algorithm – by Xiongjie Yu

Redundancy in the weight matrix of a neural network can be much reduced with a tensor network formulation. More than 99% compression rate can be achieved while maintaing accuracy. Tensor train representation of a neural network is compact. This formalism may allow neural networks to be trained on mobile devices. **Read More ›**

algorithm – by Brian Busemeyer

Simulated annealing is a method for optimizing noisy, high-dimensional objective functions using ideas from materials science. **Read More ›**

algorithm – by Benjamin Villalonga Correa

Problems in audio compression form just a subset of the general problem of signal compression, and general techniques can well be applied to solve them. However, it is possible to benefit greatly from being aware of the very particular way in which the human brain perceives and interprets sound, being able to optimize compression techniques to keep only information that is relevant to human perception. In this presentation, I focus on speech compression, and more particularly on an implementation using a Linear Predicting Model (LPM). The LPM provides a very efficient way of reconstructing a signal from a very small set of compressed data (up to 95% of data can be neglected), generating a sythesized speech that keeps the original phonemes and the quality of the voice of the speaker, who can be recognized easily. This technique has been used in telephony applications. **Read More ›**

algorithm – by Will Wheeler

Recognize a million words per second. **Read More ›**

algorithm – by Yubo 'Paul' Yang

The basic Gibbs sampler samples a joint probability distribution one variable at a time. Each random variable is sampled from its full conditional probability distribution with all other variables fixed. Independent variables can be sampled simultaneously, making the Gibbs sampler ideal for the restricted Boltzmann machine. **Read More ›**