Artificial Intelligence 101 1

How AI Works
Table of Contents

 

 

1.0 Introduction
1.1 The first AI hype cycle
1.1.1 Early successes
1.1.2 Early failures
1.2 The first AI winter
1.3 The second AI hype cycle
1.4 The second AI winter
1.5 The third AI hype cycle
1.6 No third winter in sight
1.7 What is AI?

 

2.0 Supervised Learning
2.1 Models
2.2 Assumptions
2.3 Origin of the term “supervised learning”
2.4 Model evaluation
2.4.1 Splitting the data into training and test sets
2.4.2 Adding a validation dataset
2.4.3 Cross-validation
2.4.4 Area under the curve
2.5 Underfitting and overfitting
2.6 Variance and bias
2.7 Flexibility vs. interpretability
2.8 Types of supervised learning
2.9 Classification algorithms
2.9.1 Single vs multi-category classification
2.9.2 Linear vs non-linear classification
2.9.3 Classification algorithms
2.9.4 Classification metrics
2.9.5 Input variables for vision tasks
2.9.6 Input variables for natural language processing tasks
2.9.6.1 Individual words
2.9.6.2 Word embeddings
2.9.6.3 Document embeddings
2.10 Regression algorithms
2.10.1 Polynomial regression
2.10.2 k-nearest neighbors
2.10.3 General additive algorithms
2.10.4 Decision trees
2.10.5 Neural networks
2.10.6 Regression algorithms in scikit-learn
2.11 Hyperparameters
2.12 Ensemble algorithms
2.12.1 Voting ensemble models
2.12.2 Bagging
2.12.3 Boosting
2.12.4 Random forest
2.12.5 Stacked ensemble algorithms
2.13 Generative algorithms
2.13.1 An example: building a spam filter
2.13.2 Naïve Bayes algorithm
2.13.3 Probabilistic programming
2.13.4 Mixture models
2.13.5 Energy-based models
2.14 No free lunch theorem
2.15 Regularization
2.16 Dimensionality reduction
2.16.1 Variable selection
2.17 Semi-supervised learning

 

3.0 Deep Learning
3.1 Why do we need deep learning neural networks?
3.2 Examples of deep learning networks
3.2.1 Example: Deep learning network for the housing dataset
3.2.2 Handwritten digit classification
3.2.3 Why deep learning is revolutionary
3.3 Computing the optimal weights
3.3.1 Computing the function without a deep learning network
3.3.2 Gradient descent
3.3.3 Initial weights
3.3.4 Learning rate
3.3.5 Early stopping
3.3.6 Layers
3.3.7 Forward propagation
3.3.8 Backpropagation
3.3.9 Vanishing and exploding gradients
3.3.9.1 Vanishing gradients
3.3.9.2 Exploding gradients
3.3.10 Batch vs stochastic gradient descent
3.3.11 Momentum methods
3.3.12 Learning rate optimization
3.3.13 Activation functions
3.3.13.1 Hidden layer activation functions
3.3.13.2 Output neurons
3.3.14 Reducing overfitting
3.3.15 Performance and model complexity
3.3.16 Performance and dataset size
3.3.17 Performance, computing power, and training time
3.3.18 Memorization
3.3.19 Overfitting
3.4 Convolutional neural networks
3.4.1 Input layer
3.4.2 Convolution layer
3.4.3 Pooling layer
3.4.4 Feature learning
3.4.5 CNN applications
3.4.6 CNN improvements
3.5 Recurrent neural networks
3.5.1 Long short term memory networks
3.5.2 Gated recurrent units
3.5.3 Bidirectional RNNs and LSTMs
3.5.4 Issues with long-term dependencies
3.6 Sequence-to-sequence learning
3.7 Transformer networks

 

4.0 Transfer Learning
4.1 Feature extraction
4.2 Word embeddings
4.3 Fine-tuning language models
4.3.1 Language models
4.3.2 Fine-tuning
4.3.3 Parameter-efficient fine-tuning
4.4 Cross-lingual transfer
4.5 Multi-task learning
4.5.1 Multiple NLP tasks
4.5.2 Multiple speech tasks
4.5.3 Multiple reinforcement learning and robotics tasks
4.5.4 Multiple computer vision tasks
4.6.5 Multi-modal multi-task learning
4.6 Meta-learning
4.7 Continual learning

 

5.0 Foundation models
5.1 Large language models
5.2 Emergent properties
5.3 OpenAI competitors
5.4 Instruction tuning
5.5 Measuring the performance of LLMs
5.6 Linguistic capabilities
5.6.1 Language modeling
5.6.2 Language translation
5.6.3 Multi-lingual question answering
5.6.4 Grammatical rules
5.6.5 Understanding figurative language
5.7 Prompts
5.7.1 Prompting techniques
5.7.1.1 Chain of thought prompting
5.7.1.2 Self-consistency
5.7.1.3 Least to most prompting
5.7.1.4 Learned prompts
5.7.1.5 Ask Me Anything
5.7.1.7 Refine based on external research
5.7.1.8 As-needed decomposition and planning
5.7.1.9 Optimization by prompting
5.7.1.10 System 2 attention
5.7.1.11 Limitations of iterative approaches
5.7.2 In-context learning
5.7.3 Prompt issues
5.8 Retrieval augmented generation
5.9 Democratization of LLMs
5.9.1 Smaller models
5.9.2 More efficient attention mechanisms
5.9.3 RLHF alternatives
5.10 Vision and other modalities
5.10.1 Image generation
5.10.1.1 Other commercial text-to-image generators
5.10.1.2 Image generation performance
5.10.1.3 Generating scenes
5.10.1.3 Generating 3D images and videos from text
5.10.1.4 Under the hood
5.10.1.5 Limitations of image generation models
5.10.2 Image analysis
5.10.3 Video analysis
5.10.4 Music generation
5.10.5 Speech generation
5.10.6 Multiple modality foundation models
5.10.7 Biology as a modality

 

6.0 Thinking and reasoning
6.1 World knowledge
6.1.1 Simple facts
6.1.2 Academic subjects
6.1.3 Legal expertise
6.1.4 Medical knowledge
6.1.5 Professional knowledge
6.1.6 Conceptual knowledge
6.1.7 Events
6.1.8 Scripts
6.1.9 Models of the world
6.1.10 Conclusions about LLM knowledge
6.2 Reasoning
6.2.1 Commonsense reasoning
6.2.2 Abstract reasoning
6.2.2.1 Reasoning by analogy
6.2.2.2 Logical reasoning
6.2.2.3 Mathematical reasoning
6.2.2.4 Coding
6.2.2.5 Humor
6.2.2.6 Theory of mind
6.2.3 Limited to linear pattern-matching
6.2.4 Moving the goalposts
6.3 Stochastic parrots or AGI?

 

7.0 Natural Language Processing
7.1 Natural language processing challenges
7.1.1 Literal vs. implied meaning
7.1.2 Commonsense world knowledge
7.1.3 Word ambiguity in natural language processing
7.1.4 Coreference resolution
7.1.5 Other linguistic natural language processing phenomena
7.2 Symbolic NLP
7.2.1 Representational challenges
7.2.1.1 Lexical vs. canonical representation
7.2.1.2 Breadth of the meaning representation
7.2.1.3 Breadth of topics and genres
7.2.2 Symbolic representation schemes
7.2.2.1 Syntactic processing
7.2.2.2 Shallow meaning representations
7.2.2.3 Semantic dependency representations
7.2.2.4 Abstract Meaning Representation
7.2.2.5 Alexa Meaning Representation Language
7.2.2.6 Conceptual dependency theory
7.2.3 Symbolic NLP research tools
7.2.3.1 Corpora
7.2.3.2 Knowledge representations
7.2.3.3 Lexicons for natural language processing
7.2.3.4 Specialized linguistic resources
7.2.4 NLP pipelines
7.2.4.1 Tokenization
7.2.4.2 Regular expressions
7.2.4.3 Case conversion
7.2.4.4 Spelling correction for natural language processing
7.2.4.5 Stemming
7.2.4.6 Lemmatization
7.2.4.7 Lexicons
7.3 Statistical NLP
7.3.1 Statistical NLP corpuses
7.3.2 Statistical measures and techniques
7.3.2.1 Word frequency
7.3.2.2 Collocations
7.3.2.3 Co-occurrence
7.3.2.4 TF-IDF
7.3.2.5 Lesk algorithm
7.3.2.6 Latent semantic analysis
7.3.2.7 Latent semantic indexing
7.3.2.8 Latent dirichlet allocation

 

8.0 Information Extraction
8.1 Types of IE
8.1.1 Entity extraction
8.1.2 Entity linking
8.1.3 Template extraction
8.1.3.1 Temporal extraction
8.1.3.2 Numeric extraction
8.1.4 Sentiment analysis
8.2 IE methods
8.2.2 Supervised learning
8.2.2.1 Semi-supervised learning approaches
8.2.2.2 Distant supervised learning
8.2.3 Large language models
8.2.4 Off-the-shelf tools

 

9.0 Machine Translation
9.1 Relationship of machine translation to automatic speech recognition
9.2 History of machine translation systems
9.2.1 Machine translation competitions
9.2.2 Rule-based machine translation
9.2.2.1 Direct dictionary systems
9.2.2.2 Interlingua systems
9.2.2.3 Transfer approach
9.2.3 Statistical machine translation
9.2.3.1 Parallel texts
9.2.3.2 Sentence alignment
9.2.3.3 Word alignment
9.2.3.4 How statistical machine translation works
9.2.3.5 Word-based statistical machine translation
9.2.3.6 Phrase-based machine translation
9.2.3.7 Statistical machine translation toolkits
9.2.4 Neural machine translation
9.2.5 Comparing neural machine translation performance with phrase-based machine translation
9.2.6 Google neural machine translation
9.2.7 Ensemble models
9.2.8 Out of vocabulary words
9.2.8.1 Copying out-of-vocabulary words
9.2.8.2 Dictionary back-off
9.2.8.3 Subword approaches
9.2.9 Initialization of machine translation networks with language models
9.2.10 Challenges for neural machine translation systems
9.3 Machine translation for low-resource languages
9.3.1 Characteristics of low-resource languages
9.3.2 Multilingual machine translation
9.3.3 Machine translation using pivot languages
10.0 Speech Recognition
10.1 The speech signal
10.2 Automatic speech recognition challenges
10.3 Speech processing tasks
10.4 Speech recognition datasets
10.5 Speech recognition prior to ~2010
10.5.1 Consonant- and vowel-based methods
10.5.2 Word-based methods
10.5.3 Phonemes
10.5.4 Stages of speech recognition systems
10.5.4.1 Preprocessing of the acoustic signal
10.5.4.2 Feature extraction
10.5.4.3 Decoding
10.6 Speech recognition using deep learning
10.6.1 Hybrid systems
10.6.2 End-to-end systems
10.6.3 Self-supervised learning
10.6.3.1 Contrastive predictive coding
10.6.3.2 Autoregressive predictive coding
10.6.3.3 Masked predictive encoding
10.6.3.4 Self-training
10.6.3.5 Conformer models
10.7 Multilingual models
10.8 Other speech tasks
10.8.1 Speech-to-speech translation
10.8.2 Language identification
10.9 Performance of machines compared to humans

 

11.0 Computer Vision
11.1 Handwritten digit classification
11.2 Image classification
11.2.1 Supervised learning using feature extraction methods
11.2.2 Deep learning
11.2.3 Self-supervised learning
11.2.4 Semi-supervised learning
11.2.5 Weakly supervised learning
11.3 Object detection
11.3.1 Datasets
11.3.2 Supervised learning approaches
11.3.3 Unsupervised object detection
11.3.4 Compositional object detection
11.4 Facial recognition
11.5 What is being learned by computer vision systems?

 

12.0 Reinforcement Learning
12.1 Optimal control theory
12.2 Reinforcement learning
12.3 Reward functions
12.4 Model-based reinforcement learning
12.5 Model-free reinforcement learning
12.5.1 Q-Learning
12.5.2 Policy-based approximation
12.5.3 Game playing
12.5.4 Advantages of model-free methods
12.6 Environments with sparse rewards
12.7 Supervised learning vs. reinforcement learning
12.8 Inverse reinforcement learning
12.9 Offline reinforcement learning
12.10 Unsupervised reinforcement learning
12.11 Data augmentation

 

13.0 Robotics
13.1 Optimal control theory
13.2 Reinforcement learning for robotics
13.3 Imitation learning
13.3.1 Kinesthetic feedback and teleoperation
13.3.2 Behavioral cloning
13.3.3 Inverse reinforcement learning
13.3.4 Adversarial imitation learning
13.4 Use of simulators
13.4.1 Simulation environments
13.4.2 Examples of training in simulators
13.4.4 The reality gap
13.5 Bridging the reality gap
13.5.1 Domain adaptation
13.5.2 System identification
13.5.3 Domain randomization
13.6 Generalization
13.6.1 Adaptation to different robot morphologies
13.6.2 Using old data to learn new tasks
13.6.3 Generalization to new environments
13.7 Learning from video demonstration
13.8 Following natural language instructions

 

14.0 Unsupervised Learning
14.1 Cluster analysis
14.1.1 A cluster analysis example
14.1.2 How cluster analysis works
14.1.3 Anomaly detection
14.2 Dimensionality reduction
14.3 Time series analysis
14.4 Recommender systems
14.4.1 Collaborative filtering
14.4.1.1 User-user collaborative filtering
14.4.1.2 Item-Item Collaborative Filtering
14.4.2 Content-based techniques
14.4.3 Neural network techniques
14.5 Autoencoders
14.5.1 How autoencoders work
14.5.2 Creating deepfakes using autoencoders
14.6 Generative models
14.6.1 Variational autoencoders
14.6.2 Generative adversarial networks
14.6.3 Diffusion models

     

Introduction>

     

© aiperspectives.com, 2020. Unauthorized use and/or duplication of this material without express and written permission from this site’s owner is strictly prohibited. Excerpts and links may be used, provided that full and clear credit is given with appropriate and specific direction to the original content.