DOWNLOAD RESUME

MoMos 2D

Project Overview: Algorithmic Complexity in Neural Networks

Recent research has explored the relationship between Neural Network (NN) complexity and Kolmogorov complexity (Sakabe et al., 2025; Bakhtiarifard et al., 2026).

1. Complexity and Generalization

Specifically, Sakabe et al. (2025) demonstrate that approximating the Kolmogorov complexity of Binarized NNs provides critical insights into training dynamics and correlates strongly with generalization capabilities.

2. The Mosaic-of-Motifs (MoMo) Framework

Building on these foundations, the Mosaic-of-Motifs (MoMo) framework (Bakhtiarifard et al., 2026) provides a reliable method to bound Kolmogorov complexity from above by decoupling a model's weights from its architecture.

Key Advantage: MoMo is highly flexible and can be applied to any architecture. It often yields better compression rates than standard quantization by enforcing algorithmic simplicity directly during the training phase.

3. Proposed Enhancements

The aim of this project is to enhance the MoMo framework by introducing more expressive mappings. Our research focuses on two primary directions:


References