Lecture

Advanced methods of machine learning (Fortgeschrittene Methoden des maschinellen Lernens (FKI/FML))

Staff

Language

  • German/English

Lecture

The lecture is to be held offline. Please refer to the Moodle course for further informations.

  • Tuesday, 12:00 - 14:00, BC 523
  • Start: 14. October

Exercises

The exercises contain small project, which are splitet in exercises. Start: 17. Oktober

  • Friday
    • 10:00 - 12:00, BC 523

Exam

There is an oral exam.

Topics

  • Fundamentals (neural networks) (definitions, different classes of ML, neurons, layers, neural architectures)
  • Computer vision/audition (definitions, image processing, convolution, special convolutions, CNN, special layers, networks (LeNet, AlexNet, VGG, Inception, ResNet, SqueezeNet, DenseNet, ResNeXt, MobiNet, EfficientNet, two-stage (R-NN, Fast R-CNN, Faster R-CNN), multi-stage detectors, YOLO, SSD, U-Net, audio signal preprocessing, Nyquist–Shannon, spectrogram, TasNet, audio recognition)
  • Generative/probabilistic models (U-Net, latent variable models (autoencoders (AE, CAE, DAE, VAE), GAN (DCGAN, cGAN, CycleGAN, StyleGAN, W-GAN), MiniMax-Game, Autoregressive Models (PixelCNN, WaveNet, GPT), Flow-based Models (Normalizing Flow, RealNVP, Glow, FFJORD), Energy-based Models, State-Space Models)
  • Optimization (over/underfitting, backpropagation, SGD, NAG, Adam, AdamW)
  • Explainability (XAI, explainability, interpretability, Grad-CAM, LRP, XRAI, IG, LIME, SHAP)
  • NLP/LLM (NLP, N-grams, tokenization, BOW, tf-idf, Encoding (WordPiece, Unigram-LM, SentencePiece), Markov Models, HMM, Seq2Seq, Embedding, Positional Embedding, Transformer, Attention (Multihead, Self, Cross, Masked, Flash), Mamba, LLM (Llama, Mistral, Apertus, ChatGPT, LaMDA, PALM, Gemini, Bloom, Gopher, Chinchilla, DeepSeek), scaling laws, reasoning, RAG, MoE, hallucination, agentic LLM)
  • Reinforcement learning (agent, exploitation, exploration, state, action, strategy, Bellman, strategy gradient, Monte Carlo, TD, SARSA, REINFORCE, PPO, A2C, A3C, Q-learning, DQN, DDQN, hierarchical reinforcement learning, feudal neural networks, HIRO, DIAYN, MARL)

Materials & Infos

Moodle-Course

A key is required for registration, which will be announced in the first lecture.