跳转至

ELVAE：基于证据学习的变分自编码器用于不确定性感知的生成

文章背景与核心概要

在深度生成模型领域，变分自编码器（VAE）通常通过概率隐空间表示来生成样本，但标准模型往往难以将“关于隐空间位置的不确定性”与“围绕该位置的变异性”区分开来。这限制了模型在需要主动管理风险和鲁棒性的场景中的应用。

为了弥补这一技术空白，本文提出了 ELVAE（Evidential Learning-Based Variational Autoencoder）。该方法构建了一种分层架构，其中每个隐空间坐标均由输入依赖的正态-逆伽马（normal-inverse-gamma）后验来控制。这种设计使得模型能够捕获显式的隐空间位置不确定性，并将其直接应用于生成阶段，而不是仅仅在推理后被动观察。实验表明，低不确定性锚点可用于生成更可靠的合成样本，而高不确定性锚点则可被主动利用于模型的压力测试，为不确定性感知的生成提供了一种实用的控制变量。

📌 摘要 (Summary)

Variational Autoencoders (VAEs) typically generate samples from probabilistic latent representations, but standard models struggle to separate uncertainty about the latent location from variability around it.

变分自编码器通常从概率隐空间表示中生成样本，但标准模型难以区分“关于隐空间位置的不确定性”与“围绕该位置的变异性”。

To bridge this gap, this paper introduces ELVAE (Evidential Learning-Based Variational Autoencoder). ELVAE formulates a hierarchical approach where each latent coordinate is governed by an input-dependent normal-inverse-gamma posterior. This allows the model to capture explicit latent-location uncertainty, which can be actively utilized during generation rather than merely observed after inference: * Low-uncertainty anchors are leveraged to produce more reliable synthetic samples. * High-uncertainty anchors can be intentionally exploited for stress-testing models.

为了弥合这一差距，本文推出了 ELVAE（基于证据学习的变分自编码器）。ELVAE 构建了一种分层方法，其中每个隐空间坐标由输入依赖的正态-逆伽马后验控制。这使得模型能够捕获显式的隐空间位置不确定性，该不确定性可以在生成过程中被积极利用，而不仅仅是在推理后进行报告： * 低不确定性锚点被用于产生更可靠的合成样本。 * 高不确定性锚点可以被刻意利用来进行模型压力测试。

📖 摘要原文 (Abstract)

Variational autoencoders generate samples from probabilistic latent representations but do not distinguish uncertainty about the latent location from variability around it. We formulate ELVAE, an evidential learning-based VAE in which each latent coordinate is governed by an input-dependent normal-inverse-gamma posterior. This hierarchy yields an explicit latent-location uncertainty that can be used during generation, not merely reported after inference: low-uncertainty anchors support more reliable synthetic samples, while high-uncertainty anchors can be deliberately exploited for stress testing.

变分自编码器从概率隐空间表示中生成样本，但并未区分关于隐空间位置的不确定性与围绕它的变异性。我们提出了 ELVAE，一种基于证据学习的 VAE，其中每个隐空间坐标由输入依赖的正态-逆伽马后验控制。这种分层结构产生了显式的隐空间位置不确定性，可用于生成过程中，而不仅仅是在推理后报告：低不确定性锚点支持更可靠的合成样本，而高不确定性锚点可以刻意用于压力测试。

The objective is an exact evidence lower bound, and we show that direct regularization of the full hierarchy is required, since the marginalized latent law alone cannot identify the uncertainty decomposition. In an MNIST generation pilot with a frozen external classifier, this uncertainty clearly stratified the semantic reliability of generated digits. A zero-displacement control revealed that most of the effect reflects how reliably an anchor can be re-generated, while a smaller but distinct component is attributable to uncertainty-scaled perturbation itself. The effect holds only under within-class uncertainty ranking, and its magnitude varies across seeds. These findings support the learned latent-location uncertainty as a practical control variable for uncertainty-aware generation, separating anchor reliability from perturbation-induced failure.

其优化目标是一个精确的证据下界（ELBO），并且我们证明需要对整个分层结构进行直接正则化，因为仅靠边缘化的隐空间规律无法识别不确定性分解。在一个结合了冻结外部分类器的 MNIST 生成试点中，这种不确定性清晰地分层了生成数字的语义可靠性。零位移对照实验表明，大部分效应反映了锚点能够被多可靠地重新生成，而一个较小但明显的成分则归因于不确定性缩放的扰动本身。该效应仅在类内不确定性排序下成立，且其幅度随随机种子而变化。这些发现支持了学习到的隐空间位置不确定性作为不确定感知生成的一个实用控制变量，从而将锚点可靠性与扰动引起的失效区分开来。

📋 文档详情 (Document Details)

Full-Text Links

全文链接