Flattening Images Manifold at the Nishimori Point: Exploring Linearity in Polysemantic Embedding for Human-Aware Bayesian CNN Image Classification

Efficient image classification often requires a balance between accuracy and computational complexity. In this paper, we propose a graph-based spectral embedding method that leverages improved Nishimori temperature estimation to achieve low-dimensional feature representations while maintaining high classification performance. Our approach reduces image features extracted by MobileNetV2 to a 32-dimensional manifold using graph embeddings, significantly decreasing the parameter count compared to conventional methods like bottleneck ResNet-18. On the ImageNet10 dataset, the proposed method achieved 98% accuracy with only 32 dimensions, while on the ImageNet-100 dataset, it matched the performance of a 64-dimensional bottleneck layer using half the dimensionality. Furthermore, with polysemantic multigraph embeddings exceeding 64 dimensions, the method demonstrated a 1.5% accuracy improvement. The computational efficiency, achieved by reducing parameters from 51200 to 3200, highlights the potential of this method for resource-constrained applications. These results demonstrate that the proposed graph-based embedding framework is a scalable and effective solution for image classification tasks, offering a compelling alternative to conventional deep learning architectures. Future work will extend this approach to coarsely quantized feature extracting, improved discriminators, Diffusions and Transformers neural networks.