Converge to Surprise: Evolutionary Self-supervised Image Clustering

2026-07-09 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The "Converge-to-Surprise" framework introduces a novel self-supervised image clustering approach that operates without requiring a per-step loss function. It defines a "surprise score" based on Kullback–Leibler divergence, measuring how unlikely a model's output representation is under a null hypothesis of i.i.d. random pixels. The optimization combines an evolution strategy (ES) outer loop, which directly maximizes this non-differentiable surprise score, with a periodic gradient-descent inner loop that uses already discovered surprising clusters as surrogate targets. This hybrid method achieves new state-of-the-art results in non-parametric self-supervised image clustering on standard benchmarks like MNIST and USPS, and competitive performance on Fashion-MNIST. Crucially, the model is trained from scratch and accurately infers the number of active clusters, such as 10.0 on MNIST and 10.4±0.5 on USPS, without prior knowledge of ground-truth classes.

Key takeaway

For Machine Learning Engineers developing unsupervised image clustering solutions, especially in non-parametric settings where the number of classes is unknown, you should consider hybrid optimization strategies like "Converge-to-Surprise." This framework demonstrates that combining evolution strategies with periodic gradient descent can effectively discover meaningful, hard representations from raw images without relying on traditional per-step loss functions, offering a robust alternative to purely loss-based methods.

Key insights

Converge-to-Surprise optimizes self-supervised image clustering by maximizing a non-differentiable "surprise score" via a hybrid evolution strategy and gradient descent.

Principles

• Learning can be viewed as iteratively rejecting a null hypothesis of random data.
• Maximizing "surprise" (non-randomness) naturally prevents representation collapse.
• Orderliness (neg-entropy) quantifies learning by imposing strong order on diverse marginals.

Method

The "converge-to-surprise" scheme uses an evolution strategy outer loop to maximize a non-differentiable surprise score, paired with a periodic gradient-descent inner loop that leverages discovered surprising clusters as surrogate targets.

In practice

• Employ chessboard masking for complementary image views.
• Independent augmentation of views breaks low-level feature shortcuts.
• Anisotropic augmentation helps with class-imbalanced datasets like USPS.

Topics

• Self-supervised Learning
• Image Clustering
• Evolution Strategy
• Gradient Descent
• Non-parametric Models
• Surprise Score
• Deep Learning

Code references

• canlinzhang/converge-to-surprise

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

Related on AIssential

• Converge to Surprise: Evolutionary Self-supervised Image Clustering — Self-supervised image clustering can achieve state-of-the-art results by maximizing a "surprise score" without per-step loss targets, using an evolutionary strategy.
• A Parameter-free Adaptive Resonance Theory-based Topological Clustering Algorithm Capable of Continual Learning — A new ART-based clustering algorithm autonomously estimates critical parameters for superior, adaptive performance.
• CLIMB: Centroid-Based Hierarchical Memory for Online Continual Self-Supervised Learning — CLIMB combines hierarchical centroid-based memory with knowledge distillation for robust online continual self-supervised learning.
• Singularity-aware Optimization via Randomized Geometric Probing: Towards Stable Non-smooth Optimization — S-Adam stabilizes non-smooth deep learning optimization by dynamically adjusting step sizes based on local geometric instability, preventing gradient chattering.
• Unsupervised Feature Selection via Nonnegative Orthogonal Constrained Regularized Minimization — NOCRM offers a theoretically sound unsupervised feature selection method with convergence guarantees.
• EvoForest: A Novel Machine-Learning Paradigm via Open-Ended Evolution of Computational Graphs — EvoForest evolves computational graphs and functions, guided by diagnostics, to discover optimal computations for structured prediction.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.