Research map / Label-efficient and robust learning / Recognition and long-tail

Network Inversion of Convolutional Neural Nets (Student Abstract)

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Self-explaining and visual explanation networks · 0 citations (OpenAlex)

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 a vector-matrix conditioned generator with a combined KL, cross-entropy, and cosine-similarity inversion loss to produce diverse inputs that a trained CNN classifier maps to specified labels.

Problem

Network inversion for interpretability: reconstructing diverse inputs that a trained classifier assigns to a desired output, since straightforward label-conditioned generation suffers from mode collapse and low diversity.

Approach

Replace simple label conditioning with n-dimensional softmaxed vectors and n×n one-hot matrices fed into a generator, and add a cosine-similarity loss between generated images' classifier features to encourage diversity across samples of the same class.

Key results

• Generates visually diverse samples for all 10 classes of MNIST, FashionMNIST, SVHN and CIFAR10 classifiers (qualitative only).
• Reports that higher weight on the cosine similarity loss increases both inter-class and intra-class diversity of generated samples.

Measured effects reported in the paper

Limitations

• Evaluation is purely qualitative visual inspection; no quantitative diversity or fidelity metrics.
• No baseline comparisons against prior network inversion methods are reported.
• Robustness-guided generation of training-like samples is described but not evaluated.

Experimental setup

Datasets: MNIST FashionMNIST SVHN CIFAR10

Benchmarks: MNIST FashionMNIST SVHN CIFAR10

Metrics: qualitative visual inspection of generated samples

Builds on: Generative adversarial conditioning (label embedding) Batch normalization Dropout

Closest papers by meaning

• Invertibility of Convolutional Generative Networks from Partial Measurements (NeurIPS 2018 · similarity 0.87)
• i-RevNet: Deep Invertible Networks (ICLR 2018 · similarity 0.87)
• Network Deconvolution (ICLR 2020 · similarity 0.86)
• Inverting Visual Representations With Convolutional Networks (CVPR 2016 · similarity 0.86)
• InversionView: A General-Purpose Method for Reading Information from Neural Activations (NeurIPS 2024 · similarity 0.84)
• Expand-and-Cluster: Parameter Recovery of Neural Networks (ICML 2024 · similarity 0.84)