Artificial intelligence is a swiftly growing field, and many are researching whether AI can serve as a diagnostic aid in the medical domain. However, the primary weakness of traditional machine learning for many applications is energy efficiency, and this may hamper its ability to be effectively utilized in medicine for portable or edge systems. In order to be more effective, new energy-efficient machine learning paradigms must be investigated for medical applications. In addition, smaller models with fewer parameters would be better suited to medical edge systems. By processing data as a series of "spikes" instead of continuous values, spiking neural networks (SNN) may be the right model architecture to address these concerns. This work investigates the proposed advantages of SNNs compared to more traditional architectures when tested on various medical datasets. We compare the energy efficiency of SNN and recurrent neural network (RNN) solutions by finding sizes of each architecture that achieve similar accuracy. The energy consumption of each comparable network is assessed using standard tools for such evaluation. On the SEED human emotion dataset, SNN architectures achieved up to 20x lower energy per inference than an RNN while maintaining similar classification accuracy. SNNs also achieved 30x lower energy consumption on the PTB-XL ECG dataset with similar classification accuracy. These results show that spiking neural networks are more energy efficient than traditional machine learning models at inference time while maintaining a similar level of accuracy for various medical classification tasks. With this superior energy efficiency, this makes it possible for medical SNNs to operate on edge and portable systems. / Master of Science / As artificial intelligence grows in popularity, especially with the rise of new large language models like Chat-GPT, a weakness in traditional architectures becomes more pronounced. These AI models require ever-increasing amounts of energy to operate. Thus, there is a need for more energy-efficient AI models, such as the spiking neural network (SNN). In SNNs, information is processed in a series of spiking signals, like the biological brain. This allows the resulting architecture to be highly energy efficient and adapted to processing time-series data. A domain that often encounters time-series data and would benefit from greater energy efficiency is medicine. This work seeks to investigate the proposed advantages of spiking neural networks when applied to the various classification tasks in the medical domain. Specifically, both an SNN and a traditional recurrent neural network (RNN) were trained on medical datasets for brain signal and heart signal classification. Sizes of each architecture were found that achieved similar classification accuracy and the energy consumption of each comparable network was assessed. For the SEED brain signal dataset, the SNN achieved similar classification accuracy to the RNN while consuming as little as 5% of the energy per inference. Similarly, the SNN consumed 30x less energy than the RNN while classifying the PTB-XL ECG dataset. These results show that the SNN architecture is a more energy efficient model than traditional RNNs for various medical tasks at inference time and may serve as the solution to the energy consumption problem of medical AI.

Affective computing has gained significant attention from researchers in the last decade due to the wide variety of applications that can benefit from this technology. Often, researchers describe affect using emotional dimensions such as arousal and valence. Valence refers to the spectrum of negative to positive emotions while arousal determines the level of excitement. Describing emotions through continuous dimensions (e.g. valence and arousal) allows us to encode subtle and complex affects as opposed to discrete emotions, such as the basic six emotions: happy, anger, fear, disgust, sad and neutral. Recognizing spontaneous and subtle emotions remains a challenging problem for computers. In our work, we employ two modalities of information: video and audio. Hence, we extract visual and audio features using deep neural network models. Given that emotions are time-dependent, we apply the Temporal Convolutional Neural Network (TCN) to model the variations in emotions. Additionally, we investigate an alternative model that combines a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN). Given our inability to fit the latter deep model into the main memory, we divide the RNN into smaller segments and propose a scheme to back-propagate gradients across all segments. We configure the hyperparameters of all models using Gaussian processes to obtain a fair comparison between the proposed models. Our results show that TCN outperforms RNN for the recognition of the arousal and valence emotional dimensions. Therefore, we propose the adoption of TCN for emotion detection problems as a baseline method for future work. Our experimental results show that TCN outperforms all RNN based models yielding a concordance correlation coefficient of 0.7895 (vs. 0.7544) on valence and 0.8207 (vs. 0.7357) on arousal on the validation dataset of SEWA dataset for emotion prediction.

Localization is critical for self-awareness of any autonomous system and is an important part of the autonomous system stack which consists of many phases including sensing, perceiving, planning and control. In the sensing phase, data from on board sensors are collected, preprocessed and passed to the next phase. The perceiving phase is responsible for self awareness or localization and situational awareness which includes multi-objects detection and scene understanding. After the autonomous system is aware of where it is and what is around it, it can use this knowledge to plan for the path it can take and send control commands to pursue this path. In this proposal, we focus on the localization part of the autonomous stack using camera images. We deal with the localization problem from different perspectives including single images and videos. Starting with the single image pose estimation, our approach is to propose systems that not only have good localization accuracy, but also have low space and time complexity. Firstly, we propose SurfCNN, a low cost indoor localization system that uses SURF descriptors instead of the original images to reduce the complexity of training convolutional neural networks (CNN) for indoor localization application. Given a single input image, the strongest SURF features descriptors are used as input to 5 convolutional layers to find its absolute position and orientation in arbitrary reference frame. The proposed system achieves comparable performance to the state of the art using only 300 features without the need for using the full image or complex neural networks architectures. Following, we propose SURF-LSTM, an extension to the idea of using SURF descriptors instead the original images. However, instead of CNN used in SurfCNN, we use long short term memory (LSTM) network which is one type of recurrent neural networks (RNN) to extract the sequential relation between SURF descriptors. Using SURF-LSTM, We only need 50 features to reach comparable or better results compared with SurfCNN that needs 300 features and other works that use full images with large neural networks. In the following research phase, instead of using SURF descriptors as image features to reduce the training complexity, we study the effect of using features extracted from other CNN models that were pretrained on other image tasks like image classification without further training and fine tuning. To learn the pose from pretrained features, graph neural networks (GNN) are adopted to solve the single image localization problem (Pose-GNN) by using these features representations either as features of nodes in a graph (image as a node) or converted into a graph (image as a graph). The proposed models outperform the state of the art methods on indoor localization dataset and have comparable performance for outdoor scenes. In the final stage of single image pose estimation research, we study if we can achieve good localization results without the need for training complex neural network. We propose (Linear-PoseNet) by which we can achieve similar results to the other methods based on neural networks with training a single linear regression layer on image features from pretrained ResNet50 in less than one second on CPU. Moreover, for outdoor scenes, we propose (Dense-PoseNet) that have only 3 fully connected layers trained on few minutes that reach comparable performance to other complex methods. The second localization perspective is to find the relative poses between images in a video instead of absolute poses. We extend the idea used in SurfCNN and SURF-LSTM systems and use SURF descriptors as feature representation of the images in the video. Two systems are proposed to find the relative poses between images in the video using 3D-CNN and 2DCNN-RNN. We show that using 3D-CNN is better than using the combination of CNN-RNN for relative pose estimation. / Graduate

Neural networks was shown to be very susceptible to small and imperceptible perturbations on its input. In this thesis, we study perturbations on two-layer piecewise linear networks. Such studies are essential in training neural networks that are robust to noisy input. One type of perturbations we consider is `1 norm bounded perturbations. Training Deep Neural Networks (DNNs) that are robust to norm bounded perturbations, or adversarial attacks, remains an elusive problem. While verification based methods are generally too expensive to robustly train large networks, it was demonstrated in [1] that bounded input intervals can be inexpensively propagated per layer through large networks. This interval bound propagation (IBP) approach lead to high robustness and was the first to be employed on large networks. However, due to the very loose nature of the IBP bounds, particularly for large networks, the required training procedure is complex and involved. In this work, we closely examine the bounds of a block of layers composed of an affine layer followed by a ReLU nonlinearity followed by another affine layer. In doing so, we propose probabilistic bounds, true bounds with overwhelming probability, that are provably tighter than IBP bounds in expectation. We then extend this result to deeper networks through blockwise propagation and show that we can achieve orders of magnitudes tighter bounds compared to IBP. With such tight bounds, we demonstrate that a simple standard training procedure can achieve the best robustness-accuracy tradeoff across several architectures on both MNIST and CIFAR10. We, also, consider Gaussian perturbations, where we build on a previous work that derives the first and second output moments of a two-layer piecewise linear network [2]. In this work, we derive an exact expression for the second moment, by dropping the zero mean assumption in [2].

No description available.

The thesis deals with design of novel circuit structure suitable for hardware implementations of feedforward neural networks. The structure utilizes innovative data bus structure. The main contribution of the structure is in optimization of the utilization of implemented computing units. Proposed architecture is flexible and suitable for implementations of variety of feedforward neural network structures.

Bolagsverket, a Swedish government agency receives cases both in paper form via mail, document form via e-mail and also digital forms. These cases may be about registering people in a company, changing the share capital, etc. However, handling and confirming all these papers can be time consuming, and it would be beneficial for Bolagsverket if this process could be automated with as little human input as possible. This thesis investigates if it is possible to identify whether a paper contains a signature or not by using artificial intelligence (AI) and convolutional neural networks (CNN), and also if it is possible to determine how many signatures a given paper has. If these problems prove to be solvable, it could potentially lead to a great benefit for Bolagsverket. In this paper, a residual neural network (ResNet) was implemented which later was trained on sample data provided by Bolagsverket. The results demonstrate that it is possible to determine whether a paper has a signature or not with a 99% accuracy, which was tested on 1000 images where the model was trained on 8787 images. A second ResNet architecture was implemented to identify the number of signatures, and the result shows that this was possible with an accuracy score of 94.6%.

This dissertation explores the modular learning in artificial neural networks that mainly driven by the inspiration from the neurobiological basis of the human learning. The presented modularization approaches to the neural network design and learning are inspired by the engineering, complexity, psychological and neurobiological aspects. The main theme of this dissertation is to explore the organization and functioning of the brain to discover new structural and learning inspirations that can be subsequently utilized to design artificial neural network. The artificial neural networks are touted to be a neurobiologicaly inspired paradigm that emulate the functioning of the vertebrate brain. The brain is a highly structured entity with localized regions of neurons specialized in performing specific tasks. On the other hand, the mainstream monolithic feed-forward neural networks are generally unstructured black boxes which is their major performance limiting characteristic. The non explicit structure and monolithic nature of the current mainstream artificial neural networks results in lack of the capability of systematic incorporation of functional or task-specific a priori knowledge in the artificial neural network design process. The problem caused by these limitations are discussed in detail in this dissertation and remedial solutions are presented that are driven by the functioning of the brain and its structural organization. Also, this dissertation presents an in depth study of the currently available modular neural network architectures along with highlighting their shortcomings and investigates new modular artificial neural network models in order to overcome pointed out shortcomings. The resulting proposed modular neural network models have greater accuracy, generalization, comprehensible simplified neural structure, ease of training and more user confidence. These benefits are readily obvious for certain problems, depending upon availability and usage of available a priori knowledge about the problems. The modular neural network models presented in this dissertation exploit the capabilities of the principle of divide and conquer in the design and learning of the modular artificial neural networks. The strategy of divide and conquer solves a complex computational problem by dividing it into simpler sub-problems and then combining the individual solutions to the sub-problems into a solution to the original problem. The divisions of a task considered in this dissertation are the automatic decomposition of the mappings to be learned, decompositions of the artificial neural networks to minimize harmful interaction during the learning process, and explicit decomposition of the application task into sub-tasks that are learned separately. The versatility and capabilities of the new proposed modular neural networks are demonstrated by the experimental results. A comparison of the current modular neural network design techniques with the ones introduced in this dissertation, is also presented for reference. The results presented in this dissertation lay a solid foundation for design and learning of the artificial neural networks that have sound neurobiological basis that leads to superior design techniques. Areas of the future research are also presented. / Ph. D.

The thesis is written in chapter form. Chapter 1 describes some of the history of neural networks and its place in the field of artificial intelligence. It indicates the biological basis from which neural network approximation are made. Chapter 2 describes the properties of neural networks and their uses. It introduces the concepts of training and learning. Chapters 3, 4, 5 and 6 show the perceptron and adaline in feedforward and recurrent networks particular reference is made to regression substitution by "group method data handling. Networks are chosen that explain the application of neural networks in classification, association, optimization and self organization. Chapter 7 addresses the subject of practical inputs to neural networks. Chapter 8 reviews some interesting recent developments. Chapter 9 reviews some ideas on the future technology for neural networks. Chapter 10 gives a listing of some neural network types and their uses. Appendix A gives some of the ideas used in portfolio selection for the Johannesburg Stock Exchange. / Computing / M. Sc. (Operations Research)

This research focused on the implementation of a reliable intelligent fall detection system so as to reduce accidental falls among the elderly people. A video-based detection system was used because it preserved privacy while monitoring the activities of the senior citizens. Another advantage of the video-based system is that the senior citizens are able to move freely without experiencing any hassles in wearing them as opposed to portable fall detection sensors so that they can have a more independent and happy life. A scientific research method was employed to improve the existing fall detection systems in terms of reliability and accuracy. This thesis consists of four stages where the first stage reviews the literature on the current fall detection systems, the second stage investigates the various algorithms of these existing fall detection systems, the third stage describes the proposed fall detection algorithm in detecting falls using two distinct approaches. The first approach deals with the use of specific features of the silhouette, an extracted binary map obtained from the subtraction of the foreground from the background, to determine the fall angle (FA), the bounding box (BB) ratio, the Hidden Markov Models (HMM) and the combination of FA, BB, and HMM. The second approach used is the neural network approach which is incorporated in the algorithm to identify a predetermined set of situations such as praying, sitting, standing, bending, kneeling, and lying down. The fourth stage involves the evalua- tion of the developed video-based fall detection system using different metrics which measure sensitivity (i.e. the capacity of the fall detection system to detect as well as declare a fall) and specificity (i.e. the capacity of the algorithm to detect only falls) of this algorithm. The video camera was properly positioned to avoid any occluding objects and also to cover a certain range of motion of the stunt participants performing the falls. The silhouette is extracted using an approximate median filtering approach and the threshold criteria value of 30 pixels was used. Morphological filtering methods that were dilation and erosion were used to remove any spurious noises from the extracted image prior to subsequent feature analysis. Then, this extracted silhouette was scaled and quantised using 8 bits/pixel and compared to the set of predetermined scenarios using a neural network of perceptrons. This neural network was trained based on various situations and the falls of the participants which represent inputs to the neural network algorithm during the neural learning process. In this research study, the built neural network consisted of 600 inputs, as well as 10 neurons in the hidden layer together with 7 distinct outputs which represent the set of predefined situations. Furthermore, an alarm generation algorithm was included in the fall detection algorithm such that there were three states that were STATE NULL (set at 0), STATE LYING (set at 1) and STATE ALL OTHERS (set at 2) and the initial alarm count was set to 90 frames (meaning 3 seconds of recorded consecutive images at 30 frames per second). Therefore, an alarm was generated only when the in-built counter surpassed this threshold of 90 frames to signal that a fall occurred. Following the evaluation stage, it was found that the combination of the first approach fall detection algorithm method (fall angle, bounding box, and hidden Markov) was 89% with specificity and 84.2% with sensitivity which is better than individual performance. Moreover, it was found that the second approach fall detection algorithm method (neural network performance) 94.3% of the scenarios were successfully classified whereby the specificity of the developed algorithm was determined to be 94.8% and the sensitivity was 93.8% which altogether show a promising overall performance of the fall detection video-based intelligent system. Moreover, the developed fall detection system were tested using two types of handicaps such as limping and stumbling stunt participants to observe how well this detection algorithm can detect falls as in the practical situations encountered or present in elderly people. In these cases it was found that about 90.2% of the falls were detected which showed still that the developed algorithm was quite robust and reliable subjected to these two physical handicaps motion behaviours.