Matgc04/neuralatlas-attributions-efficientnet_b4

Neural Atlas attributions — efficientnet_b4 on imagenet-pico

Precomputed attribution maps and faithfulness metrics for the torchvision efficientnet_b4 model (default pretrained weights, no fine-tuning) on imagenet-pico, a 3000-image subset of ImageNet-1k with three images for each of the 1000 classes.

This repository is part of Neural Atlas, a web tool for comparing attribution methods across vision architectures on the same image, developed as an undergraduate thesis at the Facultad de Informática, Universidad Nacional de La Plata (Argentina). The viewer reads these files directly; they can also be used on their own, without the tool.

• Code: https://github.com/midusi/NeuralAtlas
• Author: Matías Gómez Carrique
• Advisors: Dr. Franco Ronchetti, Dr. Facundo Manuel Quiroga, Mg. Oscar Agustín Stanchi

Contents

Images	3000 (1000 classes × 3)
Attribution configurations	17
Attribution maps	50,994
Top-1 accuracy on imagenet-pico	0.7820 (2346/3000)
Macro F1	0.7712

Methods: CB-RISE, Deconvolution, DeepLift, GradientShap, GuidedBackprop, GuidedGradCam, InputXGradient, IntegratedGradients, KernelShap (KMeans), KernelShap (SLIC), LayerGradCam, LayerIntegratedGradients, Lime (KMeans), Lime (SLIC), Occlusion, RISE, Saliency.

LIME and KernelSHAP appear twice, once with SLIC superpixels and once with a k-means segmentation (RGB + scaled coordinates), both with 32 segments.

Layout

images/imagenet-pico/<model>__<dataset>__<class_id>__<image_idx>__<method>.avif
runs/efficientnet_b4/imagenet-pico/summary.json   # run-level classification metrics and map encoding
runs/efficientnet_b4/imagenet-pico/images.json    # one entry per image

Each entry in images.json holds the class id, the image index within the class, the prediction (predicted_class_id, confidence), the path of every attribution map (outputs), the faithfulness metrics per method (interpretability_metrics) and the methods that failed on that image (attribution_failures).

Map encoding

Maps are 224 × 224, 8-bit grayscale AVIF. Each attribution tensor is reduced by summing over channels and taking the absolute value, then normalized with Captum's cumulative-sum threshold (the top 2 % of attribution mass is clipped) and scaled to 0–255. The colormap is applied by the viewer. The sign of the attribution is therefore not preserved in the images; the faithfulness metrics were computed on the raw signed tensors before encoding.

Faithfulness metrics

Per image and method, interpretability_metrics contains four scores, all computed on the logit of the true class. Higher is better in all of them.

• lif — least-relevant-first (LeRF) deletion of single pixels (black fill, 100 steps); area under the logit curve.
• segment — least-attributed-first deletion of 25 fixed k-means (LABXY) superpixels per image (black fill); unnormalized area under the logit curve.
• morph — MeS: the binarized map (τ = 0.5) is eroded from the border inward with a 3 × 3 element and the removed area is filled with a Gaussian blur (σ = 10); area under the logit curve.
• fidelity — 1 − INFD(a)/INFD(0) over 25 perturbations (Gaussian noise with σ = 0.2 for local methods, a black 56 × 56 square for global ones).
• fidelity_superpixel — LIME and KernelSHAP only: the same score, removing one whole superpixel of the method's own segmentation instead of a square.

lif, segment and morph are not normalized by the intact-image score, so they are meant for comparing methods on the same image and model, not across images. The exact definitions and the differences from their original formulations are documented in the thesis.

Known failures

6 maps are missing because SLIC returned a single segment for the image, which leaves LIME and KernelSHAP with nothing to estimate (insufficient_features):

class_id	image_idx	method
539	2	KernelShap (SLIC)
539	2	Lime (SLIC)
582	0	KernelShap (SLIC)
582	0	Lime (SLIC)
958	2	KernelShap (SLIC)
958	2	Lime (SLIC)

Source images

imagenet-pico was sampled from ImageNet-mini (a redistribution of the ImageNet-1k validation split) with seed 0. Where a class had fewer than three validation images, it was completed with training images of the same class (71 images across 64 classes).

License

The attribution maps, predictions, metrics and metadata in this repository are released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0).

This license does not grant any rights over the source images. ImageNet does not own the copyright of those images and restricts access to non-commercial research and educational use; the images remain subject to the ImageNet terms of access and to the rights of their respective owners.

Other models

• neuralatlas-attributions-alexnet
• neuralatlas-attributions-resnet18
• neuralatlas-attributions-resnet101
• neuralatlas-attributions-efficientnet_b0
• neuralatlas-attributions-inception_v3
• neuralatlas-attributions-mobilenet_v2
• neuralatlas-attributions-convnext_tiny