Evolutionary Self-Supervised Clustering Maximizes Image Surprise Score

Most deep learning approaches, including self-supervised image clustering models, rely on gradient descent, which necessitates a clearly defined target or loss function for every optimization step. This research challenges that paradigm by proposing a self-supervised framework that eliminates the need for such explicit per-step targets in image clustering. The core innovation is the concept of a "surprise score." Starting with the null hypothesis that each pixel is independently and identically distributed, the surprise score measures how unlikely a model's output representation would be under this random assumption. Maximizing this score forces the deep learning model to discover non-random, meaningful features within the data. Since a surprise score cannot generally be reduced to a per-step loss, the framework employs a "converge-to-surprise" scheme. This involves an outer loop driven by an evolution strategy (ES), which directly maximizes the surprise score without needing its gradient. This ES loop is paired with a periodic inner loop using gradient descent, which leverages the surprising clusters already identified by ES as surrogate targets. This hybrid approach achieves new state-of-the-art results in non-parametric self-supervised image clustering on standard benchmarks.