https://github.com/klauseduard/deep-learning-knowledge-base-demo
Demo: structured deep learning knowledge base with interactive teaching materials — built with Claude Code
https://github.com/klauseduard/deep-learning-knowledge-base-demo
backpropagation deep-learning interactive-learning knowledge-base machine-learning neural-networks obsidian teaching-materials
Last synced: 3 months ago
JSON representation
Demo: structured deep learning knowledge base with interactive teaching materials — built with Claude Code
- Host: GitHub
- URL: https://github.com/klauseduard/deep-learning-knowledge-base-demo
- Owner: klauseduard
- Created: 2026-04-17T05:37:16.000Z (3 months ago)
- Default Branch: master
- Last Pushed: 2026-04-17T07:00:28.000Z (3 months ago)
- Last Synced: 2026-04-17T07:31:52.576Z (3 months ago)
- Topics: backpropagation, deep-learning, interactive-learning, knowledge-base, machine-learning, neural-networks, obsidian, teaching-materials
- Language: HTML
- Size: 286 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Deep Learning Knowledge Base — Demo
A demonstration of building a structured knowledge base and interactive teaching materials using [Claude Code](https://claude.ai/code), with deep learning as the subject matter.
## What this is
This repository demonstrates a workflow for learning a complex technical domain with an LLM as a collaborator. The knowledge base started as AI-generated drafts seeded from François Fleuret's *The Little Book of Deep Learning* (2023), then improved through iterative review — the learner reads the articles, asks questions about what's unclear or missing, and the material gets refined or expanded in response. Interactive teaching materials emerged from this process: each one was prompted by a genuine question ("I forgot the math — show me with real numbers", "what happens when I stack 20 sigmoid layers?") rather than planned upfront.
Most content has been removed from this public repository. A knowledge base built this way is inherently personal — shaped by one learner's questions, gaps, and interests — and publishing it as a reference would misrepresent its purpose. Additionally, the LLM can generate draft articles far faster than a person can engage with them, so much of the initial material hasn't been reviewed or refined yet. What remains here is enough to demonstrate the *approach*: the article format, the teaching materials, and how they connect.
```mermaid
flowchart LR
subgraph seed ["1. Seed"]
S[Source material
book, paper, topic]
end
subgraph generate ["2. Generate"]
G[LLM drafts
concept articles]
end
subgraph learn ["3. Learn"]
L[Learner reads
and questions]
end
subgraph refine ["4. Refine"]
direction TB
R1[Clarify jargon
add links]
R2[Add worked
examples]
R3[Generate teaching
material]
R4[Create reference
entries]
end
S --> G --> L
L -- "unclear term" --> R1
L -- "forgot the math" --> R2
L -- "show me interactively" --> R3
L -- "what does this symbol mean?" --> R4
R1 & R2 & R3 & R4 -- improved article --> L
style seed fill:#f5f3ef,stroke:#e2ddd5
style generate fill:#e8f0f8,stroke:#3d6b9e
style learn fill:#fef6e8,stroke:#b87a1a
style refine fill:#e8f5ec,stroke:#27864a
```
The result combines:
- **Concept articles** — structured markdown files covering deep learning topics, interlinked with wiki-style `[[references]]`
- **Interactive teaching materials** — standalone HTML pages with step-by-step math walkthroughs, interactive simulations, and explorable visualizations
- **Reference materials** — glossary, notation guide, timeline, and key people for quick lookup
## Demo scope
**5 fully developed concept articles** demonstrate the article format and editorial approach:
| Article | Description |
|---------|-------------|
| [backpropagation](concepts/backpropagation.md) | The algorithm that makes training feasible — chain rule, BPTT, adjoint method, memory requirements |
| [gradient-descent](concepts/gradient-descent.md) | SGD, learning rates, momentum, Adam, mini-batch trade-offs |
| [loss-functions](concepts/loss-functions.md) | MSE, cross-entropy (with softmax explained), contrastive loss, autoregressive loss |
| [activation-functions](concepts/activation-functions.md) | ReLU through GELU, desirable properties, initialization interaction |
| [multilayer-perceptrons](concepts/multilayer-perceptrons.md) | The simplest deep architecture, universal approximation theorem |
The remaining ~90 concept articles are **stubs** — they retain their title, tagline, and connections section to preserve the linking structure, but their full content is not included in this demo.
## Interactive teaching materials
Each material is a self-contained HTML file — no build step, no server. Try them live via GitHub Pages or clone the repo and open locally.
### Backpropagation
- **[The Math of Backpropagation](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/backpropagation/index.html)** — partial derivatives refresher, chain rule with numbers, full forward/backward pass by hand, how autograd works. Includes synced network diagrams that highlight the active computation at each step.
- **[Vanishing & Exploding Gradients](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/backpropagation/vanishing-gradients.html)** — simulation with controls for depth (2-50 layers), activation function, initialization scheme, skip connections, and layer normalization. Watch gradient magnitudes respond in real time.
### Gradient Descent
- **[SGD by the Numbers](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/gradient-descent/sgd-math.html)** — worked mini-batch example: per-example gradients, averaging, parameter update, plus a multi-epoch training runner with adjustable learning rate and batch size.
- **[Gradient Descent — Interactive Learning](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/gradient-descent/index.html)** — 3D loss landscape exploration, learning rate comparison, momentum visualization, optimizer race (SGD vs Adam).
### Loss Functions
- **[Cross-Entropy & Softmax — Worked Math](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/loss-functions/cross-entropy-math.html)** — step-by-step: softmax computation, cross-entropy loss, gradients, parameter update, autoregressive cross-entropy with perplexity, and contrastive loss (CLIP-style).
- **[Cross-Entropy & Softmax Explorer](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/loss-functions/index.html)** — adjust logits with sliders, watch softmax probabilities and the -log(p) loss curve respond. Temperature control, gradient visualization, presets for edge cases.
### Activation Functions
- **[Activation Function Explorer](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/activation-functions/index.html)** — drag along curves to probe function values and derivatives, compare any two functions side by side, stack 1-30 layers to see how repeated activation crushes or preserves signals.
### Tensors
- **[Tensor Fundamentals](https://klauseduard.github.io/deep-learning-knowledge-base-demo/teach/tensors/index.html)** — interactive grid visualization of tensor shapes, indexing, reshaping, and broadcasting.
## Reference materials
| File | Contents |
|------|----------|
| [reference/notation.md](reference/notation.md) | Mathematical symbols and conventions ($\hat{y}$ = predicted, $\nabla$ = gradient, etc.) |
| [reference/glossary.md](reference/glossary.md) | 55+ term definitions (logits, softmax etymology, cross-entropy, epoch, backbone, etc.) |
| [reference/timeline.md](reference/timeline.md) | 42 milestones from McCulloch-Pitts (1943) to frontier models (2025) |
| [reference/key-people.md](reference/key-people.md) | ~40 researchers organized by contribution area |
## Structure
```
concepts/ 95 topic articles (5 full, 90 stubs in this demo)
teach/ Interactive HTML teaching materials
backpropagation/ 3 pages (math walkthrough, vanishing gradients)
gradient-descent/ 2 pages (SGD math, interactive landscapes)
loss-functions/ 2 pages (cross-entropy math, interactive explorer)
activation-functions/ 1 page (function explorer)
tensors/ 1 page (tensor fundamentals)
tools/ 5 framework/library articles (stubs)
reference/ 4 reference documents (notation, glossary, timeline, people)
_index.md Master index of all articles
_connections.md Cross-cutting themes and relationships
_sources.md Bibliography
_open_questions.md Unresolved questions across the field
```
## Design principles
**Concept articles** follow a consistent structure: tagline → overview → key details (intuition, etymology, technical subsections) → learn (links to teaching materials) → connections → sources → open questions. Inline `[[wiki-links]]` connect articles where concepts are naturally referenced.
**Teaching materials** use vanilla HTML/CSS/JS with CDN-hosted libraries (KaTeX for math, Plotly for charts). They follow a consistent design system (Source Serif for headings, Inter for body, JetBrains Mono for code/values). Two complementary formats:
- *Math walkthroughs* — step-through panels with concrete numbers at every step, synced diagrams
- *Interactive explorers* — parameter controls with immediate visual feedback, "what to notice" hints
**Reference materials** are look-up resources, not study materials. They support the concept articles without duplicating them.
## How to use
### Concept articles (Markdown)
The concept articles use `[[wiki-links]]` for cross-referencing and embedded Plotly charts for visualizations. To get the full experience:
1. **[Obsidian](https://obsidian.md/)** (free) — open this repository as an Obsidian vault. Wiki-links become clickable, the graph view shows connections between articles, and LaTeX math renders inline.
2. **[Obsidian Plotly plugin](https://github.com/Dmytro-Shulha/obsidian-plotly)** — install via Obsidian's community plugins to render the embedded charts in the loss-functions and activation-functions articles.
Without Obsidian, the articles are still readable as plain Markdown — `[[wiki-links]]` won't be clickable but the text is clear. The Plotly chart data is visible as YAML blocks.
### Teaching materials (HTML)
The `teach/` directory contains standalone HTML files. Just open them directly in a browser (Firefox, Chrome, etc.) — no server, no build step, no installation. They load fonts and math rendering from CDNs, so an internet connection is needed on first open.
### Reference materials (Markdown)
Plain Markdown with some LaTeX math notation. Readable in any Markdown viewer or text editor; best in Obsidian where the math renders.
## Built with
The knowledge base was built iteratively using [Claude Code](https://claude.ai/code) — article drafting, review, teaching material generation, and cross-linking were done in conversation.
Much of the repetitive structure (article templates, teaching material scaffolding, linting, visualization helpers) was encapsulated in a set of personal [Claude Code skills](https://docs.claude.com/en/docs/claude-code/skills) — for example `kb-seed` (generate article drafts from a topic list), `kb-teach` (scaffold interactive teaching materials), `kb-lint` (check article structure), and `kb-viz` (generate diagrams). These skills are not included in this repository — they're personal tooling tailored to one learner's workflow, and they'd need to be rewritten to match someone else's article conventions and design preferences. But the concept of per-domain skills is worth knowing about: a small set of invokable prompts (like `/kb-teach `) reduces the friction of consistent knowledge base work significantly.
## License
Content is provided for demonstration purposes. The teaching material HTML files can be opened directly in any modern browser.