A ground-up course preparing you for the PhD comprehensive oral exam on three research papers: Paper 1 — short-term solar power forecasting with deep learning at a utility-scale photovoltaic plant, Paper 2 — teaching robots to build simulations of themselves from raw video, and Paper 3 — an AI-driven cybersecurity framework for IoT combining LSTM anomaly detection, reinforcement learning, and post-quantum encryption.
We start from first principles — no prior background in energy systems, robotics, security, or ML assumed — and climb, module by module, until you can defend every design choice in all three papers.
How to use this course: follow the cards in the order shown below (the Next/Previous links on each page walk the same path) — Modules 12 and 13 were added for Paper 3 and slot into the sequence where shown, so the study order is 0–8, 12, 9, 10, 13, 11. Each module is short, opens with a ~2-minute audio recap, and ends with a quiz — score at least 80% before moving on, because every later module assumes the earlier ones. Quiz scores are saved on this device and shown on each card.
No energy or robotics background needed. Plain-language primers on solar farms, the power grid and why forecasting matters (Paper 1's world) and on robot arms, kinematics and self-models (Paper 2's world) — and what each paper set out to solve.
Vectors, matrices & rotations, functions, derivatives & gradients, time series notation, and basic statistics — the entire mathematical toolkit the papers rely on.
Supervised learning, regression vs classification, MSE and friends, gradient descent, chronological train/validation/test splits, overfitting, and the forecast error metrics NRMSE, MAE and MAPE.
Neurons, ReLU, feedforward networks (FFNN), backpropagation, the Adam optimizer, early stopping, and the hyperparameters — layers, units, batch size — both papers must choose.
Sensor data and its faults, elimination/interpolation/imputation, feature engineering with solar angles, one-hot encoding, normalization, sliding windows — plus video frames, segmentation and pixel imbalance.
Why sequences need memory, vanishing gradients, the LSTM gates, the streamlined GRU, sequence-to-vector forecasting, and multi-target regression for 21 steps ahead.
Learning without labels, motor babbling, query-based models, positional encoding, NeRF-style density & visibility, differentiable rendering — the machinery of the robot self-model.
Deterministic metrics, bootstrap confidence intervals, probabilistic forecasting, honest baselines — and Paper 1's four reasons why "our model is superior" claims are usually unfounded.
Systematic hyperparameter search (the 3-phase framework), K-means clustering with Euclidean distance, choosing K, inverter clustering — and gradient-based inverse kinematics and motion planning.
New for Paper 3: anomaly detection as forecasting-plus-surprise, thresholds, reinforcement learning and the Q-learning update rule, cryptographic hashing, and the quantum threat with lattice/LWE encryption.
The full study: 75 MW plant, 84 inverters, FFNN vs LSTM vs GRU, macro vs inverter-level forecasts, weather-type results, bootstrap CIs, and the "marginal improvement" conclusion.
The free-form kinematic self-model (FFKSM): architecture, training from video, morphology prediction, motion planning without kinematic equations, damage detection and recovery.
The four-component framework end to end: LSTM anomaly detection, homomorphic hashing, the learned Q-learning response policy, LWE encryption — plus results, comparisons, and the critical reading examiners expect.