IA-001 / 2018
The Building Blocks of Interpretability is spatial attributionColour laid over a picture showing which parts of it the model was reacting to.Attribution values rendered over the positions of an input image, showing which regions drove the output. A saliency map is the most common instance.glossary and an activation gridLots of tiny pictures in a grid, each showing what the model saw at that spot.Many small images laid out in a grid, each one standing for what a part of the model responded to at that position.glossary showing featuresA single thing the model has learned to recognise, like a place, a tone of voice, or a kind of mistake.Individual learned directions or concepts inside the model.glossary and attributionWorking out which inputs were responsible for an output.Assigning responsibility for an output to particular inputs or components.glossary, about feature-interpretationWorking out what one small piece of the model has learned to recognise.What an individual learned feature means.glossary.
- ID
- IA-001
- Name
- The Building Blocks of Interpretability
- Artefact type
- Interactive article
- Visual form
- Spatial attributionColour laid over a picture showing which parts of it the model was reacting to.Attribution values rendered over the positions of an input image, showing which regions drove the output. A saliency map is the most common instance.glossaryActivation gridLots of tiny pictures in a grid, each showing what the model saw at that spot.Many small images laid out in a grid, each one standing for what a part of the model responded to at that position.glossary
- Makes visible
- FeaturesA single thing the model has learned to recognise, like a place, a tone of voice, or a kind of mistake.Individual learned directions or concepts inside the model.glossaryAttributionHow much each bit of the input pushed the model towards its answer.How much each part of an input pushed the model towards its output. Distinct from a feature, which is what the model represents, and from an activation, which is how hard a component is firing.glossary
- Role of the visual
- Finding
- Runnable by a visitor
- Pre-set examples only
- Date published
- 2018-03-06
- Year
- 2018
- Link status
- Live
- Built by (people)
- Chris Olah; Arvind Satyanarayan; Ian Johnson; Shan Carter; Ludwig Schubert; Katherine Ye; Alexander Mordvintsev
- Organisation
- Google Brain
- Venue / published in
- Distill
- Sector
- Frontier lab + academia
- Country / region
- US
- Open source
- Yes
- Code repo
- https://github.com/tensorflow/lucid
- Method / technique
- Feature visualisation composed with attribution; semantic dictionaries; activation grids; spatial attribution; neuron groups
- Model(s) studied
- GoogLeNet / InceptionV1 (vision)
- Access required
- Open weights
- What it visualises
- Which neurons fire where in an image, what each one detects, and how much each contributes to the classification
- Interaction affordances
- Hover to inspect; toggle between attribution modes; explore layer by layer; composable interface demos
- Pragmatic vs Basic science
- Basic science
- Reverse-eng vs Concept-based
- Reverse-engineering
- Observational vs Interventional
- Observational
- Intended audience
- Researchers; Practitioners
- Description (card)
- The founding text for interactive interpretability. Argues that interpretability techniques studied in isolation are far weaker than the interfaces you get by composing them, and demonstrates this with a set of live, hoverable interfaces over an image classifier.
- Why it matters
- Established that the interface IS the contribution — the template every entry in this ledger inherits from.
- Visual / design notes
- Distill house style: generous whitespace, inline hoverable figures, colour-coded attribution overlays. Still the visual benchmark 8 years on.
- Tags
- modality:visionImages. Models that look at pictures.Models that process static images.glossarymethod:feature-visualisationGenerating an image that shows what a part of the model responds to most strongly.Synthesising an input that maximally activates a chosen component.glossarymethod:attributionWorking out which inputs were responsible for an output.Assigning responsibility for an output to particular inputs or components.glossaryphenomenon:feature-interpretationWorking out what one small piece of the model has learned to recognise.What an individual learned feature means.glossary
- Thumbnail URL
- https://distill.pub/2018/building-blocks/thumbnail.jpg
- Citation
- Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K. and Mordvintsev, A., 2018. The Building Blocks of Interpretability. Distill. DOI 10.23915/distill.00010
- Related entries
- IA-002
- Confidence
- High
- Source of info
- Read in full (article + Distill metadata), 2026-08-06
- Date added
- 2026-08-06
- Added by
- Claude
- Notes
- Lucid repo link is the associated tooling, not the article source — verify before publishing. Classified Observational: the composed interfaces use gradient attribution, but the model is not altered.