Reducing Network Agnostophobia

By Akshay Raj Dhamija, Manuel Günther, Terrance E. Boult. 2018

383 citations recorded by the source corpus.

Agnostophobia, the fear of the unknown, can be experienced by deep learning engineers while applying their networks to real-world applications. Unfortunately, network behavior is not well defined for inputs far from a networks training set. In an uncontrolled environment, networks face many instances that are not of interest to them and have to be rejected in order to avoid a false positive. This problem has previously been tackled by researchers by either a) thresholding softmax, which by construction cannot return "none of the known classes", or b) using an additional background or garbage class. In this paper, we show that both of these approaches help, but are generally insufficient when previously unseen classes are encountered. We also introduce a new evaluation metric that focuses on comparing the performance of multiple approaches in scenarios where such unseen classes or unknowns are encountered. Our major contributions are simple yet effective Entropic Open-Set and Objectosphere losses that train networks using negative samples from some classes. These novel losses are designed to maximize entropy for unknown inputs while increasing separation in deep feature space by modifying magnitudes of known and unknown samples. Experiments on networks trained to classify classes from MNIST and CIFAR-10 show that our novel loss functions are significantly better at dealing with unknown inputs from datasets such as Devanagari, NotMNIST, CIFAR-100 and SVHN.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

RMM: Reinforced Memory Management for Class-Incremental Learning (2023, background): “, for curriculum learning [1] and open set recognition [2].”
A Wholistic View of Continual Learning with Deep Neural Networks: Forgotten Lessons and the Bridge to Active and Open World Learning (2020, background): “•Universum Inference [136] • Confidence Calibration [139] •Objectosphere Loss [141] • SCM [142] •Discrepancy Loss [140]”
Recent Advances in Open Set Recognition: A Survey (2018, methodology): “To address these problems, researchers have looked at different approaches [25],”
Opening up Open World Tracking (2021, background): “Object detection has recently been studied in open-set conditions [20, 51].”
Generalized ODIN: Detecting Out-of-Distribution Image Without Learning From Out-of-Distribution Data (2020, methodology): “[6, 14] push the idea of utilizing OoD data further by using a carefully chosen OoD dataset to regularize the learning of class posteriors so that OoD data have much lower conﬁdence than in-distribution.”
Learning Bounds for Open-Set Learning (2021, background): “These unknown classes can emerge unexpectedly and drastically weaken the performance of existing closed-set algorithms (de O. Cardoso et al., 2017; Dhamija et al., 2018; Perera et al., 2020).”
C2AE: Class Conditioned Auto-Encoder for Open-Set Recognition (2019, background): “However, these questions are relatively unexplored in the context of deep neural networks [38], [4], [24], [8], [7].”
Stream-51: Streaming Classification and Novelty Detection from Videos (2020, methodology): “The explicit equation for γ can be found in [9] and computes performance based on novelty detection capabilities, as well as correct classification of in-distribution samples.”
Adversarial Reciprocal Points Learning for Open Set Recognition (2021, methodology): “The Correct Classification Rate (CCR) is the fraction of samples where the correct class k has maximum probability and has a probability greater than d CCRðdÞ ¼ jfx 2 D k T ^ argmaxkP ðkjxÞ ¼ k̂ ^ P ðk̂jxÞ dg jDkT j : (21) The False Positive Rate (FPR) is the fraction of samples from unknown data DU”
Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models (2022, background): “Distribution shifts can occur naturally in the real world due to variations in the environment [33] or the encounter of unseen concepts [34].”
ViM: Out-Of-Distribution with Virtual-logit Matching (2022, background): “[6] regularize samples from extra background classes to have uniform logits and to have small feature norms.”
Semantically Coherent Out-of-Distribution Detection (2021, methodology): “The metric evaluates ID classification and OOD detection simultaneously and is formalized by Equation 3 in [12].”
Distinction Maximization Loss: Fast, Scalable, Turnkey, and Native Neural Networks Out-of-Distribution Detection simply by Replacing the SoftMax Loss (2019, methodology): “ may be integrated and beneﬁt both SoftMax and IsoMax losses, all experiments were performed without relying on outlier data [25], [26]. Similar arguments hold for background samples based approaches [44]. All datasets, models, training procedures, and metrics followed the baseline established in [6”
Generalized Out-of-Distribution Detection: A Survey (2021, neutral): “…above metrics evaluate the novelty detection and ID classiﬁcation capabilities independently, some works raise some evaluation criteria for joint evaluation, such as CCR@FPR x (Dhamija et al., 2018), which calculates the class-wise recall when a certain FPR equal to x (e.g.,10 − 1 ) is achieved.”
Isotropic Maximization Loss and Entropic Score: Fast, Accurate, Scalable, Unexposed, Turnkey, and Native Neural Networks Out-of-Distribution Detection. (2019, neutral): “Similar arguments hold for background samples based approaches (Dhamija et al., 2018).”
Detecting semantic anomalies (2019, background): “Some work in the area of open-set recognition has come close to a realistic setup: in [33], detection of novel categories is tested with a set of images corresponding to different classes that were discontinued in later versions of Imagenet, but later work [34] relapsed into treating very different ”
Self-Supervised Features Improve Open-World Learning (2021, background): “This over-specialization limits the ability to detect out-ofdistribution (unknown) classes because of their projection on top of known classes, see [16].”
Learning and the Unknown: Surveying Steps toward Open World Recognition (2019, background): “More recent work (Dhamija, G¨unther, and Boult 2018) seeks to train deep networks that handle unknowns by combining softmax with a new loss function which forces known unknown samples, i.e., background classes, to have a small feature magnitude.”
A Unifying Review of Deep and Shallow Anomaly Detection (2020, methodology): “, the use of spherical models in both AD [137], [155] and OOD [512], [514]) may help them to advance together.”