Machine Learning Concepts Guide

Table of Contents

Performance Metrics

FLOPS (Floating-Point Operations)

Hardware FLOPS: Floating-Point Operations Per Second - measures computational speed

Model FLOPS: Floating-Point Operations - measures computational complexity

Simple Definition: FLOPS = how much work the model does when processing one image

Think of it as counting how many billion mathematical operations your model performs for a single forward pass (one input image).

Datasets

Dataset Size Categories Description
ImageNet 14 million images 20,000 categories General object recognition
COCO Various sizes 80 object types Common Objects in Context
CIFAR-10/100 60,000 images 10/100 categories 32×32 pixel images (Canadian Institute For Advanced Research)
DOTA Various 15 categories Dataset for Object Detection in Aerial Images
Fashion-MNIST 60,000 images 10 categories Fashion product images

Tools & Frameworks

TorchVision

TorchVision is PyTorch's computer vision library that simplifies development with:

  • Pre-trained Models: ResNet, VGG, Inception, and more
  • Built-in Datasets: CIFAR-10, CIFAR-100, COCO, MNIST, etc.
  • Data Transformations: Resizing, cropping, rotation, normalization, quality enhancement
  • Detection & Segmentation: Faster R-CNN, Mask R-CNN implementations

YAML Configuration Files

YAML (YAML Ain't Markup Language) is the preferred format for configuration files due to: - Python-like syntax with meaningful whitespace - Better readability than JSON - More user-friendly structure

Example YAML Configuration:

# Project configuration
company: spacelift
domain:
  - devops
  - devsecops

tutorial:
  - yaml:
      name: "YAML Ain't Markup Language"
      type: awesome
      born: 2001
  - json:
      name: JavaScript Object Notation
      type: great
      born: 2001
  - xml:
      name: Extensible Markup Language
      type: good
      born: 1996

author: omkarbirade
published: true

Training Concepts

Epochs

Definition: One complete pass through the entire training dataset where every sample is processed and model parameters are updated based on calculated error.

Why Multiple Epochs? The model improves iteratively, adjusting parameters with each pass to minimize prediction errors.

Weights

Purpose: Weights determine how strongly each input influences the final output.

Example: In a course grading prediction model, different factors (marital status, trade, location) are multiplied by specific weights based on their importance to the final grade.

How They Work: 1. During forward propagation, inputs are multiplied by their weights 2. Results pass through an activation function 3. Weights are updated during training via gradient descent 4. Goal: Minimize difference between predicted and actual outcomes

Bias

Analogy: Think of bias as your default emotional state before anything happens in your day. Events (inputs) push you up or down from this starting point.

Mathematical Representation:

y = mx + b
  • m = weight (slope)
  • x = input
  • b = bias (starting point)

Neural Network Fundamentals

Activation Functions

Purpose: The "decision-making" step that determines whether a neuron should activate.

Analogy: Like your brain deciding "Should I go outside?" based on total influence of weather, mood, and time. If the combined score exceeds your threshold (say, 50), you go outside.

Common Types: - ReLU (Rectified Linear Unit) - Most popular for hidden layers - Softmax - Used in output layer for classification

Neural Network Layers

Structure: 1. Input Layer: Receives raw data 2. Hidden Layers: Apply mathematical transformations 3. Output Layer: Produces final predictions

Back Propagation: The learning process where the network adjusts weights to improve accuracy.

Model Optimization

Quantization

Definition: Compressing a trained model by converting 32-bit floating-point numbers to smaller representations (typically 8-bit integers).

Why Quantize? - Faster inference - Lower memory usage - Easier deployment on edge devices

Types of Quantization:

Type What's Quantized Pros Cons
Dynamic Weights only (int8) Fast, easy to implement Moderate accuracy
Static Weights + activations (int8) Better accuracy More complex setup
QAT (Quantization Aware Training) Trained with quantization Best accuracy Slower training

Note: For YOLO models, dynamic or static quantization is commonly applied post-training.

ONNX (Open Neural Network Exchange)

Purpose: A universal format for neural networks - like exporting a Word document as PDF so any device can open it the same way.

Benefits: - Cross-platform compatibility - Framework interoperability - Simplified deployment

Advanced Architectures

Transformer Networks

What They Are: Neural networks designed for sequential data (text, audio, time series).

Key Advantages: - Process data in parallel (much faster than RNNs) - Can handle long-range dependencies - State-of-the-art performance on NLP tasks

Key Concepts:

1. Self-Attention Mechanism

Analogy: Like being in a busy room and selectively focusing on one conversation while filtering out background noise.

Example: Understanding "The trophy wouldn't fit in the suitcase because it was too big." - What does "it" refer to? - Self-attention helps the model understand "it" refers to the trophy, not the suitcase

2. Positional Encoding

Why Needed: Since Transformers process all words simultaneously, they need explicit position information.

Example: Without positional encoding: - "dog bites man" = "man bites dog" (completely different meanings!)

3. Multi-Head Attention

Analogy: Like having multiple people read the same text, each focusing on different aspects: - Person 1: Syntactic relationships (grammar) - Person 2: Semantic similarities (meaning) - Person 3: Subject-object relationships

GPT: Generative Pre-trained Transformer - a specific type of Transformer architecture

Encoder-Decoder Architecture

Components: - Encoder: Processes input and creates a representation - Decoder: Takes the representation and generates output

Neural Network Types Comparison

Type Best For Example Use Cases
CNN (Convolutional Neural Network) Images, spatial data Object detection, image classification
RNN (Recurrent Neural Network) Sequential data Time series, simple text tasks
Transformer Sequential data Language models, translation, complex NLP

Visual Question Answering (VQA) Models

What Are VQA Models?

Input: Image + Text question Output: Text-based answer

Example Questions: - "What is the color of the car in the image?" - "Is the person on the bike wearing a helmet?"

Architecture

VQA models use a multi-modal architecture combining: 1. Vision Component: CNN or Vision Transformer for image understanding 2. Language Component: LLM for text processing and generation 3. Fusion Layer: Combines visual and textual information

Popular VQA Models: - Gemini 3n - PaliGemma - LLaVA - CogVLM - InstructBLIP - DonutBase - BLIP-2

Word Embeddings

What Are Embeddings?

Definition: Converting words into numerical vectors (arrays of numbers) that capture meaning and relationships.

Why Important: Computers can only process numbers, not words directly.

Types of Embeddings

Static Embeddings

Words have fixed representations regardless of context.

Techniques: - Word2Vec: Learns word associations from large text corpus - GloVe (Global Vectors): Captures global word co-occurrence statistics

Contextual Embeddings

Word representations change based on surrounding context.

Example: - "Bank" in "river bank" vs "bank account" - Contextual embeddings understand these different meanings