akankshanc/inception-v1-microscope-data

Inception V1 Microscope Data

This dataset powers the Inception V1 Microscope, an interactive interface for exploring visual features learned by individual neurons in Inception V1.

It combines two complementary interpretability views:

1. Activation maximization: one synthesized visualization optimized to strongly activate each neuron.
2. Top dataset examples: the ten ImageNet examples producing the strongest recorded activations for each neuron, paired with crops associated with the activating regions.

These visualizations are evidence about model behavior, not definitive semantic labels for neurons.

Dataset scale

The production release covers 5,804 neurons across 12 layers.

Layer	Channels	Activation visualizations	Ranked FULL/CROP pairs
conv2d0	64	64	640
conv2d1	64	64	640
conv2d2	192	192	1,920
mixed3a	256	256	2,560
mixed3b	480	480	4,800
mixed4a	508	508	5,080
mixed4b	512	512	5,120
mixed4c	512	512	5,120
mixed4d	528	528	5,280
mixed4e	832	832	8,320
mixed5a	832	832	8,320
mixed5b	1,024	1,024	10,240
Total	5,804	5,804	58,040

Each ranked pair contains two images, giving 116,080 natural-image and crop records, in addition to the 5,804 synthesized activation visualizations.

Configurations

activation_maximization

This configuration contains one row per neuron.

Field	Type	Description
model	string	Model identifier (inception_v1)
layer	string	Selected layer
channel	int32	Zero-indexed channel number
neuron_id	string	Stable layer/channel identifier
original_filename	string	Source visualization filename
image	image	Activation-maximization visualization

The configuration uses the split name train as a Hugging Face storage label. These records are visualizations and are not a model-training set.

dataset_examples

This configuration uses one split per layer and contains ten rows per neuron. Each row keeps the full source image and its crop together.

Field	Type	Description
model	string	Model identifier (inception_v1)
layer	string	Selected layer
channel	int32	Zero-indexed channel number
neuron_id	string	Stable layer/channel identifier
rank	int16	Rank from 1 through 10 within the neuron
activation_score	float32	Raw response used for within-neuron ranking
source_image_id	string	ImageNet source identifier
full_image	image	Full natural image
crop_image	image	Crop associated with the activating region

Example

from datasets import load_dataset

activation = load_dataset(
    "akankshanc/inception-v1-microscope-data",
    "activation_maximization",
    split="train",
)

mixed4a_examples = load_dataset(
    "akankshanc/inception-v1-microscope-data",
    "dataset_examples",
    split="mixed4a",
)

neuron = mixed4a_examples.filter(
    lambda row: row["channel"] == 254
).sort("rank")

print(neuron)

Methodology

Activation maximization

Each synthesized image was produced by optimizing a parameterized input to increase the response of one selected Inception V1 channel. The resulting image provides a visual hypothesis about patterns that strongly excite that neuron under the chosen optimization procedure.

Dataset examples

Natural images were scored for each neuron. The ten highest-scoring examples were retained and ordered by raw activation score. Each full image was paired with a precomputed crop associated with its strongly activating region.

Validation

The release was checked for:

• 5,804 activation-maximization rows;
• exactly ten ranked example pairs per neuron;
• expected row counts for all 12 layer splits;
• matching model, layer, channel, and neuron identifiers;
• complete rank sets from 1 through 10;
• presence of both FULL and CROP image fields;
• correct first and last records at every layer boundary; and
• image URL availability through the Hugging Face Dataset Viewer API.

Intended use

This dataset is intended for non-commercial research and educational work on:

• neural-network interpretability;
• feature visualization;
• qualitative analysis of convolutional representations;
• interpretability interfaces and teaching demonstrations; and
• comparisons between synthesized features and natural-image evidence.

Limitations

• A visualization is an interpretability aid, not a definitive neuron label or complete causal explanation.
• Activation-maximization results depend on the checkpoint, objective, parameterization, regularization, and optimization procedure.
• Top examples characterize the evaluated image collection and may not cover every pattern that activates a neuron.
• Crops can omit contextual information that contributes to the full-image response.
• Raw activation scores are suitable for ranking examples within a neuron but should not be compared directly across layers or channels.
• The natural-image examples inherit biases and coverage limitations from ImageNet.

Data provenance and terms

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 repository does not relicense those source images.

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

References

• Szegedy, C. et al. Going Deeper with Convolutions. CVPR 2015. Paper
• 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
• OpenAI. OpenAI Microscope. Project

Citation

@misc{devkar_inception_v1_microscope_data_2026,
  author       = {Akanksha Devkar},
  title        = {Inception V1 Microscope Data},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/akankshanc/inception-v1-microscope-data}
}