We developed an unsupervised learning model for generic high dimensional data. The model demonstrated exceptional performance across diverse data modalities from neural recording to gene expression.

Constructed a two-layer interpretable unsupervised learning model based on principles from neural computation. The model remains competitive with deep learning methods.

We used sparse coding to visualize Large Language Models (LLMs). By adopting neural computation principle, we greatly improve the interpretability of LLMs. Our result implies LLMs compute in a manner akin to the human brain, where it builds representations to predict upcoming language at multiple levels of abstraction.

We model visual system with a chromatic aberration constrained optical simulation and designed a learnable cone mosaic sampling. We show that the model’s emerged cone mosaic resembles a cone mosaic found in humans, when optimized for both spatial acuity and color acuity.

We applied the principles of sparsity and slowness to build representations for natural scenes. Then used coupled Hopfield network dyanmic to learn motion patterns and shape patterns.

We proposed a stability-driven autoencoder for unsuperised learning. The method is used for extracting gene expression patterns.