ML Overview

Roles#

• Data Analysis

  • Data Scientist

• Model Development

  • AI Researcher
  • AI Specialist
  • AI Engineer

• Service (Production)

  • AI Engineer
  • MLOps

---

AI Product Development Cycle#

1. Clearly define the problem and goals
   • target audience, key stakeholders, data and resources
2. Collect and analyze data
   • Acquire high-quality and relevant data.
   • Perform exploratory data analysis(EDA) to understand the distribution, patterns, and anomalies in the data.
   • Determine the data split between training, validation, and testing.
3. Prepare the dataset
   • clean, preprocess, transform, missing or imcomplete, imbalance
   • Ensure that the data is properly formatted and normalized.
4. Choose and train an appropriate model
   • Consider the trade-off between model complexity and interpretability.
   • Ensure that the model is scalable and can handle large datasets.
   • loss function, optimization algorithm
   • Monitor and adjust the model
5. Evaluate the model’s performance with appropriate metrics
6. Refine and optimize the model
   • Compare the model’s performance on the test data with its performance on the validation data.
7. Deploy the model
   • Make sure that it is integrated with the other systems and processes as needed.
8. Monitor and maintain the model’s performance.
   • Monitor the model’s performance in real-world conditions, and adjust as necessary.
   • Continuously evaluate the model’s performance and make improvements as needed.

Main Challenges#

Data Preprocessing#

• Insufficient Quantity of Data
• Nonrepresentative Data
• Poor-Quality Data
• Imbalanced data
  • Oversampling
  • Undersampling
  • Generating synthetic
• Irrelevant Features

Training#

• Overfitting
  • Dropout
  • Monte Carlo (MC) Dropout
  • Regularization
• Underfitting
• The Vanishing/Exploding Gradients Problems
  • Glorot and He Initialization
  • Better Activation Functions
  • Batch Normalization
  • Gradient Clipping
• Hyperparameter tuning
  • Grid search
  • Random search
  • Bayesian optimization

Evaluation#

• Explainable AI (XAI)
• Data Mismatch in Testing and Validating

---

Math#

• Softmax

Linear Algebra#

Vector, Matrix(Tensor)

Probability#

Statistics#

---

Framework#

• Dataset and Dataloader

• Optimizer

• Multi-GPU

• Monitoring

  • Weights & Biases
  • Tensorboard

• Neural Network

  • PyTorch; PyG
  • JAX; Flax, Jraph
  • TensorFlow;

• Accelerator

  • TPU: XLA, ?
  • GPU: CUDA, Triton

---

Base of ML#

$\Theta^* \larr \Theta - \eta\nabla_{\Theta}{\cal L}$

• Gradient Descent

• Constraints?

Activation Functions

Weight Initializers

Metric

• Confusion Matrices
  • Decision Boundaries
• Precision and Recall, and it’s Trade-off
• ROC Curve
• F1

Loss (Cost)

• MSE, MAE
• Cross Entropy
  • KL divergence
  • Focal

Backpropagation

• Autodiff (AutoGrad)

Optimizers

• Momentum, Nesterov Accelerated Gradient, AdaGrad, RMSProp, Adam, AdaMax, Nadam, AdamW

Learning Rate Scheduler

Computational Complexity

Regularization

• The Normal Equation

• Regularized Models

  • Ridge Regression
  • Lasso Regression
  • Elastic Net Regression
  • Early Stopping

• $\ell_1$ and $\ell_2$ Regularization

• Max-Norm Regularization

• Estimating Probabilities ?

Architectures#

• 1-Stage Detector

• 2-Stage Detector

• AutoML

Layers#

• Linear Layers

• Convolutional Layers

  • CNN

• Recurrent Neurons

  • LSTM, GRU

• Attention Mechanisms

  • Transformer
    • Attention Is All You Need: The Original Transformer Architecture
  • Vision Transformers

• Pooling Layers

  • Avg, Max

• Normalization Layers

• Dropout

• Neural architecture search (NAS)

• Optimization

  • Tensor Decomposition
  • Quantization
  • Model Compression
    • Knowledge distillation

Technic#

• Using GPUs to Speed Up Computations

  • Getting Your Own GPU
  • Managing the GPU RAM
  • Placing Operations and Variables on Devices
  • Parallel Execution Across Multiple Devices

• Training Models Across Multiple Devices

  • Model Parallelism
  • Data Parallelism

Tasks#

• Supervise, Un-supervise, Semi-supervise

• Instance-Based vs. Model-Based Learning

• Classification

• Regression

• Annotation

• Computer Vision (CV)

  • Classification and Localization
  • Object Detection
  • Object Tracking
  • Semantic Segmentation
  • Optical Character Recognition(OCR)

• Natural Language Processing (NLP)

  • Bag of Words & Word Embedding
  • Forecasting a Time Series
  • Handling Long Sequences
    -Fighting the Unstable Gradients Problem
    -Tackling the Short-Term Memory Problem
  • Sentiment Analysis
    • Masking
    • Reusing Pretrained Embeddings and Language Models
  • An Encoder-Decoder Network for Neural Machine Translation
    • Bidirectional RNNs
    • Beam Search
  • KLUE
  • MRC
  • Summarize
  • Generative
    • GPT

• RecSys

• Multi-modal Learning

Models#

• X-AI

Support Vector Machines#

• Linear SVM Classification
  • Soft Margin Classification
• Nonlinear SVM Classification
  • Polynomial Kernel
  • Similarity Features
  • Gaussian RBF Kernel
  • SVM Classes and Computational Complexity
• SVM Regression
• Under the Hood of Linear SVM Classifiers
• The Dual Problem
  • Kernelized SVMs

Decision Trees#

• The CART Training Algorithm
• Gini Impurity or Entropy?
• Regularization Hyperparameters
• Sensitivity to Axis Orientation
• Decision Trees Have a High Variance

Ensemble Learning and Random Forests#

• Voting Classifiers
• Bagging and Pasting
  • Out-of-Bag Evaluation
• Random Forests
  • Feature Importance
• Boosting
• Stacking

Dimensionality Reduction#

• Projection
• Manifold Learning
• PCA
• LLE

Clustering#

• k-means and DBSCAN
• Gaussian Mixtures
  • Using Gaussian Mixtures for Anomaly Detection

Autoencoder#

• Efficient Data Representations
• Performing PCA with an Undercomplete Linear Autoencoder
• Stacked Autoencoders
• Convolutional Autoencoders
• Denoising Autoencoders
• Sparse Autoencoders
• Variational Autoencoders

Generative Adversarial Networks#

• The Difficulties of Training GANs
• Deep Convolutional GANs
• Progressive Growing of GANs
• StyleGANs

Reinforcement Learning#

• Rewards
• Policy Search
• Neural Network Policies
• Evaluating Actions: The Credit Assignment Problem
• Policy Gradients
• Markov Decision Processes
• Temporal Difference Learning
• Q-Learning
  • Exploration Policies
  • Approximate Q-Learning and Deep Q-Learning
• Implementing Deep Q-Learning
• Deep Q-Learning Variants
  • Fixed Q-value Targets
  • Double DQN
  • Prioritized Experience Replay
  • Dueling DQN

Diffusion Models#

Service#

Cloud Service#

• Microsoft Azure

• Google Cloud services: Vertex AI

• AWS

• Cloudflare

Drug Discovery Service#

• Isomorphic Labs, DeepMind

• Insilico medicien

---

Reference:

1. homl, 2nd ed.