an loss function is introduced for minimizing the intra-class compactness and simultaneously
maximizing the inter-class distance in feature space.
Fig. 2 The upper part is a standard neural network which can be divided into a deep feature extractor and a softmax classifier. The
lower part is the proposed DNCM model which inherits the feature extractor and replaces the original classifier by a NCM classifier.
In the DNCM model, the yellow concentric circles represent mean of classes that have already been integrated into model, not neuron
of network. The green concentric circles represent the mean of new class that will be integrated into model. The yellow dotted lines
are not the same as the connections in network, which means that the distance between the output of feature extractor (feature vector)
and each class mean.
The rest of this paper is organized as follows. Section II show some related work; Section III
describe the detail of proposed model; Section IV present an experimental results, analyzing the
influence of parameters on the proposed DNCM and comparing them to the current state-of-art in
incremental odor recognition problems. Finally, we give our conclusion in Section V.
II Related work
Odor recognition by machine with E-nose is a challenge task for pattern recognition since gas
sensor is sensitive to physical environment. Especially, some odor classification task is based on
dynamic incremental dataset. Below we discuss various work about odor classification and
incremental learning.
A. Odor Classification
With the development of artificial intelligence, more and more machine learning algorithms
were proposed for odor recognition. L. Zhang et al. proposed discriminative subspace learning
which target at odor recognition across multiple E-nose [1] and extreme learning machine based
self-expression for abnormal odor detection [2]. Tudu et al. proposed an incremental fuzzy
approach to evaluate black tea quality, which can incremental learn new types of tea [3]. Rodriguez
et al. use support vector machine (SVM) for the calibration of gas sensor arrays [4]. Dixon et al. [5]
compared the performance of five pattern recognition in odor recognition: Euclidean Distance to
Centroids (EDC), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA),
Learning Vector Quantization (LVQ) and Support Vector Machines (SVM). In machine olfaction
community, there are some valuable reviews which summarized the pattern classification and signal
processing methods for E-nose [6, 7].
B. Deep Learning
Conventional machine learning divide the process of learning into two separate parts: feature
extractor and classifier. Feature extractor which transformed the raw data (for example the signal of
multiple sensors) into a suitable feature vector was normally designed artificially. It required careful