[Paper Review] Conditional Gaussian Distribution Learning for Open Set Recognition

Xin Sun, Zhenning Yang+3 more|arXiv (Cornell University)|Mar 19, 2020

Domain Adaptation and Few-Shot Learning39 references37 citations

TL;DR

This paper introduces CGDL, a conditional Gaussian distribution learning framework with a probabilistic ladder network to simultaneously detect unknown samples and classify known ones in open set recognition, achieving state-of-the-art results.

ABSTRACT

Deep neural networks have achieved state-of-the-art performance in a wide range of recognition/classification tasks. However, when applying deep learning to real-world applications, there are still multiple challenges. A typical challenge is that unknown samples may be fed into the system during the testing phase and traditional deep neural networks will wrongly recognize the unknown sample as one of the known classes. Open set recognition is a potential solution to overcome this problem, where the open set classifier should have the ability to reject unknown samples as well as maintain high classification accuracy on known classes. The variational auto-encoder (VAE) is a popular model to detect unknowns, but it cannot provide discriminative representations for known classification. In this paper, we propose a novel method, Conditional Gaussian Distribution Learning (CGDL), for open set recognition. In addition to detecting unknown samples, this method can also classify known samples by forcing different latent features to approximate different Gaussian models. Meanwhile, to avoid information hidden in the input vanishing in the middle layers, we also adopt the probabilistic ladder architecture to extract high-level abstract features. Experiments on several standard image datasets reveal that the proposed method significantly outperforms the baseline method and achieves new state-of-the-art results.

Motivation & Objective

• Develop an end-to-end open set recognition method that jointly detects unknown samples and classifies known classes.
• Learn class-conditional latent representations by forcing posterior latents to approximate class-specific Gaussian models.
• Preserve high-level abstract features through a probabilistic ladder architecture to improve discrimination in latent space.
• Leverage reconstruction information alongside latent distributions to enhance unknown detection.
• Demonstrate state-of-the-art performance on standard image datasets (MNIST, SVHN, CIFAR variants).

Proposed method

• Learn conditional posterior distributions qφ(z|x,k) that approximate class-specific Gaussian priors pθ^(k)(z) = N(z; μk, I).
• Use a fully-connected layer to map one-hot class labels to latent-space means μk.
• Adopt a probabilistic ladder network in both encoder and decoder to capture high-level abstract features and enable information flow between layers.
• Train with a joint loss: reconstruction loss Lr, KL-divergence LKL (including middle-layer KL terms), and classification loss Lc; total loss L = - (Lr + β LKL + λ Lc).
• During testing, model per-class Gaussians f_k(z) = N(z; mk, σk^2) from latent representations of correctly classified samples and use reconstruction errors as an additional unknown-detection cue.
• Detect unknowns by combining CGD-based likelihoods with reconstruction-error thresholds τl and τr as described in the testing algorithm.

Experimental results

Research questions

• RQ1 Can conditional Gaussian latent priors enable simultaneous open-set detection and closed-set classification within an end-to-end framework?
• RQ2 Does incorporating a probabilistic ladder architecture help preserve high-level features and improve open-set performance over traditional VAEs?
• RQ3 How effective is combining latent-model-based unknown detection with reconstruction-based cues in CGDL?
• RQ4 What is the impact of learning class-conditioned latent distributions on closed-set accuracy compared to standard CNN baselines?
• RQ5 How does CGDL perform on standard OSR benchmarks across MNIST, SVHN, and CIFAR datasets?

Key findings

• CGDL achieves state-of-the-art open-set recognition performance on several standard image datasets.
• A probabilistic ladder architecture improves latent representation quality and open-set scores compared to plain CVAEs.
• Conditional Gaussian distributions enable discriminative latent representations for known classes while enabling unknown detection.
• Adding reconstruction-error cues to the CGD-based detector further improves performance in ablation studies.
• On Table 2, CGDL (Ours) achieves 0.837±0.055 on MNIST, 0.776±0.040 on SVHN, 0.655±0.023 on CIFAR10, 0.760±0.024 on CIFAR+10, and 0.695±0.013 on CIFAR+50, outperforming several baselines.
• CGDL also shows strong results on MNIST with Omniglot/Noise outliers and on CIFAR-10 with outliers, achieving the best macro-F1 scores among listed methods.