当前位置：网站首页>Study notes of "deep learning" (VII)

Study notes of "deep learning" (VII)

2022-04-21 12:00:00 【Bai Yanling】

This post mainly introduces AUTO-ENCODER.

List of articles

self-supervised learning Framework
Basic Idea of Auto-encoder
- Why Auto-encoder?
Feature Disentanglement
Discrete Latent Representation

self-supervised learning Framework

Insert picture description here

Basic Idea of Auto-encoder

Insert picture description here
old feature adopt NN Encoder High dimension becomes low dimension ,new feature for downstream tasks. Low dimensional through NN Decoder Become high dimensional . High dimension becomes low dimension, which is also called Dimension Reduction.

Why Auto-encoder?

Insert picture description here

Also can put the BERT Think of it as a De-noising Auto-encoder

Feature Disentanglement

Insert picture description here

This technology means that the original input audio In two parts , The previous dimension represents content information, The latter dimension represents speaker information.

Some dimensions represent content, Some dimensions represent speaker.

Discrete Latent Representation

Insert picture description here

representation It can also be text, however train It won't work when it comes to .

add GAN And then ok Of , In fact, that is cycle GAN.

The difficulty of anomaly detection is to collect data , It is difficult to collect outliers .

There are many techniques for anomaly detection , The use of auto-encoder To do anomaly detection .