MIT+ML06 Intro to Neural Networks

Intro to Neural Networks

Overall Process:

Data-> Feature Extraction -> Classifier

Neural Networks:直接从数据中获取特征进行学习，无需进行特征提取。

That is “turn a feature extraction by hand into a ML task“

简单的分类问题：

1. get a linear model: $h(x)=w^Tx+b$

2. get a linear model + sigmoid activation: $h(x)=\sigma (w^Tx+b)$

3. set threshold to do classifier :
   $$
   y_{predicted}=\begin{cases}0 && if h(x)<0.5\1 &&if h(x)\ge0.5\end{cases}
   $$
   在这里，如果是多分类问题呢，我们就去用softmax。

Simple Overview of neural network models:

Raw Data -> a bunch of experts( classifier) ->Decisions

each unit is like an expert, performing a simple task, 这个task 是什么呢，就是去做，$w^Tx+b$, a simple computational part.

Feed these decisions to the next level…

• first layer:第一层

  x->$h(x)=\sigma (w_1^Tx+b_1)$ Unit1,bias1

  x->$h(x)=\sigma (w_2^Tx+b_2)$ Unit2,bias2

  x->$h(x)=\sigma (w_3^Tx+b_3)$ Unit3,bias3

  …

  将第一层的这个计算过程，合并，即是：

  x->h(x)=$\sigma(W^Tx+b)$,这里，W是权重矩阵，b是偏置向量。

• 然后，第二层，第三层，第四层以此类推。

• 注意，权重矩阵的notation $W_{ij}$表示从第i层到第j层连接的偏执，所以说对于第i层的unit来说，他们面对的不是权重矩阵，也是权重矩阵的某一列。

• 

上面就是传说中的 feed-forward deep neural network.

Some intuition

• Not only better classification + better feature(data)encoding ,意味着更多计算力(resources)的消耗。
• NN有分治的思想：feed simple decisions into next stage for more complex decisions.
• Learn to encode on the way to learn to do feature extraction and better classificaiton.

NN architecture

注意，这里units aka “neurons”

neurons 要做的不止是linear model

• Each unit performs a task such as linear classification/detection of magma/linear transformations of its input/scaling/normalization

Jargon in NN

ReLU activation function

ReLU activiation function is also called a rectifier, and a unit that uses the activation if called a Rectified Linear Unit.

Desicion surface

NN always require a trade-off between computer resources and neurons nuts.

hyperparameter search is always a key point and also means sophisticating design of architectures.

Nonlinear separation

Representation power of NN

How power are deep models?

任何function都可以用只有一个hidden layer的NN去表示

Universal approximation theorem.

既然越深越好，那为啥不深呢:why deep instead of big?

Size<->power的 trade-off

因为size太大，由于计算力的限制，我们失去了效率。

Illustrative experiment

赏

谢谢你请我吃糖果❤️

支付宝