Please write your contents for 17/1/30.
Hands-On Machine Learning with Scikit-Learn and TensorFlow
Aurélien Géron
Preface
pp.xiii
The Machine Learning tsunami
In 2006, Geoffrey Hinton et al. published a paper1 showing how to train a deep neural network
1 Available on Hinton’s home page at http://www.cs.toronto.edu/~hinton/.
Machine Learning in your projects
3 Find many more examples at https://www.kaggle.com/wiki/DataScienceUseCases.
事例(Healthcare,Lifesciences,Agriculture)
pp.xiv
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Healthcare
Claims review prioritization クレーム審査の優先順位付け
payers picking which claims should be reviewed by manual auditors手動請求者が審査する必要がある請求者を選ぶ支払人
Medicare/medicaid fraud メディケア/メディケイド詐欺
Tackled at the claims processors, EDS is the biggest & uses proprietary tech
請求プロセッサーに取り組むEDSは最大のもので、独自の技術を使用しています
Medical resources allocation 医療資源配分
Hospital operations management 病院運営管理
Optimize/predict operating theatre & bed occupancy based on initial patient visits最初の患者の訪問に基づいて手術場とベッド占有率を最適化/予測する
Alerting and diagnostics from real-time patient data リアルタイムの患者データからの警告および診断
Embedded devices (productized algos) 埋め込みデバイス（製品化）
Exogenous data from devices to create diagnostic reports for doctors医師のための診断レポート作成デバイスからの外因性データ
Prescription compliance 処方のコンプライアンス
Predicting who won't comply with their prescriptions 処方箋に従わない者を予測する
Physician attrition 医師の麻痺
Hospitals want to retain Drs who have admitting privileges in multiple hospitals 病院は複数の病院で特権を認めているDrsを保持したい
Survival analysis 生存分析
Analyse survival statistics for different patient attributes (age, blood type, gender, etc) and treatments
異なる患者属性（年齢、血液型、性別など）および治療法の生存統計を分析
Medication (dosage) effectiveness 投薬（投薬量）の有効性
Analyse effects of admitting different types and dosage of medication for a diseas病気の薬剤種類と投与量を認める効果を分析
Readmission risk 再発リスク
Predict risk of re-admittance based on patient attributes, medical history, diagnose & treatment
患者の属性、病歴、診断および治療に基づいて再アドミタンスのリスクを予測する
Life Sciences
Identifying biomarkers for boxed warnings on marketed products 販売された製品のボックス化された警告のためのバイオマーカーの特定
Drug/chemical discovery & analysis 薬物/化学物質の発見と分析
Crunching study results 研究成果のクランチ
Identifying negative responses (monitor social networks for early problems with drugs)負反応同定（薬物による早期問題ソーシャルネットワークの監視）
Diagnostic test development 診断テスト開発
Hardware devices、Software
Diagnostic targeting (CRM) 診断ターゲティング（CRM）
Predicting drug demand in different geographies for different products 異なる製品の異なる地域における薬物需要の予測
Predicting prescription adherence with different approaches to reminding patients 異なるアプローチによる処方順守予測
Putative safety signals 推定安全信号
Social media marketing on competitors, patient perceptions, KOL feedback 競合他社のソーシャルメディアマーケティング、患者認識、KOLフィードバック
Image analysis or GCMS analysis in a high throughput manner 高スループットの画像解析やGCMS解析
Analysis of clinical outcomes to adapt clinical trial design 臨床試験設計に適応するための臨床結果の分析
COGS optimization COGS最適化
Leveraging molecule database with metabolic stability data to elucidate new stable structures 安定構造を解明する代謝データを用いた分子データベース
Agriculture
Yield management (taking sensor data on soil quality - common in newer John Deere et al truck models and determining what seed varieties, seed spacing to use etc 収量管理（土壌の質に関するセンサーデータを取得する）
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Objective and approach
This book's goal is to give you the concepts, the intuitions and the tools you need to actually implement
programs capable of learning from data.
production-ready python frameworks
scikit-learn
tensorflow (is a more complex library for distributed numerical computation using data flow graphs. multi-GPU servers.
code example
https://github.com/ageron/handson-ml.
p.xv
Prerequisites
python’s main scientific libraries, in particular NumPy, Pandas and Matplotlib.
Roadmap
1.
• What is Machine Learning?
• The main steps in a typical Machine Learning project.
• Learning by fitting a model to data.
• Optimizing a cost function.
• Handling, cleaning and preparing data.
• Selecting and engineering features.
• Selecting a model and tuning hyper-parameters using cross-validation.
• The main challenges of Machine Learning, in particular underfitting and overfitting
(the Bias/Variance tradeoff).
• Reducing the dimensionality of the training data to fight the curse of dimensionality.
• And of course we will go through the most common learning algorithms: Linear
and polynomial regression, Logistic regression, k-Nearest Neighbors, Support
Vector Machines, Decision Trees, Random Forests and Ensemble
methods.
Preface
2. Neural Networks and Deep Learning
• What are neural nets? What are they good for?
• Building and training neural nets using TensorFlow.
• The most important neural net architectures: FeedForward Neural Nets, Convolutional
Nets, Recurrent Nets, Long Short Term Memory Nets and Autoencoders.
• Techniques for training deep neural nets.
• Scaling neural networks for huge datasets.
• Reinforcement Learning.
The first part is based mostly on Scikit-Learn while the second part uses TensorFlow.
Other resources
*Andrew Ng’s ML course on Coursera4
*Geoffrey Hinton’s course on Neural Networks and Deep Learning
*Scikit-Learn’s exceptional User Guide.
*https://www.dataquest.io/ which provides very nice interactive tutorials.
*http://goo.gl/GwtU3A. http://deeplearning.net/ has a good list.
Conventions Used in This Book
Using Code Examples
Supplemental material (code examples, exercises, etc.) is available for download at
https://github.com/ageron/handson-ml.
PART I
The Fundamentals of Machine Learning
CHAPTER 1
The Machine Learning landscape
What is Machine Learning?
Machine Learning is the science (and art) of programming computers so they can
learn from data.
spam filterの場合
training set = the examples that the system uses to learn
trainint instance(or sample) = each training example
task T = to flag spam for new emails
experience E = training data
performance measure P = accuracy
Why use Machine Learning?
1.最初に、スパムが典型的にどう見えるかを調べる。
特定の単語（「4U」、「クレジットカード」、「フリー」、「アメージング」など）、送信者名、メールの本文など
2.次に、抽出単語の検出アルゴリズムを書きます。
プログラムは、単語が一定の数を超えたら電子メールをスパムとみなします
3.プログラムをテスト、ステップ1と2を繰り返し評価します。
問題は簡単ではないので、あなたのプログラムはおそらく複雑な長いリストになる。
リストを維持するのはかなり難しい。
一方、機械学習技術に基づくスパムフィルタは自動的に学習する。
プログラムははるかに短く、維持しやすく、より正確だろう。
データが更新されるとMLアルゴリズムも更新されるので、人手介入なしに処理を自動化できる。
機械学習は、人間が学習するのに役立ちます。
時にはこれは予期しない相関を明らかにして、問題の理解につながります。
大量のデータを掘り起こして見えなかったパターンを発見する事をデータマイニングと呼びます。
機械学習の良い点まとめ
• 多くの場合コードを簡素化し、優れたパフォーマンスを発揮する
• 複雑な問題に対処できる
•機械学習システムは新データに適応可能
• 大規模データや複雑な問題が解釈できるようになる
Types of Machine Learning systems
機械学習システムには学習に種類がある。
「人間」が訓練へ介入するか否か(supervised, unsupervised, semi-supervised and reinforcement learning)
逐次的訓練か否か(online vs batch learning)
新データと既知データの単純比較か、パターン検出で予測モデル構築が必要か(instance based vs model based learning)
Supervised/Unsupervised learning
Supervised learning
* Training data is labeled.
*[Classification Task] = predicting a target label (応答値はラベル)
*Class = Target label (spam or ham)
*[Regression Task] = predicting a target numeric value (応答変数は実数)
*Specific Words(Features , Predictors, Instances, Attributes)
http://archive.ics.uci.edu/ml/datasets/Iris (UCI Machine Learning Repository)
*Important algorithms (k-Nearest Neighbors, Linear Regression, Logistic Regression, Support Vector Machines(SVMs), Decision Trees and Random Forests, Neural networks)
Unsupervised learning
* Training data is unlabeled. Need to learn without a teacher.
* Important algorithms
A. Clustering (k-means, Hierarchical cluster analysis, Expectation maximization)
B. Visualization and dimensionality reduction (Principal Component Analysis:PCA, Kernel PCA, LLE, t-SNE)
C. Anomaly detection
D. Association rule learning (Apriori , Eclat)
* Example = blog's visitor clustering
Semi-supervised learning = using partially labeled training data
* Combinations of unsupervised and supervised algorithms
* A lot of unlabeled data + a little labeled data
* Google Photo service (clustering same persons first + supervised person recognition second)
* DBNs (Deep Belief Networks : DBNs) = unsupervised (pre-training) + supervised
http://www.sersc.org/journals/IJHIT/vol9_no11_2016/34.pdf
Reinforcement learning = learn by itself what is the best strategy
*agent=learning system
*rewards
*penalties=negative rewards
*policy=best strategy
Batch and Online learning
* Batch training = offline training (incapable of learning incrementally)
* Online training = capable of learning incrementally