# Encyclopedia of Artificial Intelligence --- ## Table of Contents 1. [Overview & Definitions](#overview--definitions) 2. [History of AI (1940s–2025)](#history-of-ai-1940s2025) 3. [Major AI Concepts](#major-ai-concepts) - [Machine Learning](#machine-learning) - [Neural Networks](#neural-networks) - [Symbolic AI](#symbolic-ai) - [Reinforcement Learning](#reinforcement-learning) - [Natural Language Processing](#natural-language-processing) - [General AI (AGI)](#general-ai-agi) - [Other Concepts](#other-concepts) 4. [Landmark Algorithms (with Pseudocode)](#landmark-algorithms-with-pseudocode) 5. [Philosophy of Mind](#philosophy-of-mind) 6. [AI Ethics and Alignment](#ai-ethics-and-alignment) 7. [The Next 100 Years: Predictions](#the-next-100-years-predictions) 8. [Index and Glossary](#index-and-glossary) --- ## Overview & Definitions Artificial Intelligence (AI) is the study and design of agents that perceive, reason, learn, and act in an environment to achieve goals. These agents can be formalized as functions: \[ f: \mathcal{P} \to \mathcal{A} \] where $\mathcal{P}$ is the set of possible percepts (input), and $\mathcal{A}$ is the set of possible actions (output). --- ## History of AI (1940s–2025) ### 1940s–1950s: Foundations - **1943**: McCulloch & Pitts formalize artificial neurons. - **1950**: Turing proposes the Turing Test ([Turing, "Computing Machinery and Intelligence"]). - **1956**: Dartmouth Workshop; term "Artificial Intelligence" coined (McCarthy, Minsky, Rochester, Shannon). ### 1960s: Symbolic AI - General Problem Solver (Newell, Simon) - Lisp language (McCarthy) - ELIZA chatbot (Weizenbaum) ### 1970s: Knowledge Engineering - Expert systems (MYCIN, DENDRAL) - SHRDLU (Winograd) - First "AI winter" due to unmet expectations. ### 1980s: Machine Learning & Neural Nets Revival - Backpropagation for neural networks (Rumelhart, Hinton, Williams, 1986) - Decision Trees, Rule-based systems. - Second "AI winter" (late 1980s). ### 1990s: Statistical Approaches - Support Vector Machines (Cortes, Vapnik) - Bayesian Networks (Pearl) - IBM Deep Blue defeats Kasparov (1997). ### 2000s: Big Data & Web - Boosting (Freund, Schapire) - Large-scale ML (Google, Amazon) - Watson wins Jeopardy! (2011). ### 2010s: Deep Learning Boom - ImageNet breakthroughs (Krizhevsky et al., 2012) - AlphaGo defeats Lee Sedol (2016) - Transformer architecture (Vaswani et al., 2017). ### 2020s: Foundation Models & AGI Concerns - GPT-3, GPT-4 (OpenAI) - DeepMind's AlphaFold (protein folding) - Widespread deployment & ethical debates. --- ## Major AI Concepts ### Machine Learning #### Definition Machine Learning (ML) is the study of algorithms that improve their performance at some task $T$ with experience $E$. Formally: \[ \text{Given}: D = \{(x_i, y_i)\}_{i=1}^n, \ x_i \in \mathcal{X}, \ y_i \in \mathcal{Y} \] \[ \text{Learn}:\ f: \mathcal{X} \to \mathcal{Y} \] \[ \text{Optimize}: \min_\theta \mathbb{E}_{(x,y)\sim D}[L(f_\theta(x), y)] \] Where $L$ is a loss function. #### Types - **Supervised Learning**: Labeled data $(x, y)$ - **Unsupervised Learning**: Only $x$ - **Reinforcement Learning**: Learn via rewards $r$ --- ### Neural Networks #### Definition A Neural Network is a parameterized function composed of layers: \[ f(x) = \sigma(W_L \cdots \sigma(W_2 \sigma(W_1 x + b_1) + b_2) + \cdots + b_L) \] where $W_l$ and $b_l$ are weights and biases, and $\sigma$ is an activation function. #### Example: Perceptron \[ y = \text{sign}(w^T x + b) \] --- ### Symbolic AI #### Definition Symbolic AI manipulates symbols and rules explicitly: - Knowledge base $\mathcal{K} = \{\text{facts, rules}\}$ - Inference: Apply logic ($\forall, \exists, \wedge, \vee, \neg$) --- ### Reinforcement Learning #### Definition Agent interacts with environment: - States $s \in \mathcal{S}$ - Actions $a \in \mathcal{A}$ - Rewards $r: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ - Transition $P(s'|s, a)$ Goal: Find policy $\pi(a|s)$ maximizing expected return: \[ J(\pi) = \mathbb{E}_\pi \left[\sum_{t=0}^\infty \gamma^t r_t \right] \] --- ### Natural Language Processing Formalized as learning mapping: \[ f: \text{Text} \to \text{Task-specific output} \] Tasks: parsing, translation, question answering, etc. --- ### General AI (AGI) An agent with human-equivalent or superior performance across *all* cognitive tasks. --- ### Other Concepts - **Unsupervised Learning**: Clustering, dimensionality reduction. - **Self-supervised Learning**: Predicting parts of data from other parts. - **Transfer Learning**: Reusing knowledge from one task/domain to another. - **Foundation Models**: Large models trained on broad datasets (e.g., GPT, BERT). --- ## Landmark Algorithms (with Pseudocode) ### 1. Perceptron ```python def perceptron(X, y, epochs, lr): w = zeros(X.shape[1]) b = 0 for epoch in range(epochs): for xi, yi in zip(X, y): if yi * (dot(w, xi) + b) <= 0: w += lr * yi * xi b += lr * yi return w, b ``` --- ### 2. Backpropagation ```python def backpropagation(network, X, y, lr): # Forward pass activations = [X] for layer in network: X = layer.activate(X) activations.append(X) # Compute loss gradient grad = loss_derivative(activations[-1], y) # Backward pass for i in reversed(range(len(network))): grad = network[i].backward(activations[i], grad, lr) ``` --- ### 3. Q-Learning ```python def q_learning(env, alpha, gamma, episodes): Q = defaultdict(float) for episode in range(episodes): s = env.reset() done = False while not done: a = select_action(Q, s) s_, r, done, _ = env.step(a) Q[s,a] += alpha * (r + gamma * max(Q[s_,a_] for a_ in env.actions()) - Q[s,a]) s = s_ return Q ``` --- ### 4. k-Means Clustering ```python def k_means(X, k, iterations): centers = random.sample(X, k) for _ in range(iterations): clusters = [[] for _ in range(k)] for x in X: idx = argmin([distance(x, c) for c in centers]) clusters[idx].append(x) centers = [mean(cluster) for cluster in clusters] return centers, clusters ``` --- ### 5. Decision Tree (ID3) ```python def id3(examples, attributes): if all_same_class(examples): return leaf(class_of(examples[0])) if not attributes: return leaf(majority_class(examples)) best_attr = argmax(attributes, key=lambda a: info_gain(examples, a)) tree = {} for v in values_of(best_attr): subset = [e for e in examples if e[best_attr] == v] tree[v] = id3(subset, attributes - {best_attr}) return {best_attr: tree} ``` --- ### 6. Support Vector Machine (SVM, simplified primal) \[ \min_{w, b} \frac{1}{2} \|w\|^2 \quad \text{s.t. } y_i(w^T x_i + b) \geq 1, \ \forall i \] --- ### 7. Expectation-Maximization (EM, Gaussian Mixture) ```python def em_gmm(X, k, iterations): # Initialize means, covariances, weights for _ in range(iterations): # E-step: compute responsibilities # M-step: update parameters return means, covariances, weights ``` --- ### 8. AlphaGo’s Monte Carlo Tree Search (MCTS, simplified) ```python def mcts(root, n_simulations): for _ in range(n_simulations): node = root # Selection while node.fully_expanded() and not node.is_terminal(): node = node.best_uct_child() # Expansion if not node.is_terminal(): node = node.expand() # Simulation reward = rollout(node) # Backpropagation node.backup(reward) return root.best_action() ``` --- ### 9. Transformer Attention (Vaswani et al., 2017) \[ \text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V \] ```python def attention(Q, K, V): scores = softmax(Q @ K.T / sqrt(d_k)) return scores @ V ``` --- ### 10. Generative Adversarial Network (GAN) ```python for epoch in range(epochs): # Train discriminator d_loss = train_discriminator(real_data, fake_data) # Train generator g_loss = train_generator() ``` --- ## Philosophy of Mind - **Strong AI Hypothesis**: Machines can have minds/consciousness. - **Chinese Room Argument**: Syntax vs. semantics (Searle). - **Symbol grounding problem**: How symbols acquire meaning. - **Functionalism**: Mental states as computational states. --- ## AI Ethics and Alignment ### Key Issues - **Bias & Fairness**: Data, model, and deployment biases. - **Privacy**: Data usage, surveillance. - **Transparency**: Explainability, auditability. - **Alignment**: Ensuring AI objectives match human values. - **Control**: Safe deployment, corrigibility. --- ## The Next 100 Years: Predictions ### Potential Breakthroughs - Human-level AGI by 2070–2100; possible earlier. - Autonomous scientific discovery, engineering. - Brain-computer interfaces, neuro-symbolic systems. ### Risks - Misaligned superintelligence. - Societal disruption: labor, privacy, autonomy. - Weaponization, surveillance states. ### Societal Transformations - Economic: Automation, new labor markets. - Governance: Regulation, AI rights. - Culture: New forms of art, communication, relationships. ### Unknowns - Emergence of new forms of intelligence. - AI-driven evolution of human cognition. --- ## Index and Glossary | Term | Definition/Section | |---------------------|--------------------------------------------------------| | AGI | [General AI (AGI)](#general-ai-agi) | | Backpropagation | [Backpropagation](#landmark-algorithms-with-pseudocode)| | Bias | [AI Ethics and Alignment](#ai-ethics-and-alignment) | | Deep Learning | [Neural Networks](#neural-networks) | | Decision Tree | [Landmark Algorithms](#landmark-algorithms-with-pseudocode)| | EM Algorithm | [Landmark Algorithms](#landmark-algorithms-with-pseudocode)| | Foundation Model | [Other Concepts](#other-concepts) | | GAN | [Landmark Algorithms](#landmark-algorithms-with-pseudocode)| | k-Means | [Landmark Algorithms](#landmark-algorithms-with-pseudocode)| | MCTS | [Landmark Algorithms](#landmark-algorithms-with-pseudocode)| | Machine Learning | [Machine Learning](#machine-learning) | | Neural Network | [Neural Networks](#neural-networks) | | NLP | [Natural Language Processing](#natural-language-processing)| | Perceptron | [Landmark Algorithms](#landmark-algorithms-with-pseudocode)| | Q-Learning | [Landmark Algorithms](#landmark-algorithms-with-pseudocode)| | Reinforcement Learning| [Reinforcement Learning](#reinforcement-learning) | | SVM | [Landmark Algorithms](#land