Artificial Intelligence and Machine Learning

Artificial Intelligence (AI) is the branch of computer science dedicated to building systems that perform tasks ordinarily requiring human intelligence — perception, reasoning, learning, planning, language understanding. Machine Learning (ML) is the subfield in which systems learn patterns from data without being explicitly programmed.

Brief history

Year	Milestone
1950	Alan Turing's paper "Computing Machinery and Intelligence" proposes the Turing Test
1956	Dartmouth Conference — birth of the field; McCarthy coins "Artificial Intelligence"
1965-70	ELIZA chatbot; early expert systems
1980s	Expert systems boom (MYCIN, DENDRAL)
Late 1980s	"AI winter" — funding collapses
1997	Deep Blue defeats Garry Kasparov at chess
2011	IBM Watson wins Jeopardy!
2012	AlexNet — deep learning revolution in image classification
2016	AlphaGo defeats Lee Sedol
2017	Transformer architecture introduced ("Attention Is All You Need")
2020-25	LLMs (GPT, Claude, Gemini) revolutionise NLP

Branches of AI

• Search and problem-solving — uninformed (BFS, DFS) and informed (A*, greedy) search.
• Knowledge representation and reasoning — logic, semantic networks, frames, ontologies.
• Machine learning — supervised, unsupervised, reinforcement.
• Natural Language Processing (NLP).
• Computer Vision.
• Robotics.
• Expert systems — rule-based domain reasoning.

Search algorithms

Uninformed

• BFS — complete, optimal for uniform cost; O(b^d) time and space.
• DFS — low memory; not optimal.
• Iterative Deepening DFS — combines BFS's optimality with DFS's memory profile.

Informed (heuristic)

• Greedy best-first — uses h(n); not optimal.
• A* — uses f(n) = g(n) + h(n); optimal if h(n) is admissible (never overestimates).

Adversarial

• Minimax — assumes optimal opponent; backs up values from leaves.
• Alpha-beta pruning — speeds up minimax without changing outcome.
• Monte Carlo Tree Search (MCTS) — basis of AlphaGo.

Machine learning paradigms

Supervised learning

Learn from labelled (input, output) pairs. Goal: predict outputs for new inputs.

• Classification — discrete labels (spam vs. not spam).
• Regression — continuous outputs (predicting house prices).

Common algorithms: linear/logistic regression, k-Nearest Neighbours, decision trees, random forests, gradient boosting (XGBoost, LightGBM), Support Vector Machines, neural networks.

Unsupervised learning

No labels; find structure.

• Clustering — k-means, hierarchical, DBSCAN.
• Dimensionality reduction — PCA, t-SNE, UMAP.
• Association rules — Apriori (market-basket analysis).

Reinforcement learning

Agent interacts with environment, learns policy maximising cumulative reward.

• Markov Decision Process formalism: states, actions, rewards, transition probabilities.
• Q-learning, SARSA, Policy gradients, Actor-critic methods.
• Deep RL combines neural nets with RL (DQN, PPO, A3C).
• Applications: game AI, robotics, recommendation systems.

Semi-supervised, self-supervised, transfer learning

These approaches exploit large unlabelled corpora (text from the web) to pre-train powerful models, then fine-tune on small labelled datasets — the basis of modern LLMs.

Model evaluation

Classification metrics

• Accuracy = correct / total.
• Precision = TP / (TP + FP).
• Recall (sensitivity) = TP / (TP + FN).
• F1 score = harmonic mean of precision and recall.
• ROC curve and AUC.
• Confusion matrix.

Regression metrics

• Mean Absolute Error (MAE).
• Mean Squared Error (MSE).
• Root Mean Squared Error (RMSE).
• R² (coefficient of determination).

Train-test split, cross-validation, regularisation (L1 LASSO, L2 ridge), early stopping — all techniques to combat overfitting.

Deep learning

Deep neural networks have many hidden layers; trained via backpropagation and gradient descent variants (SGD, Adam, RMSProp).

Architectures

• Multilayer Perceptron (MLP) — fully connected.
• Convolutional Neural Networks (CNN) — image and grid data; LeNet, AlexNet, VGG, ResNet.
• Recurrent Neural Networks (RNN) — sequences; LSTM, GRU handle long dependencies.
• Transformers — self-attention; basis of BERT, GPT, T5, LLaMA.
• Autoencoders — unsupervised representation learning.
• Generative Adversarial Networks (GANs) — generator vs. discriminator.
• Diffusion models — image and video generation (Stable Diffusion, DALL-E, Sora).

Natural Language Processing

Tasks:

• Tokenisation, POS tagging, parsing.
• Named Entity Recognition (NER).
• Sentiment analysis.
• Machine translation — neural sequence-to-sequence with attention.
• Question answering.
• Summarisation.
• Speech recognition / synthesis.

Pakistani relevance: NLP for Urdu, Punjabi, Pashto, Sindhi, Balochi remains under-resourced — a research frontier.

Computer vision

• Image classification, object detection (YOLO, Faster R-CNN), segmentation (U-Net, Mask R-CNN), image generation, face recognition.
• Pakistan's Safe City projects in Lahore, Islamabad and other cities deploy CCTV with face/plate recognition.

Generative AI and LLMs

Large Language Models (LLMs) trained on internet-scale text exhibit few-shot reasoning, code generation and dialogue. Examples: GPT-4/5 (OpenAI), Claude (Anthropic), Gemini (Google DeepMind), LLaMA (Meta), Mistral.

Concepts:

• Pre-training on massive unlabelled corpora.
• Fine-tuning for specific tasks.
• RLHF (Reinforcement Learning from Human Feedback) for alignment.
• Prompt engineering, Retrieval-Augmented Generation (RAG).

AI ethics and risks

Concerns include:

• Bias and fairness — models reflect biases in training data.
• Privacy — training data may leak personal info.
• Hallucinations in LLMs.
• Job displacement.
• Autonomous weapons.
• Misinformation and deepfakes.
• Concentration of compute in a few firms.
• AGI safety — long-term existential debates.

Regulatory frameworks: EU AI Act (2024) is the world's first comprehensive AI law, classifying systems by risk. Pakistan's National AI Policy (draft, 2023-24) envisions sectoral guidelines and capacity building.

Pakistan and AI

• Government's Centre for Artificial Intelligence and Computing (CENTAIC) at PAF.
• National Centre of AI with university partners (UET, NUST, ITU, Bahria, FAST).
• Universities offering BS/MS AI programmes are growing.
• Applications under development: precision agriculture, e-health diagnostics, fraud detection in banking, traffic management, electoral roll analytics at NADRA.

A CSS officer who can distinguish hype from substance and understand the data, computational and ethical demands of AI is well-placed to engage with the rapidly evolving policy frontier.