AI310: AI Technology Professional NEW

Agenda:

• Attention & Transformer Fundamentals
  • Context: Why attention? Limitations of classical sequence models, intuition of “soft lookup”
  • Self-attention: concept, process, what “attention” means mathematically
  • Queries, keys, values
  • Multi-head attention: parallelization of different relationship types in text
  • Building blocks of the original Transformer architecture:
    • Encoder/decoder blocks, feed-forward, residuals, LayerNorm
    • Positional information
  • From architecture to tasks: which parts are useful for what?
    
    
• BERT vs. GPT, LLMs & Fine-Tuning (apply, decide, plan)
  • Encoding only: BERT (encoder-only)
    • Core principle
    • Typical tasks: classification, sentiment analysis, extraction, matching/similarity
    • Fine-tuning logic for downstream tasks
  • Decoding only: GPT (decoder-only)
    • Autoregressive generation (next-token principle)
    • Generation logic
    • Typical tasks: text generation, dialogue, translation, summarization
  • Large Language Models (LLMs)
    • What makes them “large”? (scaling concept, data/parameters/compute)
    • What they do well—and typical limitations (hallucinations as a risk in applications)
  • Fine-tuning
    • When fine-tuning makes sense vs. “using the model as-is”
    • Core forms: task fine-tuning (e.g., classification) vs. generative fine-tuning (e.g., style/response format)
    • Requirements: data quality, labeling, overfitting risk, evaluation strategy

Objectives:

• Master the fundamentals of Transformers: be able to clearly explain self-attention, Q/K/V, and multi-head attention
• Architecture understanding: place the original Transformer building blocks in context and justify their purpose
• Differentiate model families: encoder-only (BERT) vs. decoder-only (GPT) — strengths, limitations, typical tasks
• Position LLMs: what defines LLMs, which capabilities emerge from this, and which risks/limitations are relevant
• Plan fine-tuning conceptually: when it’s worth it, which approach to use, and which data/evaluation are needed
• Structure use cases: sentiment analysis, translation, etc. as a repeatable decision and implementation pattern

Target audience:

The training AI310 AI Technology Professional is targeted at:

• Data Scientists, ML Engineers, AI Engineers (beginner to intermediate)
• Data/AI Architects, Tech Leads, Engineering Leads in a data/AI context
• Product/project roles with a technical focus who want to support LLM decisions in a well-founded way

Prerequisites:

For participation in the course AI310 AI Technology Professional the following prerequisites are required:

• Basic understanding of Machine Learning terms (Train/Test, Features/Labels, Overfitting) are helpful
• Basic knowledge in Python

Description:

The course AI310 AI Technology Expert provides a solid technical understanding of modern language models based on the Transformer architecture—from the core ideas of attention to the classification of encoder-only and decoder-only models (BERT vs. GPT) and the fundamental principles of large language models (LLMs).

The workshop AI310 AI Technology Expert is designed so that participants not only know the terms, but truly understand the relationships: why self-attention works, how queries/keys/values interact, what multi-head attention adds, and how this leads to the original Transformer architecture.
On this basis, the most important model families are categorized in a practical way: BERT as an encoder-only model (bidirectional contextual understanding, strong for classification/extraction) and GPT as a decoder-only model (autoregressive generation, strong for text production and generative tasks).

The course AI310 AI Technology Expert also explains what fundamentally defines a large language model (pretraining objective, scaling, tokenization, context window, inference/generation) and how fine-tuning works as a transfer principle—analogous to transfer learning with CNNs, but with typical specifics for language models. The training is rounded out with a systematic bridge to practice: participants learn to structure typical use cases (e.g., sentiment analysis, translation, and others) along the appropriate model choice and a sensible approach (data, training, evaluation, limitations).