Project description/goals

Recent work reported improved results from the sparse binary projections over the classical locality sensitivity hashing methods. Our work reported that the accuracy of sparse binary projections can be studied under a supervised setting, which can be greatly simplified via a relaxation on the winner-take-all function. The significant improvement in both speed and accuracy demonstrated high potential in applying sparse binary projection in real applications.

Importance/impact, challenges/pain points

Learning Projection Matrix Relaxation of Winner-Take-All Function Real Application and Improved Results

Solution description

The key to the solution is through the proper modeling and relaxation of the winner-take-all function, which admits a fast yet more accurate solution to obtaining the projection matrix. The work also establishes the connection between the proposed model and the classical clustering method.

Key contribution/commercial implication

In real applications, the model motivates the design of new information retrieval methods with improved results.

Next steps

Further investigations along the line will be carried out, especially applications in information retrieval.

Collaborators/partners

The Chinese University of Hong Kong, Shenzhen

Team/contributors

Wenye Li (PI)

Changyi Ma (Ph.D student, graduated in 2022)

Fangchen Yu (Ph.D student)