HiRo: A Compact Four-Directional Hierarchical Reservoir Token-Mixer for Efficient Image Classification

2026-06-13 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

HiRo, a new parameter-efficient image classification model, addresses the challenge of balancing local feature modeling, cross-window interaction, and parameter count in high-performing architectures. It integrates shifted-window partitioning with multi-directional hierarchical reservoir computing. The model processes images by dividing them into patches, applying 2D sinusoidal positional encodings, and scanning tokens in four directions within local windows. A two-stage slice-and-mix reservoir module, featuring fixed reservoirs with trainable closed-loop readouts, processes these directional sequences. Consecutive blocks alternate between regular and shifted windows for cross-window interaction. HiRo achieves 99.46% accuracy on MNIST, 85.57% on CIFAR-10, and 59.10% on CIFAR-100, all while utilizing under 1M trainable parameters and significantly less memory and time compared to transformer-style baselines.

Key takeaway

For Machine Learning Engineers developing efficient image classification models for resource-constrained environments, you should consider HiRo's approach of combining hierarchical reservoir computing with shifted-window partitioning. This design offers competitive accuracy (e.g., 99.46% on MNIST) with under 1M parameters, significantly reducing memory and computational costs compared to transformer-style baselines. Evaluate its fixed reservoir and multi-directional mixing for your next project requiring high performance with minimal overhead.

Key insights

HiRo combines hierarchical reservoir computing and shifted windows for efficient, high-accuracy image classification with few parameters.

Principles

• Fixed reservoirs with trainable readouts reduce parameters.
• Alternate regular and shifted windows for cross-window interaction.
• Multi-directional token scanning enhances feature learning.

Method

Images are patched, encoded, and processed in local windows. Tokens are scanned four directions, passed through a two-stage slice-and-mix reservoir, then fused and averaged for classification.

In practice

• Implement fixed reservoirs for parameter efficiency.
• Use shifted windows for cross-window context.
• Apply 2D sinusoidal positional encodings.

Topics

• Image Classification
• Reservoir Computing
• Token-Mixer
• Parameter Efficiency
• Computer Vision
• Shifted Windows

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• MixerCA: An Efficient and Accurate Model for High-Performance Hyperspectral Image Classification — MixerCA efficiently classifies hyperspectral images by integrating depthwise convolutions, token/channel mixing, and coordinate attention.
• RECAP: Local Hebbian Prototype Learning as a Self-Organizing Readout for Reservoir Dynamics — Hebbian co-activation prototypes in reservoir computing enable robust image classification without backpropagation.
• KAN-MLP-Mixer: A comprehensive investigation of the usage of Kolmogorov-Arnold Networks (KANs) for improving IMU-based Human Activity Recognition — Strategically combining KANs for data embedding and classification with MLPs for feature mixing significantly enhances noisy IMU-based HAR.
• Improved Knowledge Distillation for Land-Use Image Classification — Combining hard and soft supervision with KL divergence and Cosine Similarity improves knowledge distillation for land-use classification.
• HASTE: Hardware-Aware Dynamic Sparse Training for Large Output Spaces — HASTE optimizes XMC output layers via group-shared sparsity and a dense/sparse decomposition for hardware-aware speedups.
• Neural Field Tokenizations with Hierarchy and Spatial Locality Priors — LH-NeF efficiently learns general-purpose neural field representations by integrating locality and hierarchy priors via a single forward pass.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.