Tomás Maul

Mateus Karvat Camara, Adriana Postal, Tomas Henrique Maul et al.

Despite the great impact of lies in human societies and a meager 54% human accuracy for Deception Detection (DD), Machine Learning systems that perform automated DD are still not viable for proper application in real-life settings due to data scarcity. Few publicly available DD datasets exist and the creation of new datasets is hindered by the conceptual distinction between low-stakes and high-stakes lies. Theoretically, the two kinds of lies are so distinct that a dataset of one kind could not be used for applications for the other kind. Even though it is easier to acquire data on low-stakes deception since it can be simulated (faked) in controlled settings, these lies do not hold the same significance or depth as genuine high-stakes lies, which are much harder to obtain and hold the practical interest of automated DD systems. To investigate whether this distinction holds true from a practical perspective, we design several experiments comparing a high-stakes DD dataset and a low-stakes DD dataset evaluating their results on a Deep Learning classifier working exclusively from video data. In our experiments, a network trained in low-stakes lies had better accuracy classifying high-stakes deception than low-stakes, although using low-stakes lies as an augmentation strategy for the high-stakes dataset decreased its accuracy.

Chen Wen Kang, Chua Meng Hong, Tomas Maul

The combination and aggregation of knowledge from multiple neural networks can be commonly seen in the form of mixtures of experts. However, such combinations are usually done using networks trained on the same tasks, with little mention of the combination of heterogeneous pre-trained networks, especially in the data-free regime. This paper proposes multiple data-free methods for the combination of heterogeneous neural networks, ranging from the utilization of simple output logit statistics, to training specialized gating networks. The gating networks decide whether specific inputs belong to specific networks based on the nature of the expert activations generated. The experiments revealed that the gating networks, including the universal gating approach, constituted the most accurate approach, and therefore represent a pragmatic step towards applications with heterogeneous mixtures of experts in a data-free regime. The code for this project is hosted on github at https://github.com/cwkang1998/network-merging.

Weichuang Shao, Iman Yi Liao, Tomas Henrique Bode Maul et al.

Domain shift is a prominent problem in Deep Learning, causing a model pre-trained on a source dataset to suffer significant performance degradation on test datasets. This research aims to address the issue of audio classification under domain shift caused by background noise using Test-Time Adaptation (TTA), a technique that adapts a pre-trained model during testing using only unlabelled test data before making predictions. We adopt two common TTA methods, TTT and TENT, and a state-of-the-art method CoNMix, and investigate their respective performance on two popular audio classification datasets, AudioMNIST (AM) and SpeechCommands V1 (SC), against different types of background noise and noise severity levels. The experimental results reveal that our proposed modified version of CoNMix produced the highest classification accuracy under domain shift (5.31% error rate under 10 dB exercise bike background noise and 12.75% error rate under 3 dB running tap background noise for AM) compared to TTT and TENT. The literature search provided no evidence of similar works, thereby motivating the work reported here as the first study to leverage TTA techniques for audio classification under domain shift.

Mohammed Amer, Tomás Maul, Iman Yi Liao

The growing capacity of neural networks has strongly contributed to their success at complex machine learning tasks and the computational demand of such large models has, in turn, stimulated a significant improvement in the hardware necessary to accelerate their computations. However, models with high latency aren't suitable for limited-resource environments such as hand-held and IoT devices. Hence, many deep learning techniques aim to address this problem by developing models with reasonable accuracy without violating the limited-resource constraint. In this work, we use a one-shot neural architecture search model to implicitly evaluate the performance of an intractable number of multipath neural networks. Combining this architecture search with a pruning technique and architecture sample evaluation, we can model the relation between the accuracy and the latency of a spectrum of models with graded complexity. We show that our method can accurately model the relative performance between models with different latencies and predict the performance of unseen models with good precision across different datasets.

Kasra Babaei, ZhiYuan Chen, Tomas Maul

This paper proposes an autoencoder (AE) that is used for improving the performance of once-class classifiers for the purpose of detecting anomalies. Traditional one-class classifiers (OCCs) perform poorly under certain conditions such as high-dimensionality and sparsity. Also, the size of the training set plays an important role on the performance of one-class classifiers. Autoencoders have been widely used for obtaining useful latent variables from high-dimensional datasets. In the proposed approach, the AE is capable of deriving meaningful features from high-dimensional datasets while doing data augmentation at the same time. The augmented data is used for training the OCC algorithms. The experimental results show that the proposed approach enhance the performance of OCC algorithms and also outperforms other well-known approaches.

Kasra Babaei, Zhi Yuan Chen, Tomas Maul

Anomaly detection is referred to as a process in which the aim is to detect data points that follow a different pattern from the majority of data points. Anomaly detection methods suffer from several well-known challenges that hinder their performance such as high dimensionality. Autoencoders are unsupervised neural networks that have been used for the purpose of reducing dimensionality and also detecting network anomalies in large datasets. The performance of autoencoders debilitates when the training set contains noise and anomalies. In this paper, a new gradient-reversal method is proposed to overcome the influence of anomalies on the training phase for the purpose of detecting network anomalies. The method is different from other approaches as it does not require an anomaly-free training set and is based on reconstruction error. Once latent variables are extracted from the network, Local Outlier Factor is used to separate normal data points from anomalies. A simple pruning approach and data augmentation is also added to further improve performance. The experimental results show that the proposed model can outperform other well-know approaches.

Mohammed Amer, Tomás Maul

Lifelong learning is a very important step toward realizing robust autonomous artificial agents. Neural networks are the main engine of deep learning, which is the current state-of-the-art technique in formulating adaptive artificial intelligent systems. However, neural networks suffer from catastrophic forgetting when stressed with the challenge of continual learning. We investigate how to exploit modular topology in neural networks in order to dynamically balance the information load between different modules by routing inputs based on the information content in each module so that information interference is minimized. Our dynamic information balancing (DIB) technique adapts a reinforcement learning technique to guide the routing of different inputs based on a reward signal derived from a measure of the information load in each module. Our empirical results show that DIB combined with elastic weight consolidation (EWC) regularization outperforms models with similar capacity and EWC regularization across different task formulations and datasets.

Kasra Babaei, ZhiYuan Chen, Tomas Maul

Outlier detection (also known as anomaly detection or deviation detection) is a process of detecting data points in which their patterns deviate significantly from others. It is common to have outliers in industry applications, which could be generated by different causes such as human error, fraudulent activities, or system failure. Recently, density-based methods have shown promising results, particularly among which Local Outlier Factor (LOF) is arguably dominating. However, one of the major drawbacks of LOF is that it is computationally expensive. Motivated by the mentioned problem, this research presents a novel pruning-based procedure in which the execution time of LOF is reduced while the performance is maintained. A novel Prune-based Local Outlier Factor (PLOF) approach is proposed, in which prior to employing LOF, outlierness of each data instance is measured. Next, based on a threshold, data instances that require further investigation are separated and LOF score is only computed for these points. Extensive experiments have been conducted and results are promising. Comparison experiments with the original LOF and two state-of-the-art variants of LOF have shown that PLOF produces higher accuracy and precision while reducing execution time.

Mohammed Amer, Tomás Maul

Convolutional neural networks (CNNs) are known for their good performance and generalization in vision-related tasks and have become state-of-the-art in both application and research-based domains. However, just like other neural network models, they suffer from a susceptibility to noise and adversarial attacks. An adversarial defence aims at reducing a neural network's susceptibility to adversarial attacks through learning or architectural modifications. We propose the weight map layer (WM) as a generic architectural addition to CNNs and show that it can increase their robustness to noise and adversarial attacks. We further explain that the enhanced robustness of the two WM variants results from the adaptive activation-variance amplification exhibited by the layer. We show that the WM layer can be integrated into scaled up models to increase their noise and adversarial attack robustness, while achieving comparable accuracy levels across different datasets.

Mohammed Amer, Tomás Maul

Capsule network (CapsNet) was introduced as an enhancement over convolutional neural networks, supplementing the latter's invariance properties with equivariance through pose estimation. CapsNet achieved a very decent performance with a shallow architecture and a significant reduction in parameters count. However, the width of the first layer in CapsNet is still contributing to a significant number of its parameters and the shallowness may be limiting the representational power of the capsules. To address these limitations, we introduce Path Capsule Network (PathCapsNet), a deep parallel multi-path version of CapsNet. We show that a judicious coordination of depth, max-pooling, regularization by DropCircuit and a new fan-in routing by agreement technique can achieve better or comparable results to CapsNet, while further reducing the parameter count significantly.

Kien Tuong Phan, Tomas Henrique Maul, Tuong Thuy Vu et al.

In an attempt to solve the lengthy training times of neural networks, we proposed Parallel Circuits (PCs), a biologically inspired architecture. Previous work has shown that this approach fails to maintain generalization performance in spite of achieving sharp speed gains. To address this issue, and motivated by the way Dropout prevents node co-adaption, in this paper, we suggest an improvement by extending Dropout to the PC architecture. The paper provides multiple insights into this combination, including a variety of fusion approaches. Experiments show promising results in which improved error rates are achieved in most cases, whilst maintaining the speed advantage of the PC approach.