GPGPU Linear Complexity t-SNE Optimization.

Area of Science:

• Computational Science
• Data Visualization
• Machine Learning

Background:

• t-distributed Stochastic Neighbor Embedding (t-SNE) is a key technique for high-dimensional data exploration.
• Its computational complexity limits application to smaller datasets.
• Existing scalable t-SNE methods struggle with interactive visualization of large datasets.

Purpose of the Study:

• To develop a computationally efficient t-SNE algorithm for large datasets.
• To leverage graphics hardware for accelerated t-SNE computations.
• To maintain or improve accuracy while drastically reducing runtime.

Main Methods:

• A novel approach to minimizing the t-SNE objective function using graphics hardware.
• Approximation of repulsive forces via kernel texture splatting.
• Reformulation of the t-SNE minimization as tensor operations executable on GPUs.

Main Results:

• Achieved linear computational complexity for t-SNE.
• Reduced computational cost by orders of magnitude for large datasets.
• Retained or improved accuracy compared to previous approximation techniques.

Conclusions:

• The proposed method significantly enhances the scalability of t-SNE.
• Hardware acceleration enables interactive visualization of large-scale high-dimensional data.
• The technique is implemented in TensorFlow.js and C++ for broad accessibility.