Key Artificial Intelligence Techniques

Why This Matters

Artificial intelligence isn't just a buzzword—it's the foundation of how modern systems process information, make decisions, and interact with humans. On the AP exam, you're being tested on your understanding of how different AI techniques work, what problems they solve, and why certain approaches are better suited for specific tasks. The key is recognizing that AI encompasses a spectrum of approaches, from rule-based systems that mimic human expertise to learning-based systems that improve through experience.

These techniques connect directly to broader course concepts like data abstraction, algorithms, and the impact of computing on society. Whether you're analyzing how a recommendation system works or evaluating the ethical implications of facial recognition, you need to understand the underlying mechanisms. Don't just memorize definitions—know what type of problem each technique solves and how it differs from alternatives.

---

Learning from Data

These techniques share a common principle: systems improve their performance by analyzing patterns in data rather than following explicitly programmed rules. This represents a fundamental shift from traditional programming.

Machine Learning

• Enables systems to improve without explicit programming—algorithms identify patterns in training data and apply them to new situations
• Prediction and classification are core capabilities, powering everything from spam filters to credit scoring
• Training data quality directly impacts results—biased or incomplete data leads to flawed outputs, a key ethical consideration

Neural Networks

• Modeled after biological neurons—interconnected nodes process inputs, apply weights, and pass signals forward
• Learning occurs by adjusting connection weights during training until the network produces accurate outputs
• Foundation for more advanced techniques—understanding basic neural networks is essential for grasping deep learning

Deep Learning

• Uses multiple hidden layers to automatically extract increasingly abstract features from raw data
• Excels with unstructured data—images, audio, and text that traditional algorithms struggle to process
• Requires massive datasets and computational power, which explains why it only became practical in recent years

Reinforcement Learning

• Agent learns through trial and error—takes actions, receives rewards or penalties, and adjusts strategy accordingly
• Exploration vs. exploitation tradeoff is central: try new approaches or stick with what works?
• Powers game-playing AI and robotics—systems that must make sequential decisions in dynamic environments

---

Processing Human Communication

Natural language and visual data require specialized techniques because human communication is ambiguous, context-dependent, and unstructured—challenges that traditional computing handles poorly.

Natural Language Processing

• Bridges human language and computer understanding—enables machines to parse meaning from text and speech
• Tokenization, parsing, and semantic analysis break language into processable components
• Powers chatbots, translation, and sentiment analysis—applications you interact with daily

Computer Vision

• Extracts meaningful information from visual data—images and video become analyzable inputs
• Feature extraction identifies patterns like edges, shapes, and textures that distinguish objects
• High-stakes applications include medical diagnosis, autonomous vehicles, and security systems

---

Rule-Based and Logic Systems

Before machine learning dominated, AI relied on encoding human knowledge directly into systems through explicit rules and reasoning frameworks. These approaches remain valuable for specific applications.

Expert Systems

• Captures domain expertise in a knowledge base—rules like "if symptoms X and Y, then consider diagnosis Z"
• Inference engine applies logical reasoning to draw conclusions from available information
• Transparent decision-making is a key advantage—you can trace exactly why the system reached its conclusion

Fuzzy Logic

• Handles degrees of truth rather than binary yes/no—"somewhat hot" or "fairly likely" become valid inputs
• Mirrors human reasoning patterns where boundaries between categories are gradual, not sharp
• Common in control systems—thermostats, washing machines, and other devices requiring nuanced responses

---

Optimization and Physical Systems

These techniques extend AI beyond pure software into solving complex optimization problems and controlling physical machines in the real world.

Genetic Algorithms

• Mimics natural selection to evolve solutions—candidate solutions "reproduce," "mutate," and compete
• Selection, crossover, and mutation operators create new generations of increasingly fit solutions
• Ideal for problems with vast solution spaces where testing every possibility is impossible

Robotics

• Integrates AI with physical actuators and sensors—perception, decision-making, and action in the real world
• Combines multiple AI techniques—computer vision for perception, reinforcement learning for control
• Autonomy exists on a spectrum from fully human-controlled to fully autonomous operation

---

Quick Reference Table

---

Self-Check Questions

1. Which two AI techniques both process unstructured data but differ in whether that data is sequential (language) or spatial (images)?

2. A hospital wants an AI system that can explain exactly why it flagged a patient as high-risk. Should they use deep learning or an expert system, and why?

3. Compare and contrast how machine learning and reinforcement learning each "improve over time"—what's the key difference in their learning mechanisms?

4. An engineer needs to optimize a complex scheduling problem with millions of possible solutions. Which technique mimics biological evolution to find good solutions without testing every possibility?

5. If an FRQ asks you to discuss AI techniques that raise privacy concerns due to their ability to identify individuals, which two techniques from this guide would provide the strongest examples?