Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI

Human-in-the-Loop Machine Learning
brief contents
contents
foreword
preface
acknowledgments
about this book
Who should read this book
How this book is organized: A road map
About the code
liveBook discussion forum
Other online resources
about the author
Part 1: First steps
Chapter 1: Introduction to human-in-the-loop machine learning
1.1 The basic principles of human-in-the-loop machine learning
1.2 Introducing annotation
1.2.1 Simple and more complicated annotation strategies
1.2.2 Plugging the gap in data science knowledge
1.2.3 Quality human annotation: Why is it hard?
1.3 Introducing active learning: Improving the speed and reducing the cost of training data
1.3.1 Three broad active learning sampling strategies: Uncertainty, diversity, and random
1.3.2 What is a random selection of evaluation data?
1.3.3 When to use active learning
1.4 Machine learning and human–computer interaction
1.4.1 User interfaces: How do you create training data?
1.4.2 Priming: What can influence human perception?
1.4.3 The pros and cons of creating labels by evaluating machine learning predictions
1.4.4 Basic principles for designing annotation interfaces
1.5 Machine-learning-assisted humans vs. human-assisted machine learning
1.6 Transfer learning to kick-start your models
1.6.1 Transfer learning in computer vision
1.6.2 Transfer learning in NLP
1.7 What to expect in this text
Chapter 2: Getting started with human-in-the-loop machine learning
2.1 Beyond hacktive learning: Your first active learning algorithm
2.2 The architecture of your first system
2.3 Interpreting model predictions and data to support active learning
2.3.1 Confidence ranking
2.3.2 Identifying outliers
2.3.3 What to expect as you iterate
2.4 Building an interface to get human labels
2.4.1 A simple interface for labeling text
2.4.2 Managing machine learning data
2.5 Deploying your first human-in-the-loop machine learning system
2.5.1 Always get your evaluation data first
2.5.2 Every data point gets a chance
2.5.3 Select the right strategies for your data
2.5.4 Retrain the model and iterate
Part 2: Active learning
Chapter 3: Uncertainty sampling
3.1 Interpreting uncertainty in a machine learning model
3.1.1 Why look for uncertainty in your model?
3.1.2 Softmax and probability distributions
3.1.3 Interpreting the success of active learning
3.2 Algorithms for uncertainty sampling
3.2.1 Least confidence sampling
3.2.2 Margin of confidence sampling
3.2.3 Ratio sampling
3.2.4 Entropy (classification entropy)
3.2.5 A deep dive on entropy
3.3 Identifying when different types of models are confused
3.3.1 Uncertainty sampling with logistic regression and MaxEnt models
3.3.2 Uncertainty sampling with SVMs
3.3.3 Uncertainty sampling with Bayesian models
3.3.4 Uncertainty sampling with decision trees and random forests
3.4 Measuring uncertainty across multiple predictions
3.4.1 Uncertainty sampling with ensemble models
3.4.2 Query by Committee and dropouts
3.4.3 The difference between aleatoric and epistemic uncertainty
3.4.4 Multilabeled and continuous value classification
3.5 Selecting the right number of items for human review
3.5.1 Budget-constrained uncertainty sampling
3.5.2 Time-constrained uncertainty sampling
3.5.3 When do I stop if I’m not time- or budget-constrained?
3.6 Evaluating the success of active learning
3.6.1 Do I need new test data?
3.6.2 Do I need new validation data?
3.7 Uncertainty sampling cheat sheet
3.8 Further reading
3.8.1 Further reading for least confidence sampling
3.8.2 Further reading for margin of confidence sampling
3.8.3 Further reading for ratio of confidence sampling
3.8.4 Further reading for entropy-based sampling
3.8.5 Further reading for other machine learning models
3.8.6 Further reading for ensemble-based uncertainty sampling
Chapter 4: Diversity sampling
4.1 Knowing what you don’t know: Identifying gaps in your model’s knowledge
4.1.1 Example data for diversity sampling
4.1.2 Interpreting neural models for diversity sampling
4.1.3 Getting information from hidden layers in PyTorch
4.2 Model-based outlier sampling
4.2.1 Use validation data to rank activations
4.2.2 Which layers should I use to calculate model-based outliers?
4.2.3 The limitations of model-based outliers
4.3 Cluster-based sampling
4.3.1 Cluster members, centroids, and outliers
4.3.2 Any clustering algorithm in the universe
4.3.3 K-means clustering with cosine similarity
4.3.4 Reduced feature dimensions via embeddings or PCA
4.3.5 Other clustering algorithms
4.4 Representative sampling
4.4.1 Representative sampling is rarely used in isolation
4.4.2 Simple representative sampling
4.4.3 Adaptive representative sampling
4.5 Sampling for real-world diversity
4.5.1 Common problems in training data diversity
4.5.2 Stratified sampling to ensure diversity of demographics
4.5.3 Represented and representative: Which matters?
4.5.4 Per-demographic accuracy
4.5.5 Limitations of sampling for real-world diversity
4.6 Diversity sampling with different types of models
4.6.1 Model-based outliers with different types of models
4.6.2 Clustering with different types of models
4.6.3 Representative sampling with different types of models
4.6.4 Sampling for real-world diversity with different types of models
4.7 Diversity sampling cheat sheet
4.8 Further reading
4.8.1 Further reading for model-based outliers
4.8.2 Further reading for cluster-based sampling
4.8.3 Further reading for representative sampling
4.8.4 Further reading for sampling for real-world diversity
Chapter 5: Advanced active learning
5.1 Combining uncertainty sampling and diversity sampling
5.1.1 Least confidence sampling with cluster-based sampling
5.1.2 Uncertainty sampling with model-based outliers
5.1.3 Uncertainty sampling with model-based outliers and clustering
5.1.4 Representative sampling cluster-based sampling
5.1.5 Sampling from the highest-entropy cluster
5.1.6 Other combinations of active learning strategies
5.1.7 Combining active learning scores
5.1.8 Expected error reduction sampling
5.2 Active transfer learning for uncertainty sampling
5.2.1 Making your model predict its own errors
5.2.2 Implementing active transfer learning
5.2.3 Active transfer learning with more layers
5.2.4 The pros and cons of active transfer learning
5.3 Applying active transfer learning to representative sampling
5.3.1 Making your model predict what it doesn’t know
5.3.2 Active transfer learning for adaptive representative sampling
5.3.3 The pros and cons of active transfer learning for representative sampling
5.4 Active transfer learning for adaptive sampling
5.4.1 Making uncertainty sampling adaptive by predicting uncertainty
5.4.2 The pros and cons of ATLAS
5.5 Advanced active learning cheat sheets
5.6 Further reading for active transfer learning
Chapter 6: Applying active learning to different machine learning tasks
6.1 Applying active learning to object detection
6.1.1 Accuracy for object detection: Label confidence and localization
6.1.2 Uncertainty sampling for label confidence and localization in object detection
6.1.3 Diversity sampling for label confidence and localization in object detection
6.1.4 Active transfer learning for object detection
6.1.5 Setting a low object detection threshold to avoid perpetuating bias
6.1.6 Creating training data samples for representative sampling that are similar to your predictions
6.1.7 Sampling for image-level diversity in object detection
6.1.8 Considering tighter masks when using polygons
6.2 Applying active learning to semantic segmentation
6.2.1 Accuracy for semantic segmentation
6.2.2 Uncertainty sampling for semantic segmentation
6.2.3 Diversity sampling for semantic segmentation
6.2.4 Active transfer learning for semantic segmentation
6.2.5 Sampling for image-level diversity in semantic segmentation
6.3 Applying active learning to sequence labeling
6.3.1 Accuracy for sequence labeling
6.3.2 Uncertainty sampling for sequence labeling
6.3.3 Diversity sampling for sequence labeling
6.3.4 Active transfer learning for sequence labeling
6.3.5 Stratified sampling by confidence and tokens
6.3.6 Create training data samples for representative sampling that are similar to your predictions
6.3.7 Full-sequence labeling
6.3.8 Sampling for document-level diversity in sequence labeling
6.4 Applying active learning to language generation
6.4.1 Calculating accuracy for language generation systems
6.4.2 Uncertainty sampling for language generation
6.4.3 Diversity sampling for language generation
6.4.4 Active transfer learning for language generation
6.5 Applying active learning to other machine learning tasks
6.5.1 Active learning for information retrieval
6.5.2 Active learning for video
6.5.3 Active learning for speech
6.6 Choosing the right number of items for human review
6.6.1 Active labeling for fully or partially annotated data
6.6.2 Combining machine learning with annotation
6.7 Further reading
Part 3: Annotation
Chapter 7: Working with the people annotating your data
7.1 Introduction to annotation
7.1.1 Three principles of good data annotation
7.1.2 Annotating data and reviewing model predictions
7.1.3 Annotations from machine learning-assisted humans
7.2 In-house experts
7.2.1 Salary for in-house workers
7.2.2 Security for in-house workers
7.2.3 Ownership for in-house workers
7.2.4 Tip: Always run in-house annotation sessions
7.3 Outsourced workers
7.3.1 Salary for outsourced workers
7.3.2 Security for outsourced workers
7.3.3 Ownership for outsourced workers
7.3.4 Tip: Talk to your outsourced workers
7.4 Crowdsourced workers
7.4.1 Salary for crowdsourced workers
7.4.2 Security for crowdsourced workers
7.4.3 Ownership for crowdsourced workers
7.4.4 Tip: Create a path to secure work and career advancement
7.5 Other workforces
7.5.1 End users
7.5.2 Volunteers
7.5.3 People playing games
7.5.4 Model predictions as annotations
7.6 Estimating the volume of annotation needed
7.6.1 The orders-of-magnitude equation for number of annotations needed
7.6.2 Anticipate one to four weeks of annotation training and task refinement
7.6.3 Use your pilot annotations and accuracy goal to estimate cost
7.6.4 Combining types of workforces
Chapter 8: Quality control for data annotation
8.1 Comparing annotations with ground truth answers
8.1.1 Annotator agreement with ground truth data
8.1.2 Which baseline should you use for expected accuracy?
8.2 Interannotator agreement
8.2.1 Introduction to interannotator agreement
8.2.2 Benefits from calculating interannotator agreement
8.2.3 Dataset-level agreement with Krippendorff’s alpha
8.2.4 Calculating Krippendorff’s alpha beyond labeling
8.2.5 Individual annotator agreement
8.2.6 Per-label and per-demographic agreement
8.2.7 Extending accuracy with agreement for real-world diversity
8.3 Aggregating multiple annotations to create training data
8.3.1 Aggregating annotations when everyone agrees
8.3.2 The mathematical case for diverse annotators and low agreement
8.3.3 Aggregating annotations when annotators disagree
8.3.4 Annotator-reported confidences
8.3.5 Deciding which labels to trust: Annotation uncertainty
8.4 Quality control by expert review
8.4.1 Recruiting and training qualified people
8.4.2 Training people to become experts
8.4.3 Machine-learning-assisted experts
8.5 Multistep workflows and review tasks
8.6 Further reading
Chapter 9: Advanced data annotation and augmentation
9.1 Annotation quality for subjective tasks
9.1.1 Requesting annotator expectations
9.1.2 Assessing viable labels for subjective tasks
9.1.3 Trusting an annotator to understand diverse responses
9.1.4 Bayesian Truth Serum for subjective judgments
9.1.5 Embedding simple tasks in more complicated ones
9.2 Machine learning for annotation quality control
9.2.1 Calculating annotation confidence as an optimization task
9.2.2 Converging on label confidence when annotators disagree
9.2.3 Predicting whether a single annotation is correct
9.2.4 Predicting whether a single annotation is in agreement
9.2.5 Predicting whether an annotator is a bot
9.3 Model predictions as annotations
9.3.1 Trusting annotations from confident model predictions
9.3.2 Treating model predictions as a single annotator
9.3.3 Cross-validating to find mislabeled data
9.4 Embeddings and contextual representations
9.4.1 Transfer learning from an existing model
9.4.2 Representations from adjacent easy-to-annotate tasks
9.4.3 Self-supervision: Using inherent labels in the data
9.5 Search-based and rule-based systems
9.5.1 Data filtering with rules
9.5.2 Training data search
9.5.3 Masked feature filtering
9.6 Light supervision on unsupervised models
9.6.1 Adapting an unsupervised model to a supervised model
9.6.2 Human-guided exploratory data analysis
9.7 Synthetic data, data creation, and data augmentation
9.7.1 Synthetic data
9.7.2 Data creation
9.7.3 Data augmentation
9.8 Incorporating annotation information into machine learning models
9.8.1 Filtering or weighting items by confidence in their labels
9.8.2 Including the annotator identity in inputs
9.8.3 Incorporating uncertainty into the loss function
9.9 Further reading for advanced annotation
9.9.1 Further reading for subjective data
9.9.2 Further reading for machine learning for annotation quality control
9.9.3 Further reading for embeddings/contextual representations
9.9.4 Further reading for rule-based systems
9.9.5 Further reading for incorporating uncertainty in annotations into the downstream models
Chapter 10: Annotation quality for different machine learning tasks
10.1 Annotation quality for continuous tasks
10.1.1 Ground truth for continuous tasks
10.1.2 Agreement for continuous tasks
10.1.3 Subjectivity in continuous tasks
10.1.4 Aggregating continuous judgments to create training data
10.1.5 Machine learning for aggregating continuous tasks to create training data
10.2 Annotation quality for object detection
10.2.1 Ground truth for object detection
10.2.2 Agreement for object detection
10.2.3 Dimensionality and accuracy in object detection
10.2.4 Subjectivity for object detection
10.2.5 Aggregating object annotations to create training data
10.2.6 Machine learning for object annotations
10.3 Annotation quality for semantic segmentation
10.3.1 Ground truth for semantic segmentation annotation
10.3.2 Agreement for semantic segmentation
10.3.3 Subjectivity for semantic segmentation annotations
10.3.4 Aggregating semantic segmentation to create training data
10.3.5 Machine learning for aggregating semantic segmentation tasks to create training data
10.4 Annotation quality for sequence labeling
10.4.1 Ground truth for sequence labeling
10.4.2 Ground truth for sequence labeling in truly continuous data
10.4.3 Agreement for sequence labeling
10.4.4 Machine learning and transfer learning for sequence labeling
10.4.5 Rule-based, search-based, and synthetic data for sequence labeling
10.5 Annotation quality for language generation
10.5.1 Ground truth for language generation
10.5.2 Agreement and aggregation for language generation
10.5.3 Machine learning and transfer learning for language generation
10.5.4 Synthetic data for language generation
10.6 Annotation quality for other machine learning tasks
10.6.1 Annotation for information retrieval
10.6.2 Annotation for multifield tasks
10.6.3 Annotation for video
10.6.4 Annotation for audio data
10.7 Further reading for annotation quality for different machine learning tasks
10.7.1 Further reading for computer vision
10.7.2 Further reading for annotation for natural language processing
10.7.3 Further reading for annotation for information retrieval
Part 4: Human–computer interaction for machine learning
Chapter 11: Interfaces for data annotation
11.1 Basic principles of human–computer interaction
11.1.1 Introducing affordance, feedback, and agency
11.1.2 Designing interfaces for annotation
11.1.3 Minimizing eye movement and scrolling
11.1.4 Keyboard shortcuts and input devices
11.2 Breaking the rules effectively
11.2.1 Scrolling for batch annotation
11.2.2 Foot pedals
11.2.3 Audio inputs
11.3 Priming in annotation interfaces
11.3.1 Repetition priming
11.3.2 Where priming hurts
11.3.3 Where priming helps
11.4 Combining human and machine intelligence
11.4.1 Annotator feedback
11.4.2 Maximizing objectivity by asking what other people would annotate
11.4.3 Recasting continuous problems as ranking problems
11.5 Smart interfaces for maximizing human intelligence
11.5.1 Smart interfaces for semantic segmentation
11.5.2 Smart interfaces for object detection
11.5.3 Smart interfaces for language generation
11.5.4 Smart interfaces for sequence labeling
11.6 Machine learning to assist human processes
11.6.1 The perception of increased efficiency
11.6.2 Active learning for increased efficiency
11.6.3 Errors can be better than absence to maximize completeness
11.6.4 Keep annotation interfaces separate from daily work interfaces
11.7 Further reading
Chapter 12: Human-in-the-loop machine learning products
12.1 Defining products for human-in-the-loop machine learning applications
12.1.1 Start with the problem you are solving
12.1.2 Design systems to solve the problem
12.1.3 Connecting Python and HTML
12.2 Example 1: Exploratory data analysis for news headlines
12.2.1 Assumptions
12.2.2 Design and implementation
12.2.3 Potential extensions
12.3 Example 2: Collecting data about food safety events
12.3.1 Assumptions
12.3.2 Design and implementation
12.3.3 Potential extensions
12.4 Example 3: Identifying bicycles in images
12.4.1 Assumptions
12.4.2 Design and implementation
12.4.3 Potential extensions
12.5 Further reading for building human-in-the-loop machine learning products
appendix: Machine learning refresher
A.1 Interpreting predictions from a model
A.1.1 Probability distributions
A.2 Softmax deep dive
A.2.1 Converting the model output to confidences with softmax
A.2.2 The choice of base/temperature for softmax
A.2.3 The result from dividing exponentials
A.3 Measuring human-in-the-loop machine learning systems
A.3.1 Precision, recall, and F-score
A.3.2 Micro and macro precision, recall, and F-score
A.3.3 Taking random chance into account: Chance-adjusted accuracy
A.3.4 Taking confidence into account: Area under the ROC curve (AUC)
A.3.5 Number of model errors spotted
A.3.6 Human labor cost saved
A.3.7 Other methods for calculating accuracy in this book
index
A
B
C
D
E
F
G
H
I
K
L
M
N
O
P
Q
R
S
T
U
V