AI Basics: Definitions, Techniques & Real-World Impact

What is Artificial Intelligence?

Artificial Intelligence (AI) is the simulation of human intelligence in machines—programs that can perform tasks requiring human-like reasoning, learning, perception, or decision-making. Unlike traditional software (which follows fixed rules), AI systems improve with experience or adapt to new data.

Key Characteristics of AI:

mindmap
  root((AI Characteristics))
    Intelligence["Thinks & Learns"]
      Reasoning["Logical problem-solving"]
      Learning["Improves from data"]
      Self-Correction["Adjusts mistakes"]
    Human-Like["Mimics human traits"]
      Perception["Sees/hears like humans"]
      NaturalLanguage["Understands speech/text"]
      Creativity["Generates new ideas"]
    Autonomy["Operates independently"]
      DecisionMaking["Chooses actions"]
      Adaptation["Changes for goals"]

A simple neural network with 3 layers (input, hidden, output) and weighted connections. (Image: Dake, Mysid, CC BY 1.0, via Wikimedia Commons)

Subfields of AI

AI is divided into three major branches, each with distinct techniques and applications:

Why it matters: Most real-world AI (e.g., Khalti’s risk scoring) uses ML, while NLP powers customer service (e.g., Ncell helpline bots).

How AI Works: Core Techniques

1. Machine Learning (ML) – Learning from Data

ML algorithms find patterns in data to make predictions or decisions. Three main types:

flowchart TD
  A["Machine Learning"] --> B["Supervised"]
  A --> C["Unsupervised"]
  A --> D["Reinforcement"]

  B --> B1["Labeled Data\n(e.g., spam/not spam)"]
  C --> C1["Unlabeled Data\n(e.g., customer segments)"]
  D --> D1["Trial & Error\n(e.g., game AI)"]

Worked Example: Predicting Loan Defaults (Nepalese Banks)

• Problem: A bank (e.g., Nabil Bank) wants to predict if a customer will default on a loan.
• Data: Past loan records (amount, income, credit score, repayment history).
• Algorithm: Logistic Regression (a supervised ML model).
• Steps:
  1. Train: Feed historical data (labeled as "default" or "no default").
  2. Predict: For a new applicant, the model outputs a probability of default (e.g., 85% → reject).
• Real Impact: Reduces bad loans, saving banks millions of NPR/year.

2. Neural Networks – The Brain of AI

Inspired by the human brain, neural networks process data through layers of interconnected nodes (neurons). Each connection has a weight that adjusts during learning.

Worked Example: eSewa’s OCR for Bill Payments

• Problem: Users upload handwritten electricity bills (NTC) to pay via eSewa.
• Solution: A Convolutional Neural Network (CNN) scans the image, extracts text, and matches it to the NTC database.
• Why CNN?: Specialized for image recognition (unlike general neural networks).

AI in the Real World

1. eSewa: Fraud Detection with ML

• AI Idea: Anomaly Detection (unsupervised ML).
• How it works: Monitors transactions in real-time. If a user suddenly sends 10x their usual amount to an unknown account, the system flags it as fraud.
• Impact: Blocks ~30% of fraudulent transactions in Nepal annually.

2. Pathao: Dynamic Pricing & Route Optimization

• AI Idea: Reinforcement Learning + Graph Algorithms.
• How it works:
  • Pricing: Adjusts fares based on demand (e.g., surge pricing during Kathmandu traffic jams).
  • Routing: Uses A algorithm* to find the fastest path, avoiding accidents or roadblocks (real-time data from Google Maps API).
• Impact: Reduces driver idle time by 25% and improves user satisfaction.

3. NEPSE: Stock Market Prediction

• AI Idea: Time-Series Forecasting (e.g., LSTM neural networks).
• How it works: Analyzes past stock prices, news sentiment (from Nepali newspapers), and global trends to predict NEPSE index movements.
• Impact: Helps traders make data-driven decisions (though not 100% accurate!).

AI vs. Traditional Programming

Visual Comparison:

stateDiagram-v2
  [*] --> Traditional: "Fixed Rules"
  Traditional --> ATM: "Withdrawal: Check PIN → Check Balance → Dispense Cash"
  Traditional --> Error: "If PIN wrong → Block Card"

  [*] --> AI: "Learns from Data"
  AI --> Chatbot: "User: 'Hello'\nBot: 'Hi! How can I help?' (adjusts responses over time)"
  AI --> FraudDetection: "Detects new fraud patterns without reprogramming"

Advantages and Limitations of AI

✅ Advantages

• Automation: Reduces human error (e.g., NTC’s automated bill processing).
• Speed: Analyzes data faster than humans (e.g., Daraz’s inventory management).
• 24/7 Operation: No fatigue (e.g., Ncell’s customer service bots).
• Personalization: Tailors experiences (e.g., YouTube recommendations).

❌ Limitations

• Data Hunger: Needs large, clean datasets (Nepal’s AI struggles with limited data).
• Bias: Learns from biased data (e.g., facial recognition fails for darker skin tones).
• Lack of Common Sense: Struggles with ambiguous queries (e.g., "What’s the meaning of life?").
• Ethical Concerns: Privacy risks (e.g., AI tracking user behavior on Daraz).

Ethical Issues in AI

1. Privacy: AI systems (e.g., Google’s search history tracking) collect vast personal data.
2. Bias: If trained on biased data, AI can discriminate (e.g., hiring algorithms favoring certain universities).
3. Job Displacement: Automation may replace roles (e.g., self-checkout kiosks replacing cashiers).
4. Accountability: Who is responsible if an AI makes a mistake? (e.g., self-driving car accidents).

Nepalese Example:

• Problem: Nepal Police’s AI surveillance risks violating citizens’ privacy.
• Solution: Transparent laws on AI data usage are needed.

Exam Tip: How to Score Full Marks

1. Define AI clearly (1 mark):

   "AI is the field of computer science that enables machines to perform tasks requiring human intelligence, such as learning, reasoning, and problem-solving."

2. Applications (4 marks):

   • Healthcare: AI diagnoses diseases from X-rays (e.g., lung cancer detection).
   • Transportation: Self-driving cars (e.g., Tesla’s Autopilot).
   • Education: Adaptive learning platforms (e.g., Khan Academy’s AI tutors).
   • Entertainment: Deepfake videos or AI-generated music (e.g., Boomy app).

3. Worked Examples (3–5 marks):

   • Always show steps (e.g., for ML: data → training → prediction).
   • Use Nepalese context (e.g., eSewa, NTC, banks).

4. Diagrams (2 marks):

   • Draw a neural network or ML workflow if asked for a technique.
   • Label all parts (input, hidden layers, output).

5. Compare AI vs. Traditional Systems (3 marks):

   • Use the table above or a Venn diagram (if allowed).

Final Note: AI is everywhere—from your Khalti app to NTC’s smart meters. Focus on real-world examples and how AI solves problems, not just theory. Practice drawing neural networks and ML workflows for diagrams!