Skip to content

Commit

Permalink
Site updated: 2024-03-11 12:32:47
Browse files Browse the repository at this point in the history
  • Loading branch information
lihaibineric committed Mar 11, 2024
1 parent 7cdcf12 commit dee4bcd
Show file tree
Hide file tree
Showing 3 changed files with 24 additions and 6 deletions.
19 changes: 14 additions & 5 deletions 2024/02/04/dl_rec_al/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,7 @@
<meta property="og:image" content="https://gitee.com/lihaibineric/picgo/raw/master/pic/image-20240206000626255.png">
<meta property="og:image" content="https://gitee.com/lihaibineric/picgo/raw/master/pic/image-20240206000903122.png">
<meta property="article:published_time" content="2024-02-03T16:37:51.000Z">
<meta property="article:modified_time" content="2024-03-08T03:12:57.935Z">
<meta property="article:modified_time" content="2024-03-11T04:19:54.925Z">
<meta property="article:author" content="Haibin Li">
<meta property="article:tag" content="人工智能">
<meta property="article:tag" content="深度学习">
Expand Down Expand Up @@ -231,7 +231,7 @@
<span class="post-meta mr-2">
<i class="iconfont icon-chart"></i>

43k words
44k words

</span>

Expand All @@ -242,7 +242,7 @@



360 mins
363 mins

</span>

Expand Down Expand Up @@ -284,7 +284,7 @@ <h1 id="seo-header">【深度学习】推荐系统算法总结</h1>
<p class="note note-info">


Last updated on March 8, 2024 am
Last updated on March 11, 2024 pm


</p>
Expand Down Expand Up @@ -1088,6 +1088,15 @@ <h2 id="深度学习模型rank">深度学习模型(Rank)</h2>
alt="image-20240308111249239" />
<figcaption aria-hidden="true">image-20240308111249239</figcaption>
</figure>
<figure>
<img
src="https://gitee.com/lihaibineric/picgo/raw/master/pic/image-20240311120735173.png" srcset="/img/loading.gif" lazyload
alt="image-20240311120735173" />
<figcaption aria-hidden="true">image-20240311120735173</figcaption>
</figure>
<p>对于1层cross和1层deep的DCN网络输入经过embedding和stack处理后维度为d,Cross部分网络参数为2d,Deep为d*d,当MLP的层数增多时,deep部分的参数量也急速增加。DCN网路的绝大部分参数都用于对隐性交叉特征进行建模。Cross部分的表达能力反而受限。</p>
<p>DCN-v2优化了cross网络的建模方式,增加了cross网络部分的表达能力;deep部分保持不变。</p>
<p>低维空间的交叉特征建模使得我们可以利用MoE。MoE由两部分组成:experts专家和gating门(一个关于输入x的函数)。我们可以使用多个专家,每个专家学习不同的交叉特征,最后通过gating将各个专家的学习结果整合起来,作为输出。这样就又能进一步增加对交叉特征的建模能力。</p>


</div>
Expand Down Expand Up @@ -1156,7 +1165,7 @@ <h2 id="深度学习模型rank">深度学习模型(Rank)</h2>

<div class="license-meta-item license-meta-date">
<div>Updated on</div>
<div>March 8, 2024</div>
<div>March 11, 2024</div>
</div>


Expand Down
2 changes: 1 addition & 1 deletion local-search.xml

Large diffs are not rendered by default.

9 changes: 9 additions & 0 deletions search.xml
Original file line number Diff line number Diff line change
Expand Up @@ -3683,6 +3683,15 @@ src="https://gitee.com/lihaibineric/picgo/raw/master/pic/image-20240308111249239
alt="image-20240308111249239" />
<figcaption aria-hidden="true">image-20240308111249239</figcaption>
</figure>
<figure>
<img
src="https://gitee.com/lihaibineric/picgo/raw/master/pic/image-20240311120735173.png"
alt="image-20240311120735173" />
<figcaption aria-hidden="true">image-20240311120735173</figcaption>
</figure>
<p>对于1层cross和1层deep的DCN网络输入经过embedding和stack处理后维度为d,Cross部分网络参数为2d,Deep为d*d,当MLP的层数增多时,deep部分的参数量也急速增加。DCN网路的绝大部分参数都用于对隐性交叉特征进行建模。Cross部分的表达能力反而受限。</p>
<p>DCN-v2优化了cross网络的建模方式,增加了cross网络部分的表达能力;deep部分保持不变。</p>
<p>低维空间的交叉特征建模使得我们可以利用MoE。MoE由两部分组成:experts专家和gating门(一个关于输入x的函数)。我们可以使用多个专家,每个专家学习不同的交叉特征,最后通过gating将各个专家的学习结果整合起来,作为输出。这样就又能进一步增加对交叉特征的建模能力。</p>
]]></content>
<categories>
<category>深度学习</category>
Expand Down

0 comments on commit dee4bcd

Please sign in to comment.