Key findings

• Achieves competitive ImageNet Top-1 accuracy without 2D convolutions, at 50,176 input pixels.
• Attains strong performance on AudioSet with raw audio, video, or both.
• Demonstrates competitive results on ModelNet-40 point cloud classification.
• Allows very deep models by decoupling input size from depth via the latent bottleneck, with overall complexity O(MN + LN^2).
• Weight sharing across cross-attention and Transformer blocks reduces parameters by ~10x and improves generalization.
• Fourier feature position encodings enable the model to retain spatial/temporal structure without hard architectural priors.