Atrial Fibrillation (AF) is the most common cardiac arrhythmia worldwide. It is associated with reduced quality of life and increases the risk of stroke and myocardial infarction. Unfortunately, many cases of AF are asymptomatic and undiagnosed, which increases the risk for the patients. Due to its paroxysmal nature, the detection of AF requires the evaluation, by a cardiologist, of long-term ECG signals. In Colombia, it is difficult to have access to an early diagnosis of AF because of the associated costs to the detection and the geographical distribution of cardiologists. This work is part of a macro project that aims to develop a specific-patient portable device for the detection of AF. This device will be based on a Convolutional Neural Network (CNN). We aim to find a suitable CNN model, which later could be implemented in hardware. Diverse techniques were applied to improve the answer regarding accuracy, sensitivity, specificity, and precision. The final model achieves an accuracy of , a specificity of , a sensitivity of  and a precision of . During the development of the model, the computational cost and memory resources were taking into account in order to obtain an efficient hardware model in a future implementation of the device.

Convolutional Neural Networks (CNNs) are becoming increasingly popular in deep learning applications, e.g. image classification, speech recognition, medicine, to name a few. However, the CNN inference is computationally intensive and demanding a large among of memory resources. In this work is proposed a CNN inference hardware accelerator, which was implemented in a co-processing scheme. The aim is to reduce the hardware resources and achieve the better possible throughput. The design was implemented in the Digilent Arty Z7-20 development board, which is based on System on Chip (SoC) Zynq-7000 of Xilinx. Our implementation achieved a  of accuracy for the MNIST database using only 12-bits fixed-point format. The results show that the co-processing scheme operating at a conservative speed of 100 MHz can identify around 441 images per second, which is about 17% times faster than a 650 MHz - software implementation. It is difficult to compare our results against other implementations based on Field-Programmable Gate Array (FPGA), because the others implementations are not exactly like ours. However, some comparisons, regarding the logical resources used and accuracy, suggest that our work could be better than previous works.

Different seismic data compression algorithms have been developed in order to make the storage more efficient, and to reduce both the transmission time and cost. In general, those algorithms have three stages: transformation, quantization and coding. The Wavelet transform is highly used to compress seismic data, due to the capabilities of the Wavelets on representing geophysical events in seismic data. We selected the lifting scheme to implement the Wavelet transform because it reduces both computational and storage resources. This work aims to determine how the transformation and the coding stages affect the data compression ratio. Several 2D lifting-based algorithms were implemented to compress three different seismic data sets. Experimental results obtained for different filter type, filter length, number of decomposition levels and coding scheme, are presented in this work. 221|Seismic Data Compression using 2D Lifting-Wavelet Algorithms Compresión de datos sísmicos usando algoritmos lifting-wavelet 2DResumen Diferentes algoritmos para compresión de datos sísmicos han sido desarrollados con el objetivo de hacer más eficiente el uso de capacidad de almacenamiento, y para reducir los tiempos y costos de la transmisión de datos. En general, estos algoritmos tienen tres etapas: transformación, cuantización y codificación. La transformada Wavelet ha sido ampliamente usada para comprimir datos sísmicos debido a la capacidad de las ondículas para representar eventos geofísicos presentes en los datos sísmicos. En este trabajo se usa el esquema Lifting para la implementación de la transformada Wavelet, debido a que este método reduce los recursos computacionales y de almacenamiento necesarios. Este trabajo estudia la influencia de las etapas de transformación y codificación en la relación de compresión de los datos. Además se muestran los resultados de la implementación de diferentes esquemas lifting 2D para la compresión de tres diferentes conjuntos de datos sísmicos. Los resultados obtenidos para diferentes tipos de filtros, longitud de filtros, número de niveles de descomposición y esquemas de compresión son presentados en este trabajo.

Deep learning has become increasingly popular and widely applied to computer vision systems. Over the years, researchers have developed various deep learning architectures to solve different kinds of problems. However, these networks are power-hungry and require high-performance computing (i.e., GPU, TPU, etc.) to run appropriately. Moving computation to the cloud may result in traffic, latency, and privacy issues. Edge computing can solve these challenges by moving the computing closer to the edge where the data is generated. One major challenge is to fit the high resource demands of deep learning in less powerful edge computing devices. In this research, we present an implementation of an embedded facial recognition system on a low cost Raspberry Pi, which is based on the FaceNet architecture. For this implementation it was required the development of a library in C++, which allows the deployment of the inference of the Neural Network Architecture. The system had an accuracy and precision of 77.38% and 81.25%, respectively. The time of execution of the program is 11 seconds and it consumes 46 [kB] of RAM. The resulting system could be utilized as a stand-alone access control system. The implemented model and library are released at https://github.com/cristianMiranda-Oro/FaceNet_EmbeddedSystem