rejection系列3 OpenMax

paper: Towards Open Set Deep Networks. CVPR

Motivation

closed set recognition 天然的特性使得它必须选择一个类别作为预测对象。但是实际场景下， recognition system 必须学会 reject unknown/unseen classes 在 testing 阶段。

A key element of estimating the unknown probability is adapting Meta-Recognition concepts to the activation patterns in the penultimate layer of the network.

Introduction

probability/confidence scores. 比如通过对抗学习得到的 adversarial images. 作者在后面也提到了， threshold 实际上拒绝的不是 unknown, 而是 uncertain predictions.

OpenMax incorporates likelihood of the recognition system failure. This likelihood is used to estimate the probability for a given input belonging to an unknown class. For this estimation, we adapt the concept of Meta-Recognition[22, 32, 9] to deep networks. We use the scores from the penultimate layer of deep networks (the fully connected layer before SoftMax, e.g., FC8) to estimate if the input is “far” from known training data. We call scores in that layer the activation vector(AV).

A key insight in our opening deep networks is noting that “open space risk” should be measured in feature space rather than in pixel space.

We show that an extreme-value meta-recognition inspired distance normalization process on the overall activation patterns of the penultimate network layer provides a rejection probability for OpenMax normalization for unknown images, fooling images and even for many adversarial images.

Open set deep networks

Building on the concepts of open space risk, we seek to choose a layer (feature space) in which we can build a compact abating probability model that can be thresholded to limit open space risk.

multi-classes meta-recognition

. Prior work on meta-recognition used the final system scores, analyzed their distribution based on Extreme Value Theory (EVT) and found these distributions follow Weibull distribution.

from wikipedia:

It seeks to assess, from a given ordered sample of a given random variable, the probability of events that are more extreme than any previously observed.

We take the approach that the network values from penultimate layer (hereafter the Activation Vector (AV)), are not an independent per-class score estimate, but rather they provide a distribution of what classes are “related.”

Xie Pan

2018-12-11

2021-06-29