Learning in AI: Types, Methods & Neural Learning

TAKEAWAYS:

• Learning in AI is the ability to improve performance from experience, using data or feedback.
• Supervised learning uses labeled data (e.g., spam detection), while unsupervised learning finds hidden patterns (e.g., customer segmentation).
• Inductive learning generalizes from examples (e.g., medical diagnosis), while analogy-based learning solves problems by mapping known solutions (e.g., Daraz’s route optimization).
• Neural networks learn via backpropagation (adjusting weights to minimize error) and activation functions (e.g., ReLU, sigmoid).
• Overfitting (memorizing training data) and underfitting (failing to capture patterns) are key pitfalls in learning models.

1. What is Learning in AI?

Learning in AI is the process by which a system improves its performance on a task by analyzing data, identifying patterns, and making decisions with minimal human intervention. Unlike traditional rule-based systems, learning systems adapt over time.

Key Characteristics of Learning in AI:

• Experience: The system improves based on input data or feedback.
• Performance: The system’s output becomes more accurate or efficient.
• Generalization: The system applies learned knowledge to unseen data.

Types of Learning in AI

Learning can be classified into three broad categories:

2. Supervised Learning

Supervised learning is the most common type, where the model is trained on a dataset with known inputs and outputs.

How It Works:

1. Training Data: A dataset with input features (X) and corresponding labels (Y). Example: X = [email text], Y = [spam/ham].
2. Model Training: The algorithm learns a mapping function f(X) → Y.
3. Prediction: The trained model predicts Y for new X.

Example: Email Spam Detection

• Dataset: 1000 emails labeled as spam/ham.
• Features: Words like "free," "offer," "win."
• Model: Logistic Regression or Naive Bayes.
• Prediction: New email with "free offer" → classified as spam.

Real-World Application: Khalti’s Fraud Detection

Khalti uses supervised learning to detect fraudulent transactions:

• Input: Transaction amount, time, location.
• Output: Fraud (1) or Legitimate (0).
• Model: Random Forest or SVM.
• Result: Reduces false positives in payments.

3. Unsupervised Learning

Unsupervised learning works with unlabeled data to discover hidden structures.

Common Techniques:

• Clustering: Groups similar data points (e.g., K-Means).
• Dimensionality Reduction: Reduces feature space (e.g., PCA).
• Association Rule Learning: Finds relationships (e.g., Market Basket Analysis).

Example: Customer Segmentation for Daraz

Daraz uses K-Means clustering to group customers:

1. Data: Purchase history, browsing behavior.
2. Clusters: High-value buyers, occasional shoppers, bargain hunters.
3. Marketing: Target promotions based on cluster.

Visual: K-Means Clustering

graph TD
    A["Data Points"] --> B["Cluster 1: High Spenders"]
    A --> C["Cluster 2: Occasional Buyers"]
    A --> D["Cluster 3: Bargain Hunters"]
    B -->|"Promotions"| E["Personalized Offers"]
    C -->|"Discounts"| E
    D -->|"Flash Sales"| E

4. Inductive Learning

Inductive learning generalizes from specific examples to broader rules.

How It Works:

1. Observations: Specific instances (e.g., "All observed swans are white").
2. Generalization: Derives a rule (e.g., "All swans are white").
3. Prediction: Applies the rule to new cases.

Example: Medical Diagnosis

• Data: Symptoms of 1000 patients with/without diabetes.
• Rule: "If blood sugar > 126 mg/dL, predict diabetes."
• Prediction: New patient with high blood sugar → diagnosed.

Limitations:

• Overfitting: Model memorizes training data but fails on new data.
• Bias: Assumes all swans are white (ignores black swans in Australia).

5. Learning by Analogy

Analogy-based learning solves new problems by mapping them to known solutions.

Types of Analogy:

Example: Pathao’s Delivery Route Optimization

• Known Problem: Shortest path in a graph (Dijkstra’s algorithm).
• New Problem: Real-time traffic delays.
• Analogy: Adjusts Dijkstra’s algorithm with live traffic data.

6. Learning in Neural Networks

Neural networks learn by adjusting weights to minimize prediction error.

Key Concepts:

• Activation Functions: Introduce non-linearity (e.g., ReLU, Sigmoid).
• Backpropagation: Adjusts weights using gradient descent.
• Loss Function: Measures prediction error (e.g., Mean Squared Error).

Example: Neural Network for Handwritten Digit Recognition

1. Input Layer: 28×28 pixels (MNIST dataset).
2. Hidden Layers: Extract features (edges, curves).
3. Output Layer: Predicts digit (0–9).
4. Training: Adjusts weights to minimize misclassification.

Visual: Neural Network Layers

graph TD
    A["Input Layer\n(784 neurons)"] --> B["Hidden Layer 1\n(256 neurons, ReLU)"]
    B --> C["Hidden Layer 2\n(128 neurons, ReLU)"]
    C --> D["Output Layer\n(10 neurons, Softmax)"]
    D -->|"Prediction"| E["Digit: 5"]

Real-World Application: Ncell’s Churn Prediction

Ncell uses neural networks to predict customer churn:

• Input: Call duration, data usage, complaints.
• Output: Churn probability (0–1).
• Action: Retention offers for high-risk users.

7. Challenges in Learning

In the Real World

1. Khalti’s Fraud Detection

   • Uses supervised learning (SVM/Random Forest) to classify transactions as fraudulent or legitimate.
   • Example: A transaction from Kathmandu to Pokhara at 3 AM → flagged as high-risk.

2. Daraz’s Recommendation System

   • Uses collaborative filtering (unsupervised learning) to suggest products.
   • Example: If User A buys a phone case, Daraz recommends similar cases to User B.

3. NTC’s Traffic Prediction

   • Uses time-series forecasting (neural networks) to predict congestion.
   • Example: During Dashain, NTC predicts delays on Ring Road and reroutes buses.

4. Pathao’s Delivery Optimization

   • Uses reinforcement learning to adjust delivery routes dynamically.
   • Example: Avoids traffic jams by learning from past deliveries.

Exam Tip

• Define clearly: Always start with definitions (e.g., "Supervised learning is...").
• Use examples: Relate concepts to real-world apps (e.g., Khalti for fraud detection).
• Compare methods: Draw tables for supervised vs. unsupervised learning.
• Diagrams: Sketch neural network layers or decision trees in exams.
• Pitfalls: Mention overfitting/underfitting when discussing learning challenges.

Worked Example for Exam: Question: Explain supervised learning with an example from NEPSE. Answer: Supervised learning uses labeled data to train a model. For NEPSE:

1. Data: Historical stock prices (input) and market trends (output).
2. Model: Linear Regression or Decision Tree.
3. Prediction: Forecasts stock prices for the next trading day. Visual:

graph TD
    A["Historical Data\n(2010-2023)"] --> B["Train Model\n(Linear Regression)"]
    B --> C["Predict 2024\nStock Price: NRS 5000]"]