AI / ML

Pure AI/ML concepts — Transformers, attention, KV cache, RL, LLM serving, diffusion, GNNs. Each node carries a mental model, mathematical model, failure modes, and an AI×Networks connection across 4 rounds.

Complete — 6 nodes, Round 4 ✓

Transformer Attention Internals Round 4 ✓

Self-attention, QKV, scaled dot-product, multi-head, the O(n²) wall. The foundational AI node feeding traffic prediction & LLM serving.

KV Cache Mechanics & Inference Memory Layout Round 4 ✓

How the O(n²) compute wall becomes an O(n) memory wall — the constraint that governs edge inference and serving economics.

RL Policy Gradients (REINFORCE → PPO) Round 4 ✓

The log-derivative trick, baselines, and PPO's clipped trust region — the engine behind RL control of network resources.

LLM Serving on Kubernetes Round 4 ✓

Latency vs throughput, continuous batching, autoscaling, cold starts — the direct AI→operations bridge to RIC/CU workloads.

Diffusion Models (DDPM, Score Matching) Round 4 ✓

Generative modeling via learned denoising; the score connection, and offline synthesis of scarce RAN rare-event data.

Graph Neural Networks (Message Passing) Round 4 ✓

Message → aggregate → update; permutation invariance and the WL ceiling. The model class whose bias matches network topology.