Unit 4: Evolution and Modern AI Paradigms

I. Orientation

Artificial Intelligence has evolved from systems that explicitly manipulate symbols and rules into systems that learn statistical patterns from data, and then into large, multimodal, adaptive models. Modern AI combines algorithms, data, computing hardware, and feedback to perform tasks such as prediction, generation, perception, planning, and control.

• Representation: Intelligence depends on how knowledge is encoded, such as rules, vectors, probability distributions, neural weights, or sensor signals.
• Learning: A system improves by adjusting parameters from examples, experience, or feedback.
• Inference: The system uses learned or programmed information to produce predictions, decisions, explanations, or actions.
• Generalization: A useful model performs well on previously unseen data rather than merely memorizing training examples.
• Objective: AI systems optimize a goal, such as minimizing prediction error or maximizing cumulative reward.
• Deployment constraint: Accuracy must be balanced against cost, speed, privacy, safety, energy use, and interpretability.

II. Symbolic and Statistical Artificial Intelligence

A. Symbolic and statistical Artificial Intelligence

Symbolic AI represents knowledge explicitly, whereas statistical AI learns regularities from data; modern systems often combine both approaches.

• Symbolic AI: Uses human-defined rules, logic, ontologies, and search.
  • A rule such as IF temperature > 38°C THEN fever = true is transparent and directly inspectable.
  • Expert systems, theorem provers, and classical planning systems are symbolic applications.
• Statistical AI: Represents knowledge through probabilities and learned parameters.
  • A spam classifier may estimate (P(\text{spam}\mid\text{words})) from labelled emails.
  • Neural networks learn numerical weights rather than storing explicit rules.
• Key contrast: Symbolic systems are usually explainable and data-efficient but brittle outside encoded cases; statistical systems handle noisy data but may be opaque and data-hungry.
• Hybrid direction: Neuro-symbolic AI can use neural perception to extract facts and symbolic reasoning to apply constraints or logical rules.

III. Overview of Machine Learning

A. Overview of Machine Learning

Machine learning is the study of algorithms that learn a function from data instead of receiving every decision rule explicitly.

• Model: A parameterized function (f_\theta(x)), where (x) is an input and (\theta) represents learned parameters.
• Training: Parameters are adjusted to reduce a loss function.

repeat:
    predict y_hat = f_theta(x)
    calculate loss L(y_hat, y)
    update theta using the gradient of L

• Generalization: Training data, validation data, and test data evaluate learning, tuning, and final performance respectively.
• Overfitting: A model performs well on training examples but poorly on new examples; regularization, more data, or simpler models can reduce it.
• Evaluation: Classification may use accuracy, precision, recall, or F1-score; regression commonly uses mean squared error.

IV. Supervised Learning

A. Supervised learning

Supervised learning learns from examples containing both input features and known target labels.

• Classification: Predicts a discrete class, such as fraud or legitimate.
  • Binary classification has two classes; multiclass classification has more than two.
• Regression: Predicts a continuous value, such as house price in dollars.
• Training objective: For data pairs ((x_i,y_i)), the model minimizes

theta* = argmin_theta (1/n) sum from i=1 to n L(f_theta(x_i), y_i)

where (n) is the number of examples, (L) is the loss, and (\theta^*) is the selected parameter set.

• Applications: Medical-image diagnosis, speech recognition, credit-risk estimation, and demand forecasting.
• Limitation: Labels may be expensive, inconsistent, biased, or unavailable for rare events.

V. Unsupervised Learning

A. Unsupervised learning

Unsupervised learning discovers structure in data without target labels supplied for each example.

• Clustering: Groups similar observations, such as customer segments.
  • In (k)-means, each point is assigned to one of (k) centroids and centroids are repeatedly recomputed.
• Dimensionality reduction: Compresses features while preserving important variation.
  • Principal Component Analysis projects data onto directions of maximum variance.
• Representation learning: Learns useful features directly from raw inputs, including embeddings for words or images.
• Anomaly detection: Identifies observations that differ substantially from the learned normal pattern, such as an unusual network login.
• Limitation: Discovered groups require human interpretation; minimizing a mathematical objective does not guarantee meaningful categories.

VI. Reinforcement Learning

A. Reinforcement learning

Reinforcement learning trains an agent to select actions in an environment so that long-term cumulative reward is maximized.

• Interaction cycle: At time (t), the agent observes state (s_t), chooses action (a_t), receives reward (rt), and reaches (s{t+1}).
• Policy: A policy (\pi(a\mid s)) specifies the probability of taking action (a) in state (s).
• Return: Future rewards are discounted using

G_t = r_(t+1) + gamma*r_(t+2) + gamma^2*r_(t+3) + ...

where (G_t) is the return and (0 \leq \gamma < 1) is the discount factor.

• Exploration and exploitation: The agent must try uncertain actions while also using actions already known to produce good rewards.
• Applications: Robot control, game playing, traffic-signal management, and resource allocation.
• Limitation: Poorly designed rewards can produce unsafe or unintended behavior; real-world exploration may be costly.

VII. Introduction to Deep Learning

A. Introduction to Deep Learning

Deep learning uses neural networks with multiple computational layers to learn hierarchical representations from large datasets.

• Neuron computation: A unit calculates (z = w^\top x+b), then applies an activation function such as ReLU, ( \max(0,z) ).
• Backpropagation: The chain rule calculates how each parameter contributes to error, allowing gradient-based updates.
• Architecture: Early layers may detect edges or simple patterns, while later layers combine them into objects, concepts, or language structures.
• Strength: Deep models reduce the need for manually designed features in images, audio, text, and video.
• Requirements: They commonly need substantial data, accelerated hardware, careful optimization, and monitoring for bias or instability.

VIII. Types and applications of Deep Learning

A. Types and applications of Deep Learning

Different deep-learning architectures are designed for different data structures and operational tasks.

• Convolutional neural networks: Use local filters and shared weights for spatial patterns; they support image classification, object detection, and medical scans.
• Recurrent and sequence models: Process ordered inputs such as speech or time series; LSTM networks use gates to preserve relevant information.
• Transformers: Use attention to relate tokens or input positions, enabling translation, language generation, document analysis, and vision tasks.
• Autoencoders: Encode data into a compact representation and reconstruct it; reconstruction error can support compression or anomaly detection.
• Generative models: Produce new text, images, audio, or video by learning a data distribution; diffusion models generate samples through iterative denoising.
• Operational concerns: Applications require metrics suited to the task, such as intersection-over-union for object detection or word-error rate for speech recognition.

IX. Foundation models

A. Foundation models

Foundation models are large models pretrained on broad, often heterogeneous datasets and adapted to many downstream tasks.

• Pretraining: A language model may learn by predicting a missing or next token from billions of text examples.
• Transfer: The same model can be adapted through prompting, fine-tuning, or retrieval-augmented generation for specialized work.
• Scale: Performance often improves with model size, data, and computation, although gains are not guaranteed and costs increase.
• Capabilities: Foundation models can summarize, translate, classify, generate code, answer questions, and extract information.
• Risks: They may hallucinate, reproduce training-data bias, expose sensitive information, or generate insecure code.
• Governance: Evaluation, access controls, provenance checks, human review, and domain-specific validation are required before high-stakes deployment.

X. Multimodal Artificial Intelligence

A. Multimodal Artificial Intelligence

Multimodal AI processes or generates more than one data modality, such as text, images, audio, video, or sensor readings.

• Fusion: Inputs can be combined early at the feature level, jointly through attention, or late by combining separate model decisions.
• Alignment: A model learns relationships between modalities, such as associating the sentence “a red car” with corresponding image regions.
• Example: A vision-language model can inspect a radiograph and produce a text description, but clinical decisions still require qualified review.
• Applications: Accessibility tools, visual question answering, robotics, education, autonomous vehicles, and content search.
• Challenge: Modalities differ in timing, resolution, noise, and reliability; a misleading image or corrupted sensor can distort the combined output.

XI. Agentic Artificial Intelligence

A. Agentic Artificial Intelligence

Agentic AI refers to systems that pursue goals by planning, using tools, maintaining state, and taking multiple actions within an environment.

• Core loop: The system observes context, plans a step, invokes a tool or action, evaluates the result, and revises its plan.
• Components: Typical components include a language or reasoning model, memory, a planner, tool interfaces, permissions, and an execution monitor.
• Example: A software agent may read a support request, query an order database, draft a response, and request human approval before issuing a refund.
• Difference from chat: A chat model mainly generates responses; an agent can cause external state changes through APIs or devices.
• Safety controls: Tool permissions, action limits, audit logs, sandboxing, confirmation gates, and rollback mechanisms reduce harmful autonomy.
• Failure mode: A vague goal can cause “goal drift,” while incorrect tool results can propagate through several later actions.

XII. Edge Artificial Intelligence

A. Edge Artificial Intelligence

Edge AI runs inference near the location where data is produced, such as on a phone, camera, vehicle, or factory controller.

• Latency: Local inference can respond in milliseconds without sending every sensor event to a distant cloud server.
• Privacy: Keeping raw speech, images, or biometric data on-device can reduce exposure during transmission.
• Connectivity: Edge systems continue operating during weak or absent network connections.
• Trade-off: Edge hardware has less memory and computing power, so models may need quantization, pruning, or hardware acceleration.
• Example: A factory camera can detect a defective component locally and transmit only an alert rather than continuous video.
• Limitation: Updating models, securing distributed devices, and monitoring performance across different hardware are operational challenges.

XIII. Tiny Artificial Intelligence

A. Tiny Artificial Intelligence

Tiny AI, or TinyML, places machine-learning inference on extremely resource-constrained microcontrollers and embedded devices.

• Hardware: A microcontroller may have kilobytes of RAM, limited flash storage, and a low-power processor rather than a GPU.
• Model compression: Quantization converts 32-bit floating-point weights to 8-bit integers; pruning removes low-value connections.
• Power objective: Devices such as wearables or sensors may operate for months on a battery by processing short data windows locally.
• Applications: Wake-word detection, vibration-based machine maintenance, gesture recognition, environmental monitoring, and fitness sensing.
• Example: A sensor can classify a motor’s vibration as normal or abnormal without uploading the raw waveform.
• Constraint: Smaller models may lose accuracy, and limited memory restricts input size, model complexity, and update mechanisms.