- This is the Spring 2025 Seminar in Artificial Intelligence special topics class.
week01 Word Embeddings
- Word tokenization and embeddings with PCA, Word2Vec, and GloVe. 2D projection with t-SNE.
week02 Text Classification
- Translation with linear regression. Sampling strategies for language models: temperature, top-k, beam-search, nuclear-k.
week03 Language Modelling
- Seq2seq model and attention mechanism for translation. Evaluation of language models by perplexity. Subword segmentation: BPE tokenization. BLUE score.
week04 Transformers
- Transformers.
week05 Transfer Learning
- Masked Language Modelling and Next Sentence Prediction objectives for BERT. GPT-2 with fixed positioning encoder.
week06 BERT Models
- BERT-like models. Architecture details: layer norm, pretrained positional encodings, rotary and ALiBi embeddings, gated FFN. Training tips for transformers: learning rate “warm-up”, large batch size, layer norm vs. batch norm. Special tokens [CLS], [SEP], [MASK]. Finetuning BERT.
week07 GPT Models
- GPT-like models. System tokens. Creating and training GPT-2 model with PyTorch transformer layers. Few-shot prompt engineering. Efficiency problems, large-scale training and parallelization.
week08 Fine-Tuning
- Parameter Efficient Fine-Tuning (PEFT): prompt tuning and low-rank adaptors (LoRA).
week09 Reinforcement Learning from Human Feedback
- LLMs alignment with reinforcement learning from human feedback (RLHF). Conversation systems. Instruction fine-tuning.
week10 Quantization
- Model compression and acceleration. Quantization of LLMs.
week11 Retrieval Augmented LMs
- Retrieval-Augmented Generation (RAG).
week12 Multimodal LLMs
- Image captioning and interpretation.
week13 ASR
- Audio LLM