AI/ML Mini Course
ML Background (Day 1 / Hour 1)
Introduction & Key Tech Shifts
Technological Eras
WWW / Internet: Enabled global connectivity.
Android / iOS / Mobile & Wearable: Brought computing to portable and wearable devices.
TensorFlow / PyTorch: Put machine learning frameworks into the hands of developers, making ML ubiquitous.
Programming Paradigms
Empirical vs. Declarative vs. AI Programming
Empirical: Based on experimentation and observed data.
Declarative: Describes what outcome is desired, not how to compute it.
AI Programming: Systems learn from data rather than following explicitly programmed rules.
AI, ML, DL, and ANNs
Artificial Intelligence (AI)
Aim: Simulate human intelligence in machines.
Foundational Claim: Human intelligence can be precisely described and replicated by a machine.
Machine Learning (ML)
Definition: The study of algorithms that learn from data without explicit instructions.
Relationship to AI: ML is a subset of AI.
Deep Learning (DL)
Definition: Part of ML using artificial neural networks (ANNs) with multiple layers.
Learning modes: Supervised, semi-supervised, or unsupervised.
Artificial Neural Networks (ANN)
Loosely modeled on biological neurons (nodes/neurons connected by weights/biases).
Difference Between ML and AI
ML: Focuses on building models from data for specific tasks.
AI: Encompasses the broader goal of achieving intelligent behavior (ML is a major approach within AI).
Examples of ML & AI
Image recognition, speech recognition, recommendation systems, predictive analytics, etc.
Understanding Deep Learning
Layers (input, hidden, output)
Nodes (neurons)
Weights & Biases
Activation Functions (ReLU, sigmoid, etc.)
Review of Key Terms
Feature: Measurable property of data (input).
Label: Desired output or classification category.
Layer: A group of nodes in an ANN.
Node: The computational unit (neuron).
Training: Creating/learning the model from data.
Inference: Using a trained model on new data.
Regression: Predicts continuous values.
Classification: Predicts discrete categories.
ANN Structure
Nodes: Organized in layers.
Weights: Connection strengths between layers.
Biases: Offset terms in each node.
Activation Functions: Transform outputs (e.g., ReLU, sigmoid).
Common Neural Network Types
Deep Neural Networks (DNNs)
Typically feed-forward; data moves from input to output without cycles.
Recurrent Neural Networks (RNNs)
Allow information to flow backward, useful for sequential data.
LSTM: A special RNN designed for long-term dependencies.
Convolutional Neural Networks (CNNs)
Key for image processing and visual tasks.
Landscape of ML Methods
Supervised Learning: Learning from labeled data (input-output pairs).
Unsupervised Learning: Finding patterns in unlabeled data.
Reinforcement Learning: Learning through rewards and penalties.
Introduction to Recommendation Systems
Basic overview of how ML is used to recommend products or content.
Resources
Machine Learning For Dummies®
ML Training (Day 2 / Hour 2)
How to Train a Machine Learning Model
ML Creation & Configuration
Define model architecture and hyperparameters.
Prepare training data.
Important Hyperparameters
Learning Rate: How quickly weights are updated during optimization.
Epochs: Number of passes through the entire training dataset.
Batch Size: Number of samples processed before the model updates.
Activation Function: Function used to activate neurons (e.g., ReLU).
Hidden Layers: More layers → more complex model (increased expressiveness).
Weight Initialization: Often small random values.
Dropout: Randomly ignoring some neurons to prevent overfitting.
Learning & Optimization
Softmax: Turns raw model outputs into probabilities.
Backpropagation: Algorithm to compute gradients for weight updates.
Gradient Descent: Minimizes loss by adjusting weights in the opposite direction of the gradient.
Loss Functions: Quantify how far the model’s predictions are from desired outputs.
Performance Metrics: Accuracy, precision, recall, F1-score, etc.
Data Handling
Data Partitioning: Split into training, validation, and test sets.
Cross Validation: Multiple train/validation splits to ensure robust performance estimates.
Feature Engineering: Creating or transforming features (e.g., synthetic features).
Overfitting: When the model learns noise in training data; mitigations include dropout, regularization.
Resources
Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow by Aurélien Géron
Python Data Science Handbook
Machine Learning Crash Course (Google)
Udacity Intro to TensorFlow for Deep Learning
Homework
TensorFlow Exercises: Neural Network Playground (playground.tensorflow.org)
ML Prediction Engine (Day 3 / Hour 3)
Building a Basic ML Prediction Engine
Focus on how a model makes inferences using trained parameters.
Gradient Computation & Differential Equations
Partial Differential Equations (PDEs): Sometimes relevant in advanced ML for modeling.
Numerical vs. Symbolic vs. Automatic Differentiation:
Numerical: Approximate derivatives from finite differences.
Symbolic: Manipulate mathematical expressions to get exact derivatives.
Automatic: Systematically applies chain rule to compute derivatives (used internally by ML frameworks).
Practical Steps
Set up Colab or another environment (Python, TensorFlow).
Walkthroughs:
Official TensorFlow fashion MNIST example
Clothing classification demos
Resources
ML & Fashion (Day 4 / Hour 4)
Applying ML to Fashion
Fashion-Specific Challenges:
Attribute prediction (color, style, texture, etc.).
Trend forecasting.
Personalization and recommendation.
Competitions & Datasets
iMaterialist Challenge
Focus: Recognizing product categories and attributes.
Tracks: Furniture/home goods & fashion (attribute prediction).
Dataset: Over 1 million fashion images labeled with various attributes.
Next Steps in ML & Fashion
Combining deep learning with large-scale image datasets (like iMaterialist).
Identifying innovative ways to automate design, production, and styling processes.