Abstracts Track 2021

Area 1 - Theory and Methods

Area 1 - Theory and Methods

Nr: 3
Title:

Adversarial Minority-class Re-sampling for Imbalanced Sequence Classification
Authors:

Shin Ando, Yusuke Hatae, Muhammad F. Fadjrimiratno, Qingpu Yang, Yuanyuan Li, Tetsu Matsukawa and Einoshin Suzuki
Abstract: Class imbalance is a pervasive problem in applications of classification models, including deep neural networks. Standard countermeasures, such as re-sampling, add emphasis on learning from the relatively smaller minority class. However with a discrepancy in the scale required for deep learning, such a strategy can induce over-fitting. In this paper, we propose an adversarial learning framework for re-sampling synthetic minority class samples in class-imbalanced, sequence classification problems. We train the generator network to produce synthetic feature vectors from stochastically modified sequences, which the classifier network is likely to predict as those from the minority class. The classifier network, on the other hand, is trained to discriminate the synthetic vectors from those of the minority class, and is prevented from relying on few limited features so as not to make the deception by the generator easier. We further attempt to gain robustness by producing a diverse set of synthetic feature vectors using both the minority and the majority class sequences as input to the generator. We present proof-of-concept experiments with classification of texts and caption sequences. The results show that the proposed framework can substantially improve the recall for the minority class and the overall retrieval performance as well.
Nr: 4
Title:

Least Action Classifier with Dynamically Configured Structure
Authors:

Roman Malashin and Arina Boiko
Abstract: Real objects follow trajectories that correspond to the minimum of a functional incorporating two components. In physics this law is known as the least action principle. Accordingly, we propose image analysis systems that unfold a computational graph with the “trajectory” of maximum expected accuracy and minimum computational expenses. Instead of building the system from scratch we consider an agent learning to perceive images through the set of CNN classifiers learned under supervision. The goal of the agent is to select few networks and based on their response make a prediction. We show that there are two important functions the agent has to learn: classifier selection and state update. Optimal functions are intractable, and we investigate the opportunity to use neural networks learned by reinforcement signals as approximators. Reward in this case is the accuracy of the decision minus the normalized number of floating point operations associated with the classifiers selected by the agent to produce the decision. The problem is a special case of dynamic algorithm configuration, because the agent has to learn how to iteratively adjust its action based on the context of the concrete image. We conduct experiments with CIFAR datasets and consider several ways of learning initial pool of classifiers including boosting and bagging. We show that by the usage of the dynamic computational structure the agent is able to successfully incorporate knowledge of the image context through classifiers responses and achieve measurable improvement in the final classification accuracy, when computational resources are limited. We consider both on-policy and off-policy methods and observe complex dynamics of learning. To achieve the desired effect we used a specifically designed architecture of neural network and loss function. Advantage of the proposed configuration is that networks learned under the supervised signals can be deeper than ones learned in RL. At the same time agent manipulates probabilities distributions of object classes, and, therefore, is able to partially abstract from data features and learn policies possibly transferable to other domains. Side effect of computational restriction is a regularization effect.
Nr: 5
Title:

Biogas Recognition over Landfill with MOS Gas Sensors Array and PCA-quantile Regression
Authors:

Eric Martial Taguem and Anne-Claude Romain
Abstract: The monitoring of biogas emissions has become a concern regarding the greenhouse effect of methane emitted over the waste treatment plants. Besides, chemical sensors array as metal oxide semiconductor (MOS) one has the technical potential to manage, in real-time, these emissions. The use of these devices could help to organise fast, easy and regular controls regarding their low cost and their low power consumption. The data treatment of MOS sensors array signal is a crucial aspect for their use. In the purpose of performing biogas monitoring over a landfill, a prediction model based on PCA and quantile regression has been developed and tested over a landfill. The results showed that it was possible to recognise areas easily with biogas emissions and those with no emissions. Following these promising results, a mapping of methane concentrations over the landfill could capture possible emissions trends or patterns.