Notes about image processing.

Modules

• Datasets
• Data Augmentation
• Metrics
• Loss Function
• Optimization

Papers read during research forward deep learning algorithm.

Quick notes

• 正推，逆推
  • 底层精度高=> 高层精度高
  • 相同结构网络，精度高的网络，从中间截取进行分类是否精度也高？
• 不同粒度特征提取的结合
  • 粒度，卷积层数？ 层数不同的实质是什么不同？
• 在调整了一些Hyper-paramenter后，大致上可以发现影响更大的参数，比如卷积核数
  • 调整优先级：欠拟合 > 过拟合
• 用CIFAR-10训练时，测试集上的loss会在某次迭代中突然丢失，然后又恢复，形成一个尖刺？
• 将问题分割成子问题，但试图用深度学习解决的问题，都不太好分割成子问题

Orthogonal Bipolar Target Vectors¹

Can OBV construct a middle target for CNN?

A kind of target representation.

• conventional
  • BNV - binary: \((0, 0, 1, 0, 0)\)
  • BPV - bipolar?: \((-1, -1, 1, -1, -1)\)
• OBV - orthogonal bipolar vectors
• NOV - Non-Orthogonal Vecotrs
  • For fail comparision
  • \(V_i=(\overbrace{-1 , \cdots , -1}^{i-1}, 1, \overbrace{-1 , \cdots , -1}^{n-i})\)
  • \(cos \theta = \frac{n-2}{n}\)
• degraded characters?
  • They use degraded license plate images as expirement data. (车牌号)

How to generate OBV from conventional target?

Learning notes of Hinton’s Neural Networks for Machine Learning in Coursera.

1. Introduction

1a. Why do we need machine learning

What, Why

• We don’t know how to program to solve some problems or the program might be very complicated.
• Rules may need to change frequently, like recognizing fraud.
• Cheap computation

How

• Collect lots of cases with inputs and correct outputs.
• ML algorithms takes examples and produces a program to do the job.

Good at

• Recognizing patterns
  • Objects
  • Face
  • Spoken words
• Recognizing anomalies (unusual)
  • Transactions
  • Sensor readings in a nuclear power plant
• Prediction
  • Stock prices, exchange rates
  • Movie recommendations