Learning path instead of animation soup
The dense lab keeps a tabular learning path aligned to the active replay frame so learners can track what stage is done, what is current, and what comes next.
Built for learning by inspection
See the network think, fail, recover, and update.
Neural Network Desktop Tutor is a standalone PyQt app for teaching neural networks through replay, targeted inspection, ablation, and family-specific labs. It is not a slide deck about backpropagation. It is a lab where learners can watch gradients move, click a node, and ask what changed and why.
Dense replay is fully live today. CNN, LSTM, Seq2Seq, and Transformer each have dedicated family labs.
What the learner sees
A replayable network scene, formula pane, causal inspector, learning path, and selected-neuron evidence all in one view.
Replay-first
Forward pass, error, backpropagation, update preview, and weight commit stay visible as a sequence.
Click-driven
Nodes, edges, and hidden layers populate `Selection`, `Change`, and `Why` views immediately.
Ablation-backed
Layer and neuron importance can be tested rather than guessed from narrative.
Family-scoped
Dense, CNN, recurrent, sequence, attention, and transformer workflows live in separate labs.
Made for the gap between textbooks and black-box demos.
Most neural-network teaching tools stop at intuition theater: arrows, gloss, and static diagrams. This one treats the learner like an engineer. It exposes the actual training sequence, the local evidence behind a parameter update, and the consequences of interventions on the same screen.
Every click turns into evidence
Selection is first-class. Clicking an edge or neuron updates inspector tables with identity, deltas, gradients, contribution, and the strongest inbound or outbound drivers.
Preset the lesson, then let the learner roam
Microscope, teaching presets, stability controls, interventions, and comparisons let you tune the lesson without freezing the app into a fixed script.
One product, multiple model families.
The app keeps one active family at a time so the user is not juggling unrelated controls. Each family inherits a common visual shell and adds its own data type, inspection surface, and comparison language.
The core teaching lab
Tabular datasets, replay, challenge checkpoints, focus modes, interventions, and post-training ablation.
- Interactive network scene with trace overlays
- Formula and signal story tied to the live frame
- Selection, Change, and Why inspector tabs
Filter and feature-map inspection
Tiny image datasets, learned replay, kernel selection, feature maps, and filter ablation.
- Within-family comparison presets
- Replay-backed training queue with caching
- Visual shell shared with the dense lab
Gate and memory timelines
Sequence datasets with hidden-state replay, timestep and unit ablation, and recurrent comparison presets.
- Single-sequence family-scoped workflow
- Train replay off the main UI thread
- Gate-focused inspection surfaces
Encoder-decoder flow that can be replayed
Source-target tasks, learned alignment views, decoder-state inspection, and token ablations.
- Embedded reusable attention surface
- Replay-backed comparison presets
- Family-specific datasets only
The reusable alignment explorer
Standalone retrieval and alignment lab for head, token, and key-slot reasoning.
- Live attention maps
- Head and token ablation preview
- Shared surface reused in sequence labs
Residual stream, blocks, heads, and replay
Prompt-output language-model tasks with learned replay, token/head/block ablation, and comparison presets.
- Residual stream and block views
- Replay-backed family comparison queue
- Shared attention tooling plus transformer-specific state
Real interface, not mockup theater.
These captures come from the current app state, including the manual verification matrix used to check dense-lab readability and interaction quality.
Microscope preset
A tiny network built for slow, inspectable updates
The microscope flow strips the network down so one selected weight can carry a full local story through the replay, inspector, and explanation panes.
62 passing tests
The desktop workflow is covered by the local regression suite, including replay, family pages, and comparison paths.
Manual UI matrix
Empty pre-train state, microscope preset, and a medium dense network were manually captured and audited for overlap and selection clarity.
Family-specific training
Dense, CNN, LSTM, Seq2Seq, and Transformer all train inside their own visual labs instead of borrowing dense controls.
What the learner actually does.
The product loop is simple: choose the family, load a scenario, train, scrub replay, click a part of the model, then verify hypotheses with ablation or comparison.
Start with a lesson
Use the start page, microscope preset, or a teaching scenario to constrain the problem before training.
Watch replay
Replay automatically advances through the learning path instead of hiding the training sequence behind a final metric.
Click the model
Selection pins one weight, neuron, or layer into the right-hand inspector so the lesson becomes local and testable.
Confirm with interventions
Use ablations, comparisons, or scenario shifts to check whether the story still holds once the model is perturbed.
Run it locally in minutes.
This is a desktop-first project. No cloud account, no browser dependency, and no remote dataset required to start exploring the labs.
Quick start
Install the base app and launch the desktop tutor.
UV_CACHE_DIR=.uv-cache uv sync --extra dev
UV_CACHE_DIR=.uv-cache uv run neural-network-desktop-tutorOptional torch backend
If you want the optional PyTorch path for supported trainers, add the `torch` extra first.
UV_CACHE_DIR=.uv-cache uv sync --extra dev --extra torch- Python `>=3.12,<3.15`
- Base dependencies: PyQt6, NumPy, Pandas, Matplotlib, platformdirs
- All generated artifacts stay local and visible from the UI
Teach models like systems, not magic.
Neural Network Desktop Tutor is for classrooms, self-study, demos, and internal teaching sessions where “trust me, gradients flow backward” is not enough. The point is not just to show a network. The point is to let a learner inspect the reasons a network changed.