akankshanc/inception-v1-microscope

Inception V1 Microscope

An interactive microscope for exploring visual features learned by individual neurons in Inception V1.

[Open the interactive microscope](https://akankshanc-inception-v1-microscope.static.hf.space/index.html) · [Try a featured neuron](https://akankshanc-inception-v1-microscope.static.hf.space/index.html?layer=mixed4b&channel=233) · [View the dataset](https://huggingface.co/datasets/akankshanc/inception-v1-microscope-data)

The microscope combines two complementary views of each neuron:

1. 1.Activation maximization — a synthesized image optimized to strongly activate the selected neuron.
2. 2.Top dataset examples — the ten ImageNet examples producing the strongest recorded activations, shown alongside the corresponding triggering crops.

Together, these views provide evidence about the visual patterns a neuron may respond to. They should be treated as interpretability aids rather than definitive semantic labels.

Dataset scale

The current release covers 5,804 neurons across 12 Inception V1 layers.

Each ranked example pair contains one full image and one crop, giving 116,080 dataset-example images, in addition to 5,804 activation-maximization visualizations.

Using the microscope

• —Hover over a layer to preview its channel grid.
• —Click a layer to select and pin it.
• —Search channels by entering a channel number.
• —Click a channel to load its activation-maximization image and ten ranked dataset examples.
• —Use Previous and Next, or the left and right arrow keys, to move between neurons.
• —Use Random to sample a neuron from the complete microscope.
• —Use Copy link to create a direct, shareable URL for the selected neuron.
• —Select Paired, Full only, or Crops only to change the example view.
• —Click any image to enlarge it and access the original image URL.

The selected neuron is encoded in the URL:

?layer=mixed4a&channel=254

This makes individual neurons directly linkable and preserves browser Back and Forward navigation.

What the visualizations mean

Activation maximization

Activation maximization begins with a parameterized input image and optimizes it to increase the response of one selected channel. The resulting visualization is a hypothesis about patterns that strongly excite that neuron under the optimization procedure.

These images can contain optimization artifacts and should not be interpreted as natural photographs or as complete descriptions of a neuron's behavior.

Top dataset examples

For each neuron, the interface displays the ten highest-scoring examples found in the evaluated ImageNet collection. Each row contains:

• —rank within the selected neuron;
• —raw activation score;
• —source image identifier;
• —full source image; and
• —crop associated with the strongest activating region.

The examples are useful for checking whether features visible in the synthesized image also appear in natural images.

Activation scores

Activation scores are raw responses used to rank examples within a neuron. Their scale can vary substantially between layers and channels. A score from one layer should therefore not be interpreted as directly larger or stronger than a score from another layer.

Dataset organization

The companion dataset uses two configurations:

Activation-maximization fields:

model, layer, channel, neuron_id, original_filename, image

Dataset-example fields:

model, layer, channel, neuron_id, rank, activation_score,
source_image_id, full_image, crop_image

The name train is a Hugging Face split label used for storage; the activation-maximization records are visualizations, not a model-training set.

Implementation

This is a client-side Static Hugging Face Space built with HTML, CSS, and JavaScript. It performs no model inference in the browser. When a neuron is selected, the application requests only:

• —one activation-maximization record; and
• —ten ranked dataset-example records.

The records and image URLs are served through the Hugging Face Dataset Viewer API. Successful responses are cached during the browser session. Independent loading states, record-identity checks, and retry controls protect against stale responses and transient network or Dataset Server failures.

Reliability checks

Before this release, the published data and application were checked for:

• —expected row counts for every configuration and split;
• —correct first and last neurons in all 12 layers;
• —matching layer, channel, neuron ID, and ranks 1–10;
• —activation-image availability;
• —all FULL/CROP field references;
• —cross-layer Previous/Next navigation;
• —direct links and browser history;
• —random-neuron navigation;
• —rapid-selection race conditions;
• —image enlargement and keyboard dismissal; and
• —responsive desktop and narrow-screen layouts.

Limitations

• —Feature visualization is evidence about model behavior, not a definitive explanation or neuron label.
• —Activation-maximization results depend on the optimization objective, parameterization, regularization, and selected checkpoint.
• —Top examples describe the evaluated image collection and may not capture all situations that activate a neuron.
• —Crops can omit contextual information that contributes to the full-image response.
• —Raw activation scores should not be compared directly across layers.
• —Because the Space fetches public data at runtime, temporary Dataset Server or network failures can delay an image. The interface provides retry controls.

Data provenance and responsible use

The natural-image examples are derived from ImageNet and are provided for non-commercial research and educational interpretability work. ImageNet does not own the copyright in the underlying images; individual images may remain subject to their original copyright and applicable ImageNet access terms. This project does not relicense those source images.

Please review the ImageNet terms of access before downloading, redistributing, or repurposing the natural-image examples.

💻 Compute Contributions

Activation maximization and dataset-example generation across InceptionV1's layers were distributed among the following contributors:

References

• —Deng, J. et al. ImageNet: A Large-Scale Hierarchical Image Database. CVPR 2009. ImageNet
• —Olah, C. et al. The Building Blocks of Interpretability. Distill, 2018. Article
• —Olah, C. et al. Zoom In: An Introduction to Circuits. Distill, 2020. Article
• —OpenAI. OpenAI Microscope. Project

Citation

If this microscope is useful in your work, please cite the Space and dataset:

@misc{akankshanc_inception_v1_microscope_2026,
  author       = {Akanksha Devkar, Mayank Bhaskar, Joyce Jiang, Anu Adesina, Ayesha Imran, Almustapha Yusuf, Joana da Matta},
  title        = {Inception V1 Microscope},
  year         = {2026},
  howpublished = {Hugging Face Space},
  url          = {https://huggingface.co/spaces/akankshanc/inception-v1-microscope}
}

Links

• —Interactive Space
• —Direct application
• —Dataset repository