Papers, patents, collaborative research

In machine learning, bootstrap methods used to estimate uncertainty are generally quite costly to apply directly in deep learning due to issues such as training time. This paper presents a technique that approximates the effect of bootstrapping in deep learning without increasing the number of parameters or requiring repetitive training, thereby enabling efficient and effective uncertainty estimation.

A model that generates a talking face from speech input. It applies a technique that extracts, stores, and uses shared information between audio features and lip-shape features through an Audio-Lip Memory module.

In general, when performing speech recognition, it is common for various types of noise to be mixed into the speech, which becomes a major obstacle to applying speech recognition models in real-world settings. This paper presents a technique that, in such situations, takes facial input of the speaker along with the audio and uses lip information from the image to enable stable speech recognition even under severely noisy conditions.

Bootstrap, which is used to estimate uncertainty in machine learning, is generally too costly to apply directly in deep learning due to issues such as training time. This paper presents a technique that can approximate the effect of bootstrapping in deep learning without increasing the number of parameters or requiring repeated training, enabling efficient and effective uncertainty estimation.