Learning and the Unknown: Surveying Steps toward Open World Recognition

Abstract

As science attempts to close the gap between man and machine by building systems capable of learning, we must embrace the importance of the unknown. The ability to differentiate between known and unknown can be considered a critical element of any intelligent self-learning system. The ability to reject uncertain inputs has a very long history in machine learning, as does including a background or garbage class to account for inputs that are not of interest. This paper explains why neither of these is genuinely sufficient for handling unknown inputs – uncertain is not unknown, and unknowns need not appear to be uncertain to a learning system. The past decade has seen the formalization and development of many open set algorithms, which provably bound the risk from unknown classes. We summarize the state of the art, core ideas, and results and explain why, despite the efforts to date, the current techniques are genuinely insufficient for handling unknown inputs, especially for deep networks.

CAMV: Class Activation Mapping Value Towards Open Set Fine-Grained Recognition

“…The traditional mapping function is to use SoftMax, which is suitable for an end-to-end deep network and ensures efficient inference speed. However, due to the closedset attribute of the SoftMax function [16], SoftMax must produce a classification score, which may be high in the unknown category [11]. Recently, many methods [15], [16], [31] based on Extreme Value Theory fill in the blanks towards open set deep networks.…”

“…However, the SoftMax layer has translation invariance, which will confuse known categories and unknown categories through our analysis. For example, CAV 1 is [1, -1, -4, 5, 2] and CAV 2 is [11,9,6,15,12]. They have the same SoftMax score and the maximum probability score with 0.9338.…”

“…OSFGR needs to effectively recognize different known categories and reject unknown category. However, most FGVC methods [5]- [10] apply the structure of closed set recognition with the implicit closed world assumption [11]. For unknown categories, most FGVC systems only get the most-likely known categories.…”

“…The traditional method on the open set directly adopts the result of SoftMax to judge whether the input is an unknown category during inference time. However, the SoftMax function forces classifier to choose from known categories due to its closed-set attribute [11], [16]. For the SoftMax function, we study the process of feature mapping from the original feature to the final score and discover the translation invariance for the original feature.…”

Open Set Recognition Through Deep Neural Network Uncertainty: Does Out-of-Distribution Detection Require Generative Classifiers?

“…Access to these uncertainties comes with the promise of allowing to separate what a model is truly confident about through output variability. However, misclassification is not prevented and in a Bayesian approach uncertain inputs are not necesssarily unknown and vice versa unknowns do not necessarily appear as uncertain [3]. This has recently been observed on a large empirical scale [19] and figure 1 illustrates this challenge.…”