# Perceiver The Perceiver is a deep learning model architecture that processes arbitrary input modalities using a fixed-size latent array, enabling efficient scaling to large inputs via cross-attention and iterative processing. The Perceiver is a [neural network](https://www.wikiprompt.org/wiki/neural-network) architecture introduced by researchers at [Google DeepMind](https://www.wikiprompt.org/wiki/google-deepmind) in 2021. It is designed to handle a wide range of input types - including images, audio, video, and point clouds - without requiring domain-specific preprocessing. Unlike standard [transformers](https://www.wikiprompt.org/wiki/transformer) that scale quadratically with input length, the Perceiver uses a fixed-size latent array to achieve linear computational complexity, making it suitable for high-dimensional data such as raw pixels or audio waveforms. The architecture was first described in the paper "Perceiver: General Perception with Iterative Attention," published in March 2021 by Andrew Jaegle, Felix Gimeno, Andrew Brock, and colleagues. The model builds on the [transformer](https://www.wikiprompt.org/wiki/transformer) framework but replaces the self-attention over the full input with a two-stage process: cross-attention from the latent array to the input, followed by self-attention within the latent array. This design allows the model to process inputs of arbitrary length while maintaining a constant memory footprint. ## Architecture The Perceiver consists of three main components: an input encoder, a latent array, and a decoder. The input encoder transforms raw data into a sequence of feature vectors, often using a simple Fourier feature mapping for positional information. The latent array is a fixed set of learned vectors (typically 256 or 512) that serve as a compressed representation of the input. Cross-attention layers allow each latent vector to attend to all input positions, while subsequent self-attention layers enable interactions among the latent vectors. This iterative process is repeated several times, with the latent array refining its understanding of the input across multiple passes. A key innovation is the use of Fourier features for positional encoding, which allows the model to handle inputs of varying dimensions and resolutions without explicit positional embeddings. The Perceiver also incorporates a technique called "weight tying" across iterations, where the same cross-attention and self-attention weights are reused, reducing parameter count and improving generalization. ## Variants and Extensions The original Perceiver was followed by the Perceiver IO, introduced in June 2021, which extends the architecture to handle arbitrary output structures, such as classification labels, segmentation masks, or language tokens. Perceiver IO uses a query array to decode outputs of any shape, making it a general-purpose sequence-to-sequence model. Another variant, Perceiver AR, adapts the architecture for autoregressive generation, enabling it to produce sequential outputs like text or audio. These variants have been applied to tasks such as image classification, optical character recognition, and multimodal learning. The Perceiver's ability to process raw audio waveforms directly, without spectrogram preprocessing, has been demonstrated in audio classification benchmarks, achieving competitive results with specialized models. ## Applications Perceiver models have been used in several practical domains. In computer vision, they have been applied to image classification and object detection, often matching or exceeding the performance of convolutional networks on standard datasets like ImageNet. In audio processing, the Perceiver can handle long audio sequences, making it useful for speech recognition and music generation. The architecture has also been explored for reinforcement learning, where it processes high-dimensional sensory inputs such as camera feeds and lidar data. In the context of [artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) research, the Perceiver is notable for its generality, as it does not rely on domain-specific inductive biases like convolution or recurrence. This aligns with the broader trend toward unified models that can handle multiple modalities, a direction also pursued by [large language models](https://www.wikiprompt.org/wiki/large-language-model) and multimodal systems. ## Comparison with Other Architectures Compared to standard [transformers](https://www.wikiprompt.org/wiki/transformer), the Perceiver offers significant computational advantages for long inputs. A transformer with N input tokens requires O(N^2) attention operations, while the Perceiver reduces this to O(N * L), where L is the latent size (typically much smaller than N). This makes it feasible to process inputs with millions of tokens, such as high-resolution images or long audio clips, on a single GPU. However, the Perceiver's iterative attention process can be slower in practice for short inputs due to the overhead of multiple passes. It also requires careful tuning of the latent size and number of iterations. Compared to convolutional networks like [ResNet](https://www.wikiprompt.org/wiki/residual-network), the Perceiver lacks spatial locality, which can make it less sample-efficient on small datasets, but it compensates with flexibility in input formats. ## Impact and Legacy The Perceiver has influenced subsequent research in efficient attention mechanisms and general-purpose perception models. Its ideas have been incorporated into later architectures, such as the Perceiver-based components in multimodal models and the use of latent bottlenecks in other domains. While not as widely deployed as [transformers](https://www.wikiprompt.org/wiki/transformer) in production systems, the Perceiver remains an important reference point for handling arbitrary data types in [deep learning](https://www.wikiprompt.org/wiki/deep-learning). As of 2025, the Perceiver is primarily used in academic research and specialized applications, with no major commercial deployments by large tech companies. Its principles continue to inform work on scalable and modality-agnostic models in the field of [machine learning](https://www.wikiprompt.org/wiki/machine-learning). --- Source: https://www.wikiprompt.org/wiki/perceiver License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Last updated: 2026-09-12T16:28:56.219114+00:00