Artificial Intelligence

To keep things easy this page will act as an index for all AI related resources.

External Resources

• Machine Learning Tutorials
• Machine Learning Getting Started Ideas
• Machine Learning Resources
• Top 100 Data Science Resources
• Learn Math Fast

News:

• Inside AI

Contests:

• Kaggle Competitions

Articles

• Deep Learning Performance with AutoAugment
  • Augmenting computer vision datasets by generating new images based on existing data in the set
• Deep Reinforcement Learning Doesn’t Work Yet - February 14th, 2018
  • Examples of why RL is hard to put into practice, the state of the industry, DeepRL
• The Evolution and Core Concepts of Deep Learning & Neural Networks - August 2016
  • Good primer on NN
• AI at Google: our principles
  • Google’s principles for AI development
• Understanding the backward pass through Batch Normalization Layer
  • Good explanation of how the derivative of the Batch Normalization layer is computed
• A Beginner’s Guide to Recurrent Networks and LSTMs
  • Good explanation of LSTMs and GRUs, plus an intuition of backpropagation through time
• Optimization techniques comparison in Julia: SGD, Momentum, Adagrad, Adadelta, Adam
  • Short explanations and performance comparison for SGD, SGD with momentum, AdaGrad, AdaDelta and Adam optimizers
• googlenet in keras
  • An implementation of GoogLeNet in Keras
• Learning Rate Schedules and Adaptive Learning Rate Methods for Deep Learning
  • Compares different methods of adjusting the learning rate of a NN dynamically
• Building powerful image classification models using very little data
  • Data augmentation in Keras; use VGG16 as a base for our image classifier
• Profiling Top Kagglers: Martin Henze (AKA Heads or Tails), World’s First Kernels Grandmaster
  • The journey to becoming a Kaggle Kernel grandmaster, with lots of tips
• Machine learning mega-benchmark: GPU providers (part 2)
  • Comparison of Cloud GPU compute providers
• Learning from humans: what is inverse reinforcement learning?
  • Inverse reinforcement learning: given a policy or a set of data about an expert performing a task, try to extract a reward function
  • Apprenticeship learning: given an expert policy, use it as a baseline from which to improve
• Retro Contest: Results
  • Short presentation of the winners of a reinforcement learning contest (they attempted to generalize from previous knowledge)
• CNNs from different viewpoints
  • Short post describing how to view CNNs as matrix multiplications or dense neural nets; informative
• 1(https://eng.uber.com/deconstructing-lottery-tickets/)
  • Using a binary mask to zero out many of the model’s weights leads to better accuracy
• 2(https://eng.uber.com/coordconv/)
  • Introducing coordinates to convolution filters to give CNNs a way to model the coordinate transform task (increased performance in RL, Object detection)

Tutorials and Guides

Tutorials:

• A Guide to TF Layers: Building a Convolutional Neural Network
• Introduction to CNN Keras - Acc 0.997 (top 8%)
  • A guide to a small but powerful CNN with a 99.671% accuracy on MNIST; uses data augmentation techniques
• A Guide to TF Layers: Building a Convolutional Neural Network
  • Tutorial about building CNNs in Tensorflow

Guides

• Tutorial: Optimizing Neural Networks using Keras (with Image recognition case study
  • Widen your net, deepen your net, add dropout layers, change activation function, change network arhitecture
• A guide to convolution arithmetic for deep learning
• Which GPU(s) to Get for Deep Learning
  • Basically aim for a GTX 1080/1080Ti if you can afford it, if not go for a GTX 1070
• TF Estimators
  • Amongst other things you can convert Keras models to TF estimators
• Understanding and Coding Inception Module in Keras

Papers

• Generative Temporal Models with Spatial Memory for Partially Observed Environments
  • Videos of the agent in action
  • Generative Temporal Model with Spatial Memory, an action-conditioned generative model that uses a scalable non-parametric memory to store spatial and visual information; DND memory
  • The agent learns an environment then can predict how the environment will look like after a number of time steps (hundreds) and moves
• Synthesizing Programs for Images using Reinforced Adversarial Learning
  • Video of demonstration
  • Using GANs to train a RL one a cluster of machines; the algorithm learns to understand the creation of and recreate images it sees (drawing the digits of the MNIST dataset, characters from the POLIGLOT dataset) and even reconstruct 3D scenes
• Backdrop: Stochastic Backpropagation
  • Backdrop, a flexible and simple-to-implement method, intuitively described as dropout acting only along the backpropagation pipeline
• Efficient Estimation of Word Representations in Vector Space
  • Study of the quality of vector representations of words derived by various models on a collection of syntactic and semantic language tasks
• Rethinking the Inception Architecture for Computer Vision
  • Ideas for reducing computational complexity of CNNs, different Inception architectures
• Curiosity-driven Exploration bySelf-supervised Prediction
  • RL with intrinsic curiosity reward manages to learn how to play Mario without extrinsic rewards, is able to generalize knowledge well; uses an Intrinsic Curiosity Module
• L2 Regularization versus Batch and Weight Normalization
  • These neural networks use L2 regularization, also called weight decay, ostensibly to prevent overfitting. However, we show that L2 regularization has no regularizing effect when combined with normalization
• Understanding the difficulty of training deep feedforward neural networks
• Deep Residual Learning for Image Recognition
  • ResNet architecture
• Wide Residual Networks
  • Wide ResNet architecture
• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• How Does Batch Normalization Help Optimization?
  • An investigation into why Batch Normalization works. It is also compared to L losses. Apparently it works because it smoothens the optimization landscape; it doesn’t have a large impact on ICS.
• Verification Of Non Linear Specifications For Neural Networks
  • Mathematical tools for proving the robustness of a model against adversarial attacks
• Going deeper with convolutions
• ImageNet Classification with Deep Convolutional Neural Networks
• Practical Recommendations for Gradient-Based Training of Deep Architectures
• CoupleNet: Paying Attention to Couples with Coupled Attention for Relationship Recommendation
• DARTS: Differentiable Architecture Search
• Learning Cognitive Models using Neural Networks
• Going deeper with convolutions

Courses and Books

Courses:

• Coursera Machine Learning - taught by AndrewNG
  • ML course, presents the mathematics and concepts behind supervised and unsupervised learning algorithms, and has a good introduction to neural networks
  • Great primer into machine learning
• Udacity Deep Learning
  • Tensorflow Deep Learning course
  • Good course for introducing a number of advanced concepts, but I feel it doesn’t spend enough time on some of them
  • Has a lot of good assignments in Tensorflow
• Udemy Zero to Deep Learning
  • Keras Deep Learning course
  • Good course with many interesting assigments; provides an intro to data preprocessing and augmentation as well

Books:

• Intelligence: A Modern Approach
• Deep Learning with Python

Tools

• Google Colaboratory - go there now
• ConvNet Drawer
• Keras Visualization Toolkit
• Tensorboard

Existing work

• LeNet5

Internal Resources

Learning ML

• Loss and Accuracy
• Debugging ML algorithms
  • Bias vs Variance
  • Precision, Recall and F1 score
  • Cross Entropy
  • Learning Curve
  • Learning rate and number of epochs

• Reinforcement Learning

• Neural Networks
  • Basics
    • Simple Neural Network
    • Data Preprocessing
    • Weight Initialization
    • One hot encoding
    • Softmax
  • Activation Functions
    • Step
    • Sigmoid
    • Tanh
    • ReLU
    • Softplus
    • Softmax
    • Optimization
      • (Batch) Gradient Descent
      • Stochastic Gradient Descent
      • Convergence
      • Online Learning
      • Map Reduce and Parallelism
      • SGD with Momentum
      • AdaGrad
      • AdaDelta
      • RMSProp
      • Adam
      • Choosing the Best Optimizer
  • Regularization
    • Batch Normalization
    • Local Response Normalization
  • Embeddings
    • Skip Gram
    • CBOW
  • Architectures
• Types of Neural Networks
  • Recurrent Neural Network
    • LSTM
    • GRU
  • Convolutional Neural Networks
    • CNN Visualizations
      • Activation Maximization
      • Saliency Maps
      • Class Activation Maps
• Types of Machine Learning Problems
  • Supervised Learning
    • Regression problems
    • Classification problems
    • Support Vector Machines
    • Natural Language Processing
  • Unsupervised Learning
    • Anomaly Detection
    • Clustering
  • Reinforcement Learning
• Machine Learning Advice
  • Dataset Management
  • Pipeline
  • Ceiling Analysis
  • Increasing Your Dataset Size

Tools:

• Tensorboard

Projects

• Iris classifier
• MNIST classifier

Discussion:

• Artificial Intelligence Explained

• Tensorflow vs Keras

Guides:

• Installing Tensorflow on ArchLinux
• Installing Keras on ArchLinux
• Installing Jupyter notebook with a virtualenv kernel