Tensorial Biometric Signal Recognition Based on Feed Forward Neural Networks with Random Weights

Most biometric signals are naturally multi-dimensional objects, which are formally known as tensors. How to classify this kind of data is an important topic for both pattern recognition and machine learning. Commonly, these biometric signals are often converted into vectors in the process of recognition. However, the vectorization usually leads to the distortion of the potential spatial structure of the original data and high computational burden. To solve this problem, in this paper, a novel classifier as a tensor extension of neural networks with random weights (NNRW) for tensorial data recognition is introduced. Due to the proposed solution can classify tensorial data directly without vectorizing them, the intrinsic structure information of the input data can be reserved. Moreover, compared with the traditional NNRW, much fewer parameters need to be calculated through the proposed tensor based classifier. Extensive experiments are carried out on different databases, and the experiment results are compared against state-of-the-art techniques. It is demonstrated that the new tensor based classifier can get better recognition performance with an extremely fast learning speed.