Abstract:Semi-supervised learning techniques utilize both labeled and unlabeled images to enhance classification performance in scenarios where labeled images are limited. However, challenges such as integrating unlabeled images with incorrect pseudo-labels, determining appropriate thresholds for the pseudo-labels, and label prediction fluctuations on low-confidence unlabeled images, hinder the effectiveness of existing methods. This research introduces a novel framework named Interpolation Consistency for Bad Generative Adversarial Networks (IC-BGAN) that utilizes a new loss function. The proposed model combines bad adversarial training, fusion techniques, and regularization to address the limitations of semi-supervised learning. IC-BGAN creates three types of image augmentations and label consistency regularization in interpolation of bad fake images, real and bad fake images, and unlabeled images. It demonstrates linear interpolation behavior, reducing fluctuations in predictions, improving stability, and facilitating the identification of decision boundaries in low-density areas. The regularization techniques boost the discriminative capability of the classifier and discriminator, and send a better signal to the bad generator. This improves the generalization and the generation of diverse inter-class fake images as support vectors with information near the true decision boundary, which helps to correct the pseudo-labeling of unlabeled images. The proposed approach achieves notable improvements in error rate from 2.87 to 1.47 on the Modified National Institute of Standards and Technology (MNIST) dataset, 3.59 to 3.13 on the Street View House Numbers (SVHN) dataset, and 12.13 to 9.59 on the Canadian Institute for Advanced Research, 10 classes (CIFAR-10) dataset using 1000 labeled training images. Additionally, it reduces the error rate from 22.11 to 18.40 on the CINIC-10 dataset when using 700 labeled images per class. The experiments demonstrate the IC-BGAN framework outperforms existing semi-supervised methods, providing a more accurate classification solution with smoother class label estimates, especially for low-confidence unlabeled images.

Interpolation consistency training for semi-supervised learning

Video Shadow Detection via Spatio-Temporal Interpolation Consistency Training

Boosting Interpolation Consistency with Difference and Adaptive Threshold

Interpolated Adversarial Training: Achieving robust neural networks without sacrificing too much accuracy

A novel interpolation consistency for bad generative adversarial networks (IC-BGAN)

Deep Network Interpolation for Continuous Imagery Effect Transition

In-Context Translation: Towards Unifying Image Recognition, Processing, and Generation

Variation Learning Guided Convolutional Network For Image Interpolation

OCAI: Improving Optical Flow Estimation by Occlusion and Consistency Aware Interpolation

Interpolation for Robust Learning: Data Augmentation on Wasserstein Geodesics

A PAC-Bayesian Perspective on the Interpolating Information Criterion

AID: Attention Interpolation of Text-to-Image Diffusion

Interpolation-Based Contrastive Learning for Few-Label Semi-Supervised Learning

Unsupervised Video Interpolation Using Cycle Consistency

Interpolating neural network: A novel unification of machine learning and interpolation theory

Improved CNN-based Learning of Interpolation Filters for Low-Complexity Inter Prediction in Video Coding

Improving Shift Invariance in Convolutional Neural Networks with Translation Invariant Polyphase Sampling

TITAN: Bringing The Deep Image Prior to Implicit Representations

Revisiting Latent-Space Interpolation via a Quantitative Evaluation Framework

Deep Pyramid Variation Learning for Image Interpolation.

Interpolation variable rate image compression