发现任务索引帖
待完善
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
可解释性学习索引帖
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
降维模型索引帖
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
聚类模型索引帖
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
非参数贝叶斯模型索引帖
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
隐因子模型索引帖
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
状态空间模型索引帖
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
概率图学习索引帖
#refplus, #refplus li{
padding:0;
margin:0;
list-style:none;
};
document.querySelectorAll(".refplus-num").forEach((ref) => {
let refid = ref.firstChild.href.replace(location.origin+location.pathname,'');
let refel = document.querySelector(refid);
let refnum = refel.dataset.num;
let ref_content = refel.innerText.replace(`[${refnum}]`,'');
tippy(ref, {
content: ref_content,
...
多视图表示学习概览
【摘 要】 表示学习是一种训练机器学习模型的特殊类型,它学着将原始的输入数据变换为对实现未来任务更有用的新形式。近年关于表示学习的研风头日胜,因为在很多实际工作中,增加预训练以学习有用的表示,确实提升了很多下游任务的性能。本文主要对表示学习的门类和方法做一概述,文章内容主要来自 Murphy 的《Machine Learning: Advanced Topics》第 32 章。
【参 考】 李沐老师讲论文系列
自监督学习有生成式学习和对比学习,对比学习需要从无标注的数据中学习特征表示,并用于下游任务中。指导原则是: 通过构造相似实例和不相似实例,学习一个表示学习模型,使得相似的实例在投影空间中较接近,不相似的实例在投影空间中距离较远。
对比学习有三个关键问题:
正负样本的构造
编码器的设计
Loss函数的选取。
过去几年,尤其是2018年开始到现在,对比学习在计算机视觉领域的发展可以划成四个阶段:
2018~2019年中,Inst Disc、CPC、CMC等方法和模型都还没有统一,目标函数和代理任务也还没有统一;
2019~2020年中,SimCLR、Moco、CP ...