Debiasing AI Using Amazon SageMaker

Introduction

JDAI AI

1. Crime-Fighting Case Study

  • Predictive ploicing
    • Hotspot Map
      • Crime is likely to occur
      • Officers patrol
      • Reduces crime
    • This Course
      • Crime is a great real-world use case
      • Excellent case study of bias in Machine Learning
      • Take on the bias of their creators
      • Address bias and ways to prevent it
  • Overview of crime-fighting case study
    • Minority
      • Precrime concept
      • Arresting criminals before crime
      • Foreknowledge using psychic technology
    • The Model
      • All knowing brain
      • Mathenmatical model
    • Protected Attributes
      • Identifying attributes about a person
      • Race, gender, disablity, family status
  • Architecture diagram
    • S3
    • Lambda
    • SageMaker

2. Building the model with SageMacker

  • Use Amazon SageMaker Built-in Algorithms
    • BlazingText(NLP, word2Vec)
      • 日本語も利用可能
      • MeCabで文字を区切りする
    • Linear Learner Algorithm
      • linear model
        • Locistic regression
          • Claasification porblems
        • Linear regression
          • Numeric values
    • K-Means Algorithm
    • Binary_classifier hyperparameter
      • yes
      • no
  • Deploy Model Deploy Model

  • Machine Learning process MLProcess
    1. Obtian Data
    2. Prepare Data
      • Outsouring it to Amazon Mechanical Turk
    3. Train Model LearningAlgorithm
      • Learing Algorithm
      • Loss Fucntion
      • SGD (確率的勾配降下法)
      • Optimization technique
          model.compile(loss='mean_squared_error', optimizer='sgd')
        
      • Taring is also an iteration process. Training_iteration_process
    4. Evaluate Model
      • Epoch
        1. valication_score
        2. quality_metric
        3. Epoch: Loss improved. Updating best model Traing_Validation_Test
    5. Deploy Model
    6. Integrate it with APP
  • Inspect and visualize data
  • Prepare the data
  • Training the model
  • Deploy the model

3. Deploying and Testing the Model via DeepLens

  • What is DeepLens
  • Deploy model to AWS DeepLens
    • we can use the same way to depoly our model to iOS, Mac using CoreML
  • Extend AWS DeepLens
  • Retrieve attibutes via AWS Rekognition
  • Invoke the crime model
  • Set up model alerts
  1. Explaining the Model
    • What is explainable AI
    • Trust and transparency issues
    • Making algorithms explainable

Conclusion