AracNet: Revealing Debiasing Signals across Layers with Shallow Monitors

ECCV 2026 — Vito Paolo Pastore, Massimiliano Ciranni, Enzo Tartaglione, Vittorio Murino

Deep neural networks often show poor generalization when trained on biased datasets presenting spurious relations between samples and target labels. AracNet is an unsupervised debiasing framework that replaces the when to stop paradigm with a where to look one. By attaching one linear classifier (Shallow Monitor) to each block of a pre-trained network, AracNet automatically identifies the layer where bias is most separable and uses it to reweight samples during a self-mitigation phase — without requiring bias labels, annotated validation sets, or heuristic early stopping.

pip install torch torchvision tqdm numpy matplotlib seaborn wandb requests gdown

A CUDA-capable GPU is required. All experiments were run on an NVIDIA A100 (20 GB VRAM).

Datasets are downloaded automatically on first run under ./data/. Each run trains a seed-specific biased backbone (Phase 1) then runs the full AracNet debiasing pipeline (Phase 2). Results are logged via Weights & Biases and checkpoints are saved under ./hist/SIMPLE/<dataset>/.

bash Waterbirds.sh

Equivalent single-seed invocation:

python main_aracnet.py \
    --dataset waterbirds \
    --epochs_base_model 50 \
    --epochs_debiasing 50 \
    --seed 0 \
    --rho 95 \
    --train_body_aracnet

bash BFFHQ.sh

Equivalent single-seed invocation:

python main_aracnet.py \
    --dataset BFFHQ \
    --epochs_base_model 50 \
    --epochs_debiasing 70 \
    --seed 0 \
    --rho 99.5 \
    --train_body_aracnet

Note: omitting --train_body_aracnet skips Phase 1 and loads an existing backbone checkpoint for the given seed. The checkpoint must have been produced by a prior run with the same --seed, --dataset, and --rho.

Backbone and debiased model checkpoints are saved per-seed under:

./hist/SIMPLE/<dataset>/<rho>/<backbone>/
    aracnet-<backbone>-False_<dataset>_<rho>_seed<seed>-biased-final.pth   # biased backbone
    debiased_<dataset>_<rho>_seed<seed>.pth                                 # debiased model

@inproceedings{pastore2026aracnet,
  title={AracNet: Revealing Debiasing Signals Across Layers with Shallow Monitors},
  author={Pastore, Vito Paolo and Ciranni, Massimiliano and Tartaglione, Enzo and Murino, Vittorio},
  booktitle={European Conference on Computer Vision},
  pages={1--18},
  year={2026},
  organization={Springer}
}