Research map / Architectures and efficiency / CNN design
PICNN: A Pathway towards Interpretable Convolutional Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PICNN, an interpretation pathway that groups late conv-layer filters into class-specific clusters via Bernoulli sampling from a learnable filter-class correspondence matrix, with a reparameterization trick for end-to-end training.
Problem
Interpretability of CNNs is hampered by filter-class entanglement: filters in late conv-layers encode many-to-many relations with image classes, and posthoc methods do not remove it.
Approach
A learnable filter-class probability matrix parameterizes Bernoulli sampling of a binary filter-cluster assignment; a novel reparameterization makes the non-differentiable sampling trainable, and a pseudo-label sampled from the discrimination output avoids label leakage. Filters are thus grouped into one-to-one class-specific clusters during training.
Key results
- On CIFAR-10 with ResNet-18, PICNN reaches ACC1 0.951 (STD 0.952) and ACC2 0.948 versus 0.588 for STD, with ACC3 lowered to 0.333 from 0.950.
- PICNN improves interpretability on ten architectures; on ResNet-50 ACC2 rises from 0.457 to 0.950 and MIS from 0.115 to 0.231.
- Bernoulli sampling outperforms Gumbel-Softmax on CIFAR-10 in ACC3 (0.333 vs 0.910) and MIS (0.243 vs 0.240), and PICNN adds at most about 5% computational overhead.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · CIFAR-10 | ACC2: 0.588 -> 0.948 | Helped |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · CIFAR-10 | ACC3: 0.950 -> 0.333 | Helped |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · CIFAR-10 | ACC1: 0.952 -> 0.951 | Hurt |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · PASCAL VOC Part | ACC1: 0.782 -> 0.800 | Helped |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · STL-10 | MIS: 0.148 -> 0.224 | Helped |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · CIFAR-100 | ACC1: 0.755 -> 0.724 | Hurt |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · CIFAR-100 | ACC2: 0.039 -> 0.707 | Helped |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · TinyImageNet | ACC1: 0.619 -> 0.555 | Hurt |
| Add the PICNN interpretation pathway to ResNet-18 | ResNet-18 · TinyImageNet | ACC3: 0.619 -> 0.082 | Helped |
| Use Bernoulli sampling instead of Gumbel-Softmax for filter-cluster assignment (ablation) | PICNN with ResNet-18 · CIFAR-10 | ACC3: 0.910 -> 0.333 | Helped |
| Use Bernoulli sampling instead of Gumbel-Softmax for filter-cluster assignment (ablation) | PICNN with ResNet-18 · CIFAR-10 | ACC2: 0.946 -> 0.948 | Helped |
| Use Bernoulli sampling instead of Gumbel-Softmax for filter-cluster assignment (ablation) | PICNN with ResNet-18 · CIFAR-10 | ACC1: 0.950 -> 0.951 | Helped |
Surprising findings
- PICNN does not work well with DenseNet-121 as backbone on PASCAL VOC Part, attributed to dense connectivity mixing low-level concepts into the target layer.
- Gumbel-Softmax, the natural differentiable alternative, gives nearly the same MIS and ACC2 as Bernoulli sampling but much worse ACC3.
Limitations
- Performance degrades for DenseNet-121 backbone on PASCAL VOC Part because the target layer mixes low-level concepts.
- On CIFAR-100 and TinyImageNet, ACC1 drops relative to STD (0.724 vs 0.755 and 0.555 vs 0.619).
Experimental setup
Datasets: CIFAR-10 STL-10 PASCAL VOC Part CIFAR-100 TinyImageNet ImageNet
Benchmarks: CIFAR-10 STL-10 PASCAL VOC Part CIFAR-100 TinyImageNet
Metrics: ACC1 ACC2 ACC3 MIS training time inference time
Baselines: STD (standard CNN) ICCNN CSG Gumbel-Softmax filter assignment
Builds on: Gumbel-Softmax Class-Specific Gate (CSG) Interpretable Compositional CNN (ICCNN) Grad-CAM
Closest papers by meaning
- Network Dissection: Quantifying Interpretability of Deep Visual Representations (CVPR 2017 · similarity 0.90)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.88)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.88)
- Interpretable Convolutional Neural Networks (CVPR 2018 · similarity 0.88)
- B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable (NeurIPS 2024 · similarity 0.87)
- NeSyFOLD: A Framework for Interpretable Image Classification (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.