🏛️ Westminster College, London | wcc.co.uk

🧠 Artificial General Intelligence · Foundations to Frontier

🗓️ 6 Saturdays | 2pm–4pm 🎯 Beginner → Advanced 💻 Online · Live + Research‑Oriented
Explore the quest for human‑level AI: cognitive architectures, neural‑symbolic systems, meta‑learning, alignment, safety, and future AGI applications. Theory, frameworks (ACT-R, Soar, PyTorch, RLHF), and a final research project.
🧩 Module 1 — Foundations of Artificial General Intelligence Week 1 · 2h

📜 From Narrow AI to AGI

Narrow AI excels at specific tasks (image recognition, game playing, language translation). AGI aims for human‑like general intelligence: ability to learn, reason, plan, and adapt across diverse domains without retraining. Milestones: Turing Test (1950), Dartmouth workshop (1956), expert systems, deep learning revolution, and recent large language models (GPT‑4) as early precursors.

🧠 Intelligence Theories & Human Cognition

Fluid vs crystallised intelligence, multiple intelligences (Gardner). Cognitive psychology models: working memory, long‑term memory, executive functions. Computational models of reasoning: deductive, inductive, analogical, and causal reasoning.

⚖️ Ethical & Societal Implications

AGI could solve grand challenges (climate, disease) but also poses existential risks. Key topics: job displacement, economic inequality, arms races, loss of human autonomy. Responsible AGI development frameworks: Asilomar principles, EU AI Act, and ongoing global governance discussions.
💡 Discussion: Should AGI development be paused or regulated? Read “Superintelligence” by Nick Bostrom.
🏛️ Module 2 — Cognitive Architectures & Human‑Like Reasoning Week 2 · 2h

📚 ACT‑R, Soar & CLARION

ACT‑R (Adaptive Control of Thought – Rational): models human cognition using modules (visual, declarative, procedural) and a central production system. Soar: problem‑space architecture with universal subgoaling and chunking for learning. CLARION: integrates implicit (sub‑symbolic) and explicit (symbolic) knowledge. These architectures simulate learning, memory, and decision‑making.
# ACT‑R conceptual production rule (pseudo)
IF goal: add numbers
AND retrieval: retrieve fact "2+3"
THEN respond "5" and clear goal

🧩 Memory, Attention & Planning

Memory systems: sensory, working (limited capacity), episodic, semantic. Attention mechanisms (bottom‑up, top‑down) filter information. Planning: hierarchical task networks (HTN), means‑ends analysis, and reinforcement learning based planning. Hybrid cognitive agents combine symbolic planning with neural perception.
🔗 Module 3 — Neural‑Symbolic AI & Knowledge Integration Week 3 · 2h

🧠 Symbolic AI + Deep Learning

Symbolic AI uses logic, rules, and knowledge graphs for reasoning and transparency. Neural networks excel at pattern recognition but lack explicit reasoning. Neural‑symbolic integration: embed symbolic rules into neural architectures (e.g., Logic Tensor Networks, DeepProbLog). Use neural networks to extract symbols from raw data.
# Example: Logic Tensor Networks (conceptual)
import ltn
Predicate = ltn.Predicate.FromLogits(neural_network)
rule = ltn.Forall(ltn.Implies(ltn.And(Predicate(x), Axiom), Predicate(y)))

🗺️ Knowledge Graphs & Commonsense Reasoning

Knowledge graphs (Wikidata, ConceptNet) store entities and relations. Use graph neural networks (GNNs) to reason over them. Commonsense reasoning (e.g., “if you drop glass, it might break”) remains challenging. Approaches: pre‑trained language models (COMET), neuro‑symbolic inference, and physical simulation. Explainable AI (XAI): SHAP, LIME, counterfactual explanations, and symbolic rule extraction from neural networks.
🔄 Module 4 — Meta‑Learning & Autonomous Adaptation Week 4 · 2h

📈 Learning to Learn (Meta‑Learning)

Meta‑learning algorithms learn across tasks to adapt quickly to new tasks with few examples. Approaches: optimisation‑based (MAML, Reptile), metric‑based (Siamese networks, Prototypical networks), and model‑based (memory augmented networks).
# MAML pseudo-code (PyTorch style)
for task in tasks:
# clone model, compute inner update
adapted_params = model_params - alpha * grad(L_task(model_params))
# meta-update on adapted_params across tasks
meta_grad = grad(L_meta(adapted_params))

🧠 Reinforcement Learning & Continual Learning

Reinforcement learning (RL): agent learns via rewards. For AGI, RL must handle sparse rewards, long horizons, and partial observability. Continual learning (lifelong learning) avoids catastrophic forgetting. Techniques: elastic weight consolidation (EWC), replay buffers, progressive networks. Transfer learning adapts knowledge from source to target tasks. Multi‑task learning shares representations. Self‑supervised learning (e.g., contrastive learning) leverages unlabelled data.
⚠️ Module 5 — Reward Modelling, Alignment & AGI Safety Week 5 · 2h

🎯 Reward Modelling & RLHF

Reinforcement Learning from Human Feedback (RLHF) trains a reward model from human preferences, then optimises policy via RL (e.g., PPO). Used in ChatGPT. Preference learning: Bradley‑Terry model, pairwise comparisons. Challenges: reward hacking (agent exploits misspecified reward), scalarisation of multiple objectives.
# RLHF simplified loop
1. Collect human comparisons (a > b)
2. Train reward model r(s,a) to predict preferences
3. Optimise policy π with RL using r

🔒 Value Alignment & AGI Safety

Alignment problem: ensure AGI pursues goals that are beneficial to humans. Approaches: inverse reinforcement learning (IRL), cooperative IRL, and scalable oversight. Safety engineering: robustness, corrigibility (allow shutdown), transparency, and specification (avoid side effects). Governance: OpenAI, DeepMind safety research, Centre for Human‑Compatible AI (CHAI), and Future of Life Institute (FLI).
📖 Recommended reading: “Human Compatible” by Stuart Russell, and the “Alignment problem” by Christian.
🔮 Module 6 — Future AGI Systems & Applied Research Week 6 · 2h

🌐 Large‑Scale Foundation Models & Multimodal AI

Foundation models (GPT‑4, Gemini, Claude) display emergent generalisation: few‑shot learning, reasoning, tool use. Multimodal models integrate text, image, audio (Flamingo, GATO). Embodied AGI: agents that act in simulated/physical environments (VirtualHome, Habitat, robotic platforms). World models for planning (Dreamer).
# Using GPT‑4 for reasoning tasks
response = openai.ChatCompletion.create(model="gpt-4", messages=[{"role": "user", "content": "Solve this puzzle..."}])

🚀 Applied AGI Research: Healthcare, Robotics, Finance

  • Healthcare: diagnostic assistants, drug discovery (AlphaFold), personalised treatment planning.
  • Robotics: general‑purpose manipulation, navigation, and task composition.
  • Finance: autonomous trading, risk assessment, economic modelling.
  • Cybersecurity: intelligent threat detection, adaptive defence.
  • Quantum AGI: quantum machine learning may accelerate AGI capabilities.

🎓 Final AGI Research Project

Project options:
  • Survey and critique of a cognitive architecture (ACT‑R, Soar, or CLARION) with a small simulation.
  • Implement a neural‑symbolic reasoning system (e.g., Logic Tensor Networks on a toy dataset).
  • Build a meta‑learning agent for few‑shot classification (MAML in PyTorch).
  • Develop a safe RL environment and demonstrate alignment (e.g., using the “Reward Modelling” gym).
  • Analyse alignment techniques in large language models (e.g., fine‑tuning with RLHF).
Deliverables: Research paper (4‑6 pages), code repository, and a 10‑minute presentation. Final session includes peer feedback and faculty evaluation.
🏆 Capstone focus: clarity of AGI problem statement, implementation depth, ethical analysis, and future impact discussion.