1. この資料は「第3回 IEEE SIGHT ハックチャレンジ 」のために作
られました。 別の目的での使用には、下記の引用が必要です:
Tejero-de-Pablos A. (2018). 機械学習の基礎 [PowerPoint slides].
Retrieved from
https://www.slideshare.net/AntonioTejerodePablo/machine-learning-
fundamentals-ieee
This material was originally created for the “3rd IEEE SIGHT Hack
Challenge” event. If used for a different purpose, the following
citation is necessary:
Tejero-de-Pablos A. (2018). 機械学習の基礎 [PowerPoint slides].
Retrieved from
https://www.slideshare.net/AntonioTejerodePablo/machine-learning-
fundamentals-ieee
17. 最急降下法
• モデルパラメーターwを更新し、ロスを徐々に減少
• Regression問題では、loss vs weightの凸関数
• Gradient(勾配): ロスを小さくするwの更新方向を示す
• 2乗誤差などの簡単なロスの勾配は簡単に計算できる
w
ロス
初期値
(ランダム)
gradient: 方向と大きさ
learning rate
What if the learning rate is too big?
What if the learning rate is too small?
What is the ideal learning rate?
16
61. 参考文献
本
• 機械学習
• C. Bishop, “Pattern Recognition and Machine Learning”
• ディープラーニング
• I. Goodfellow, “Deep learning”
• コンピュータビジョン
• 原田達也, “画像認識”
オンラインコース
• Google
• https://developers.google.com/machine-learning/crash-course/
• Coursera
• https://www.coursera.org/learn/machine-learning 60
Notas do Editor
・Artificial Intelligence is the broader concept of machines being able to carry out tasks in a way that we would consider “smart”.
・Machine Learning is a current application of AI based around the idea that we should really just be able to give machines access to data and let them learn for themselves.
・Deep learning: The machine is able to understand a broader set of cases. Greater generalization.
Deciding if a mail is a spam or not
Not solvable by people: Predicting the stock market
Generalizable: Same model can distinguish, “dogs from cats” and “birds from flowers”
Think as a scientist: Think the fundamentals of the problem instead of the implementation
Predicting learned data is 当たり前
In this lecture, we will focus on supervised learning
For example: predicting the cost of a house would be classification or regression? Predicting if a movie will be successful or not?
学習のプロセスを詳しく見てみましょう
How do you train a model? How do you decide these w values?
Shuusoku
Data is the fuel (nenryou) to our machine learning model
Getting 100% accuracy with 3 instances is not meaningful
You cannot keep low values only in your training set and try to predict high values
sengen
Learning English from a teenager
Seizokuka
Doing trial and error (chousei) with the test data is not good
Suuchi
Seikika
By knowing your data you can strategize better:
Is the dataset imbalanced?
Should I normalize?
Hint: This is a non-linear problem
Katayori ga aru
The lower the threshold, the better the recall. The higher the threshold, the better the precision Tradeoff
Kasseika kansuu
From binary to multiclass
What is the size of the input for fc1?
Gradient descent strategy for updating our model. What algorithm is used for learning? Backpropagation
Do you know what it is called deep?
Konnan
Kernel varies in sizes and jumps
Well-known datasets allow researchers to compare their methods fairly