Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

B-cos Networks: Alignment is All We Need for Interpretability

About

We present a new direction for increasing the interpretability of deep neural networks (DNNs) by promoting weight-input alignment during training. For this, we propose to replace the linear transforms in DNNs by our B-cos transform. As we show, a sequence (network) of such transforms induces a single linear transform that faithfully summarises the full model computations. Moreover, the B-cos transform introduces alignment pressure on the weights during optimisation. As a result, those induced linear transforms become highly interpretable and align with task-relevant features. Importantly, the B-cos transform is designed to be compatible with existing architectures and we show that it can easily be integrated into common models such as VGGs, ResNets, InceptionNets, and DenseNets, whilst maintaining similar performance on ImageNet. The resulting explanations are of high visual quality and perform well under quantitative metrics for interpretability. Code available at https://www.github.com/moboehle/B-cos.

Moritz B\"ohle, Mario Fritz, Bernt Schiele• 2022

Related benchmarks

TaskDatasetResultRank
Explanation FaithfulnessImageNet 2015 (test)
AOPC0.717
22
SegmentationImageNet segmentation
Pixel Accuracy77.92
22
DGA Detection2020–2025 (test)
Accuracy94.8813
12
DGA DetectionDGA 2024 (test)
FPR9.8267
12
DGA DetectionDGA 2020 (test)
FPR3.964
12
DGA DetectionDGA 2021 (test)
False Positive Rate0.0382
12
DGA DetectionDGA 2022 (test)
False Positive Rate (FPR)4.1114
12
DGA DetectionDGA 2023 (test)
FPR4.8349
12
DGA DetectionDGA 2025 (test)
FPR10.8946
12
Attribution EvaluationHard MNIST
IoU0.465
8
Showing 10 of 10 rows

Other info

Follow for update