An Obsidian-style knowledge base for the Deep Learning Do It Yourself course, generated from 35 video lecture transcripts.

transcripts/
├── knowledge_base/               # 318 Obsidian-style concept notes
├── knowledge_base_graph.html     # interactive graph visualization
└── transcript_to_module.json     # transcript filename → module ID mapping

An Obsidian-compatible vault of 318 concept notes covering 20 lecture modules, 9 practicals, and 6 bonus lectures. Each note contains:

• YAML frontmatter — aliases, source modules, tags
• Summary — short explanation with [[wikilinks]] to related concepts
• Professor's quotes — exact words with timestamps, organized by module
• See also — links to related concepts

# softmax

The softmax function converts a vector of raw scores ([[logits]]) into a
[[probability distribution]]. It is the standard output activation for
multi-class [[classification]].

## What the professor says

### Module 3 — Loss Functions for Classification

> [29:12 → 29:58] "This function is called the softmax function. Why is
> that? It's because if all the theta values here are large, you see that
> this softmax function will concentrate on the maximum. It is a soft
> version of the max function."

## See also

- [[cross-entropy]] — the loss function paired with softmax
- [[sigmoid]] — the binary classification equivalent
- [[logits]] — the raw network outputs before softmax

• Obsidian — Open knowledge_base/ as a vault in Obsidian to explore the graph visually
• Interactive graph — Open the knowledge base graph in your browser

This knowledge base is designed to be used as a resource by dataflowr-tools. Two MCP tools can leverage it:

• search_knowledge_base(query) — fuzzy-match against note names, aliases, and tags
• get_knowledge_note(concept) — fetch a specific note for the AI tutor to read

This enables queries like "What did the professor say about softmax?" to return a focused ~1 KB note with exact quotes and timestamps, instead of dumping a 47 KB raw transcript into the LLM context.

See LICENSE.