2
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Discipline
A new volterra neural network layer lıbrary usıng tensorflow
Deep learning algorithms have garnered much attention recently, their success in enhancing the accuracy of automatic speech recognition systems has caused an increase in their usage. These models have the ability to identify patterns within input data and generate predictions based on these patterns. However, a limitation of these models is their inability to capture nonlinear relationships within input data. This study aimed to enhance the performance of automatic speech recognition through the incorporation of a second-order, one-dimensional Volterra Convolution (Volt1D) layer into deep learning models. The Volt1D layer is a custom TensorFlow layer that is founded on the Volterra series convolution, a mathematical tool capable of representing a wide array of nonlinear functions. To evaluate the efficacy of the Volt1D layer, we compared its performance to that of the standard Conv1D layer using the speech commands dataset v0.01, which consists of 20 classes of spoken words. Our results indicated that the Volt1D layer achieved an accuracy of 64.91% over 10 epochs, a significant improvement over the baseline accuracy of 60.02% using the Conv1D layer over 10 epochs. This demonstrates the Volt1D layer's effectiveness in extracting nonlinear relationships within speech data. We talked about the advantages and disadvantages of these approaches. with regard to the main objective of the study, which was to capture nonlinear relationships in speech data through the use of the Volt1D layer. We found that the Volt1D layer is a promising approach for speech recognition due to its ability to effectively capture nonlinear relationships and enhance the effectiveness of deep learning models. One of the primary strengths of the Volt1D layer is its capability to represent a wide range of nonlinear functions, making it suitable for capturing complex relationships within input data, a crucial factor in accurately transcribing spoken words into written text. In addition, the Volt1D layer is computationally efficient, allowing for its utilization in real-time speech recognition applications without incurring excessive computational overhead. However, there are also some limitations to the Volt1D layer. One limitation is that these models require a vast amount of training data to effectively learn the intricate nonlinear relationships within input data, which can be problematic for smaller datasets or applications with limited access to ample training data. Additionally, the Volt1D layer may not be as effective at capturing long-term dependencies within input data as other methods like RNN or LSTM models In conclusion, the Volt1D layer represents a promising approach for improving the performance of automatic speech recognition through deep learning models. It is able to effectively capture nonlinear relationships within input data, exhibiting superior accuracy and computational efficiency compared to other methods. However, the model may be constrained by the need for a substantial amount of training data and its capacity to identify long-term dependencies within input data.
A new deep learning based object detection system for increasing salesman performance
Food distribution companies must be flexible and responsive to provide the best services to their customers. Operational costs and expenses are the biggest concern for these companies, as they rely on many employees to serve customers and achieve the company's goals. Increasing tasks creates continuous and growing pressure on the sales team and the rest of its supporting departments, which affects the time and effort of the salesperson in achieving daily tasks and thus achieving the expected monthly goal. Deep learning technologies provide solutions to help reduce the effort expended by the sales team and, at the same time, contribute to reducing operating costs. In this thesis, an integrated system is proposed to use deep learning techniques to reduce the representative's effort and compensate the human support staff with deep learning algorithms to provide the necessary support to the representative. YOLO algorithms are used to discover objects in real-time and Faster-RCNN algorithm is used to solve planogram issues on shelf. The results extracted from images are used from these algorithms and converted into data stored in the MySQL database. Google API techniques are used to transfer the results from the cloud to the company's servers and thus analyze the results received using the Microsoft Power BI tool. The results show that proposed system provides a complete solution to reduce the effort of the sales team and significantly reduce costs. Also, this system can be used to study the competitor productions visibility on shelves.