Online Continual Learning in Acoustic Scene Classification: An Empirical Study.

Numerous deep learning methods for acoustic scene classification (ASC) have been proposed to improve the classification accuracy of sound events. However, only a few studies have focused on continual learning (CL) wherein a model continually learns to solve issues with task changes. Therefore, in this study, we systematically analyzed the performance of ten recent CL methods to provide guidelines regarding their performances. The CL methods included two regularization-based methods and eight replay-based methods. First, we defined realistic and difficult scenarios such as online class-incremental (OCI) and online domain-incremental (ODI) cases for three public sound datasets. Then, we systematically analyzed the performance of each CL method in terms of average accuracy, average forgetting, and training time. In OCI scenarios, iCaRL and SCR showed the best performance for small buffer sizes, and GDumb showed the best performance for large buffer sizes. In ODI scenarios, SCR adopting supervised contrastive learning consistently outperformed the other methods, regardless of the memory buffer size. Most replay-based methods have an almost constant training time, regardless of the memory buffer size, and their performance increases with an increase in the memory buffer size. Based on these results, we must first consider GDumb/SCR for the continual learning methods for ASC.

Analysis and Validation of Cross-Modal Generative Adversarial Network for Sensory Substitution.

Visual-auditory sensory substitution has demonstrated great potential to help visually impaired and blind groups to recognize objects and to perform basic navigational tasks. However, the high latency between visual information acquisition and auditory transduction may contribute to the lack of the successful adoption of such aid technologies in the blind community; thus far, substitution methods have remained only laboratory-scale research or pilot demonstrations. This high latency for data conversion leads to challenges in perceiving fast-moving objects or rapid environmental changes. To reduce this latency, prior analysis of auditory sensitivity is necessary. However, existing auditory sensitivity analyses are subjective because they were conducted using human behavioral analysis. Therefore, in this study, we propose a cross-modal generative adversarial network-based evaluation method to find an optimal auditory sensitivity to reduce transmission latency in visual-auditory sensory substitution, which is related to the perception of visual information. We further conducted a human-based assessment to evaluate the effectiveness of the proposed model-based analysis in human behavioral experiments. We conducted experiments with three participant groups, including sighted users (SU), congenitally blind (CB) and late-blind (LB) individuals. Experimental results from the proposed model showed that the temporal length of the auditory signal for sensory substitution could be reduced by 50%. This result indicates the possibility of improving the performance of the conventional vOICe method by up to two times. We confirmed that our experimental results are consistent with human assessment through behavioral experiments. Analyzing auditory sensitivity with deep learning models has the potential to improve the efficiency of sensory substitution.

Accelerating On-Device Learning with Layer-Wise Processor Selection Method on Unified Memory.

Recent studies have applied the superior performance of deep learning to mobile devices, and these studies have enabled the running of the deep learning model on a mobile device with limited computing power. However, there is performance degradation of the deep learning model when it is deployed in mobile devices, due to the different sensors of each device. To solve this issue, it is necessary to train a network model specific to each mobile device. Therefore, herein, we propose an acceleration method for on-device learning to mitigate the device heterogeneity. The proposed method efficiently utilizes unified memory for reducing the latency of data transfer during network model training. In addition, we propose the layer-wise processor selection method to consider the latency generated by the difference in the processor performing the forward propagation step and the backpropagation step in the same layer. The experiments were performed on an ODROID-XU4 with the ResNet-18 model, and the experimental results indicate that the proposed method reduces the latency by at most 28.4% compared to the central processing unit (CPU) and at most 21.8% compared to the graphics processing unit (GPU). Through experiments using various batch sizes to measure the average power consumption, we confirmed that device heterogeneity is alleviated by performing on-device learning using the proposed method.

Multi-Path and Group-Loss-Based Network for Speech Emotion Recognition in Multi-Domain Datasets.

Speech emotion recognition (SER) is a natural method of recognizing individual emotions in everyday life. To distribute SER models to real-world applications, some key challenges must be overcome, such as the lack of datasets tagged with emotion labels and the weak generalization of the SER model for an unseen target domain. This study proposes a multi-path and group-loss-based network (MPGLN) for SER to support multi-domain adaptation. The proposed model includes a bidirectional long short-term memory-based temporal feature generator and a transferred feature extractor from the pre-trained VGG-like audio classification model (VGGish), and it learns simultaneously based on multiple losses according to the association of emotion labels in the discrete and dimensional models. For the evaluation of the MPGLN SER as applied to multi-cultural domain datasets, the Korean Emotional Speech Database (KESD), including KESDy18 and KESDy19, is constructed, and the English-speaking Interactive Emotional Dyadic Motion Capture database (IEMOCAP) is used. The evaluation of multi-domain adaptation and domain generalization showed 3.7% and 3.5% improvements, respectively, of the F1 score when comparing the performance of MPGLN SER with a baseline SER model that uses a temporal feature generator. We show that the MPGLN SER efficiently supports multi-domain adaptation and reinforces model generalization.

An Energy-Efficient Method for Human Activity Recognition with Segment-Level Change Detection and Deep Learning.

Human activity recognition (HAR), which is important in context awareness services, needs to occur continuously in daily life, owing to which an energy-efficient method is needed. However, because human activities have a longer cycle than HAR methods, which have analysis cycles of a few seconds, continuous classification of human activities using these methods is computationally and energy inefficient. Therefore, we propose segment-level change detection to identify activity change with very low computational complexity. Additionally, a fully convolutional network (FCN) with a high recognition rate is used to classify the activity only when activity change occurs. We compared the accuracy and energy consumption of the proposed method with that of a method based on a convolutional neural network (CNN) by using a public dataset on different embedded platforms. The experimental results showed that, although the recognition rate of the proposed FCN model is similar to that of the CNN model, the former requires only 10% of the network parameters of the CNN model. In addition, our experiments to measure the energy consumption on the embedded platforms showed that the proposed method uses as much as 6.5 times less energy than the CNN-based method when only HAR energy consumption is compared.