---
title: Deep Learning for everyone
tags: APAC HPC-AI competition
---

[TOC]

[Python Module(有Numpy、Pandas)](https://hackmd.io/le4IndLnQqSPfzQldPJ8iw?view)

## Machine Learning
Data -> Linear Algebra -> Answer

## How the Brain Works
- Each neuron receives input from other neurons
- Weight can be `+ | -`
- Different areas perform different functions using same structure(**Modularity**)

Our brain is very modular!!
Able to adapt to various input.

**Motivation: We are to build an artificial brain!**

## Some terms
Some terms of input
- Features
- Prediction
- Attributes
- Predictable Variables

Machine
- Algorithm
- Technique
- Models

Some terms of output
- Classes
- Responses
- Targets
- Dependant Variables

## Perceptron
inputs->mutiply(the respective weights)->outputs

## Training Procedure
- If **correct**, do nothing.
- If **incorrectly** outputs 0, **add** inputs to weight vector.
- If **incorrectly** outputs 1, **subtract** inputs to weight vector.

**Guaranteed to converge**, if a correct set of weights exists.
finding and choosing the right features -> ==MAGIC==!!

---
### Resources reference
[Github stuff](https://github.com/DataForScience/DeepLearning)

#### 1. Perceptron
**Logic function**

---
#### Q&A
Q: **Why XOR requires 2 different layers?**
A: (X AND Y) OR (NOT~X AND NOT~Y)
> 2 steps(layer) to simulate XOR.

---
## Optimization Problem
> Keep calm and start optimizing

Have three distinct pieces:
1. The constraints
2. THe function to optimize
3. THe optimization algorithm

(p.25)

## Linear Regression
What is linear regression?
> Linear Regression is a predictive algorithm which provides a Linear relationship between Prediction (Call it ‘Y’) and Input (Call is ‘X’).
>
(p.27)

**(to be studied)**

lasso vs. ridge vs. linear regression

:::danger
The line for linear regression must be straight!!
:::

## Learning Procedure

## Logistic Regression (Classification)
- Predict the P of instance belonging to the given class

- Error function
- plug in the sigmoid function as ø()
- Gradient

**comparison**
Linear Regression v.s. Logistic Regression

Activation function - Sigmoid
[Why We Use the Sigmoid Function in Neural Networks for Binary Classification](https://www.youtube.com/watch?v=WsFasV46KgQ)
[video](https://www.youtube.com/watch?v=aircAruvnKk)

Q: What are we doing in the hidden layer?
A: Creates different representations of the input, transforming the input to something that's more abstract that still contains the information you need to improve your results by adding some processing layers.

### HW
- [ ] Benchmark
- [ ] [Prove Cross Entropy](https://towardsdatascience.com/understanding-binary-cross-entropy-log-loss-a-visual-explanation-a3ac6025181a)
- [ ] Gradient Descent

(Time 156 min)

---

neurons: 帶有數字，一般為0~1 -> (Data) -Input-> Function -Output-> (Number)

What parameter should exists?
- Weights
- $w_1a_1 + w_2a_2 + ... w_na_n$
- 可能為任意數
- 使用activation fun壓進特殊範圍，例: sigmoid fun -> [0, 1]
- Bias
- 限制權重和要到哪個限度才有意義
- 像是一個開關

二
9
14
17

2
8

neurons that
fire together
wire together