PICNN Pathway: Interpretable CNN Architecture

• The PICNN pathway is a CNN modification that enforces class-specific filter clusters to decouple class information, enhancing model transparency.
• It employs a learnable correspondence matrix with differentiable Bernoulli sampling to ensure end-to-end gradient flow and robust training.
• Empirical evaluations on benchmarks like CIFAR-100 demonstrate improved interpretability metrics and maintained classification accuracy.

The PICNN Pathway refers to a structural modification in deep convolutional networks that enforces an explicit, class-specific filter clustering in late convolutional layers, thereby enabling intrinsic interpretability of CNN decisions without major tradeoffs in classification accuracy. Originating in the work "PICNN: A Pathway towards Interpretable Convolutional Neural Networks" (Guo et al., 2023), the pathway resolves filter–class entanglement by probabilistically assigning filters to unique class-specific clusters using a learnable correspondence matrix and differentiable Bernoulli sampling. The resulting architecture supports a dual-path training regime enforcing both discriminative performance and interpretability, which is validated across various architectures and benchmarks.

1. Motivation and Theoretical Framework

Conventional CNNs achieve high discriminative power, but their internal representations distribute class information across late-layer filters in an entangled superposition, leading to poor interpretability. This prevents a one-to-one or one-to-few mapping between high-level semantic concepts (classes) and network components (filters), complicating model attribution and undermining trust in safety-critical deployments. The PICNN pathway addresses this by enforcing that each filter subgroup (cluster) in a late convolutional layer is mapped to a single class, effectively decoupling class-specific information at the feature-map level (Guo et al., 2023).

2. PICNN Pathway Architecture

The core architectural adaptation is the introduction of a learnable, real-valued filter–class correspondence matrix $P \in \mathbb{R}^{K \times N}$, with $K$ as the number of classes and $N$ as the number of filters. Each $p_{y,i}$ parameterizes the probability that filter $i$ is assigned to the cluster for class $y$. At inference/training time, binary assignment matrices $B \in \{0,1\}^{K \times N}$ are generated by

$b_{y,i} \sim \mathrm{Bernoulli}(p_{y,i})$

where $b_{y,i}=1$ indicates activation of filter $i$ for class $K$0.

For each input sample of (possibly pseudo-)class $K$1, feature maps $K$2 are extracted and masked via

$K$3

so that only the filters assigned to class $K$4 are active in the interpretation pathway. This dual-path framework is depicted below.

Pathway	Filters Used	Class Prediction
Discriminative	All $K$5 filters	$K$6
Interpretation	Only class-$K$7 cluster	$K$8

3. Bernoulli Sampling and Reparameterization

Direct Bernoulli sampling precludes gradient-based learning due to non-differentiability. PICNN employs a novel uniform-noise-based reparameterization for differentiable assignment: $K$9 This construction ensures that $N$0 while maintaining explicit dependence on $N$1, thus enabling end-to-end gradient flow (Guo et al., 2023).

To prevent trivial overfitting—where the network trivially indexes $N$2 by the true class label—the pathway substitutes the true label $N$3 with a pseudo-label $N$4 sampled from the model’s current softmax distribution $N$5, using an analogous uniform-noise reparameterization for categorical samples.

4. Objective Function and Dual-Path Optimization

The PICNN pathway enforces interpretability via a composite loss: $N$6 where $N$7 is the cross-entropy loss, $N$8 denotes the output of the discrimination pathway (full-filter classification), $N$9 is the output from the interpretation pathway (cluster-masked classification), and $p_{y,i}$0 balances the two objectives. The discrimination pathway preserves task accuracy, while the interpretation pathway drives cluster-class alignment.

5. Empirical Methodology

Datasets and Architectures

Evaluation encompasses CIFAR-10, STL-10, PASCAL VOC Part, CIFAR-100, and TinyImageNet, across ten architectural backbones including VGG-11, AlexNet, ResNet-18, DenseNet-121, MobileNetV2, EfficientNet-B0, ResNet-50, ResNet-152, Wide-ResNet, and ViT-b-12 (Guo et al., 2023).

Metrics

• ACC1: Discrimination pathway accuracy using all filters (higher is better)
• ACC2: Interpretation pathway accuracy using only the class cluster (higher is better)
• ACC3: Accuracy using all filters except the class-specific cluster (lower is better; <random indicates effective class separation)
• MIS: Mean mutual information between each filter and class labels (higher reflects better alignment)

Key Results

Method	ACC1	ACC2	ACC3	MIS
STD	.952	.588	.950	.191
PICNN	.951	.948	.333	.243

On CIFAR-100, ACC2 rises from ≈0.04 (STD) to ≈0.71 (PICNN), ACC3 drops from 0.753 to 0.131, and MIS improves from 0.020 to 0.026. Robustness to filter–class ratio ($p_{y,i}$1) and $p_{y,i}$2 hyperparameters is empirically supported.

6. Ablation and Analysis

Ablation confirms that the proposed uniform-noise reparameterization outperforms alternatives such as Gumbel-softmax, particularly in suppressing ACC3 and enhancing MIS. The pseudo-label mechanism precludes trivial solutions. DenseNet-121 exhibits comparatively weaker clustering due to dense skip-connections, which retain low-level feature mixing even in late convolutional layers. Performance remains stable for filter–class ratios in [4,25] and $p_{y,i}$3 in 0.5,10.

7. Limitations and Future Directions

Present limitations include diminished interpretability on architectures with extensive cross-layer skip connectivity. Extensions involve applying the PICNN pathway to multi-label and segmentation contexts, exploring more expressive reparameterization strategies for discrete sampling, and integrating structured class priors or domain-specific constraints to further enhance class–cluster disentanglement.

The PICNN pathway thus provides a direct architectural solution to the interpretability bottleneck in CNNs, validated across architectures and datasets and opening a broad avenue for transparent discriminative models (Guo et al., 2023).