2 A Brief Overview of AI
A history recap of AI
What Is AI? — History Recap (Key Points)
1. Foundations of AI (1950s–1960s)
- 1950 – Alan Turing
- Paper: “Computing Machinery and Intelligence”
- Introduced the Turing Test and the question “Can machines think?”
- 1956 – Dartmouth Workshop
- The term Artificial Intelligence was coined
- Key figures: John McCarthy, Claude Shannon, Marvin Minsky
- Early Symbolic AI
- Intelligence described using logic, rules, and symbols
- Chess and language viewed as ideal testbeds
2. Symbolic AI & Expert Systems (1970s–1980s)
- John McCarthy
- Invented LISP (1950s), the main AI language of the time
- AI based on formal logic and rule-based reasoning
- Expert Systems
- Large collections of human-written if–then rules
- Specialized hardware (LISP machines)
- Created a billion-dollar industry
- First AI Spring
- Strong optimism about rule-based intelligence
3. AI Winters & Limitations
- First AI Winter (1970s)
- Early neural networks (perceptrons) failed on complex tasks
- Second AI Winter (late 1980s–early 1990s)
- Expert systems proved brittle and expensive
- Collapse of the LISP machine market
- Often described as the “end of pure symbolism”
- Core Problem
- Hard to precisely describe perception and patterns using rules
4. Statistical Machine Learning (1990s–2000s)
- Shift from rules to data-driven learning
- Models learned from data but relied on human-designed features
- Key advances:
- Boosting (1990)
- Support Vector Machines (1993)
- Random Forests (1995)
- Tom Mitchell’s definition
- A program learns from experience E on tasks T if performance improves
5. Probabilistic & Bayesian AI
- Judea Pearl (Turing Award 2011)
- Introduced Bayesian networks and causal reasoning
- Unified:
- Probability
- Human knowledge (priors)
- Machine learning
- Marked the rise of probabilistic graphical models
6. The Deep Learning Revolution (2012– )
2012: The Magic Year
-
AlexNet (Krizhevsky, Sutskever, Hinton)
- Won ImageNet with a huge margin
- First modern SOTA model using deep neural networks
- End-to-end learning from raw data using GPUs
-
Humans design architectures, not features
Language Domain
- 2003 – Neural Language Model (Bengio)
- First neural approach to language modeling
- 2014 – Seq2Seq (Google)
- First end-to-end neural machine translation model
- 2016 – Google Neural Machine Translation
- 2017 – Transformer
- Became the dominant architecture for sequence modeling
- Achieved SOTA across NLP, vision, and beyond
Reinforcement Learning
Motivation: Beyond Supervised Learning
- Many intelligent tasks require a sequence of decisions, not a single prediction.
- Agents must interact with an environment and adapt based on feedback.
- This setting goes beyond standard machine learning → Reinforcement Learning.
Core Framework of Reinforcement Learning
-
Agent: the decision-maker, defined by a policy (\pi(a|s)) or (\mu(s;\theta))
-
Environment: the external system the agent interacts with
-
State ((s_t)): the current situation of the environment
-
Action ((a_t)): the choice made by the agent
-
Reward ((r_t)): feedback signal evaluating the action
-
Learning paradigm: trial and error
-
Suitable for temporal decision-making problems
Foundation of modern RL
- Policy Gradient Methods for Reinforcement Learning with Function Approximation
- Richard S. Sutton, David McAllester, Satinder Singh, Yishay Mansour
- NIPS 1999