Electrical and Electronics Engineering Institute

University of the Philippines - Diliman

Thesis/Dissertation Presentations

The graduate programs of the Institute provide advanced training and specialization in a broad range of areas in electrical and electronics engineering and its allied fields to prepare students to solve complex technological problems and to contribute new knowledge to the field. As a requirement for graduation, our students present and defend their respective thesis/dissertation. Below is a list of the defense schedule of our students for the 2nd Semester AY 2019-2020.

Date and Time Venue Name Title Abstract

27 May 2020 (2:00PM)

will be streamed online

Alexis Czezar C. Torreno

MS Thesis: Power and Area Oriented Implementations of Lightweight Cryptographic Algorithms for FPGA based Wireless Sensor Networks

Security is a concern in wireless sensor networks, which are inherently prone to third party attacks. As such, cryptography is used to make a secure mode of communication among nodes and/or between nodes and base stations. However, conventional algorithms are resource-hungry and therefore not fit for small devices, hence the creation of a new category of cryptography known as Lightweight Cryptographic Algorithms. These algorithms are still continuously being improved to fit in the decreasing sizes of small scale devices like FPGA-based wireless sensor networks. Different optimizations have been used to improve area and power efficiency of well known ciphers. However, these ciphers have a limit on how much they can be improved. 

In this study, we quantify the effects of Round Unrolling, and Data Width Reduction on area, and power consumption. These are tested on three candidate ciphers: LiCi, ANUII, and QTL. Round Unrolling reduces cipher latency which lessens the contribution of static power for each computation. Results show that round unrolling improves power efficiency by 25.97%, 3%, and 14% in LiCi, ANUII, and QTL, respectively. This comes at 299%, 414%, and 52% increase in area. Data Width Reduction improves area by only using a fraction of the original datapath. Data width reduction was found to be ineffective in reducing area in the candidate ciphers, increasing area by 5.64%, 6%, and 9.2% in LiCi, ANUII, and QTL, respectively. The results allow newer and better lightweight ciphers to be further improved for small scale devices. Round Unrolling can be used for power oriented systems, and while Data Width Reduction did not work on these ciphers, the effect can still be tested on others first before using on area oriented systems.

27 May 2020 (3:00PM)

will be streamed online

Ryan Albert G. Antonio

MS Thesis: Post-training Bit-selection Control for Energy-Efficient Hyperdimensional Computing Architecture

Hyperdimensional computing (HDC) is a brain-inspired computing framework that provides simple and convenient methods to perform cognitive tasks like classification. Its foundation lies in the properties of very high dimensional vectors called hypervectors (HV). The first attempts to create an energy-efficient HDC hardware contains massive bit-wise operations. State-of-the-art designs focus on either optimizing computations or developing new devices; however, some of these may directly affect the accuracy performance of the algorithm. 

After a careful investigation of the HDC algorithm, there exist redundant bit locations in the class HVs that do not contribute any significant information during classification. These irrelevant bits can be shut-off to improve its energy efficiency. However, the amount of redundancy is dependent on the cross-similarity of the original data set of a particular application. This thesis presents two major contributions. First are mathematical models relating to the redundancy and the cross-similarity of a given data set. Second, a bit-selection control that is added to the generic hardware designs, which disables the redundant bit-wise operations and improves the overall energy efficiency without sacrificing accuracy. Results show a 10% – 67% energy reduction at the cost of 8% – 20% increase in area and < 10% energy overhead cost. Additionally, mathematical models describing resulting energy savings were also developed. 

4 June 2020 (3:00PM)

will be streamed online

Daryl L. Peralta

MS Thesis: Next-Best View Policy for 3D Reconstruction

Creating 3D models of large structures requires capturing monocular or depth images of the target structure at different viewpoints using aerial drones. The selection of these viewpoints has a significant effect on the quality of the output 3D model. Manually selecting viewpoints or using commonly available flight paths like a circular path often results in insufficient viewpoints and incomplete 3D models. On the other hand, adding more viewpoints results in longer processing time and longer flight path. Recent works have relied on hand-engineered heuristics such as maximizing the information gain to select the Next-Best View (NBV) and acquire an optimal path. In this work, we cast the problem of view planning to a reinforcement learning setting where an agent learns an NBV policy to scan houses optimally by maximizing a reward. We call this learning-based algorithm Scan-RL. To train and evaluate our algorithm, we created Houses3K, a dataset of textured 3D house models. Our experiments show that using Scan-RL, the trained NBV policy can be used to scan houses with fewer number of steps and a shorter distance compared to the baseline circular path. Experimental results using Houses3K demonstrate that a single NBV policy can be used to scan multiple houses including those that were not seen during training.

5 June 2020 (9:00AM)

will be streamed online

Rangel D. Daroya

 

MS Thesis: REIN: Flexible Mesh Generation from Point Clouds

Efficient 3D object reconstruction is important for several computer vision tasks. Objects in 3D can be digitally represented as a point cloud, an occupancy grid, or a mesh. Lidar sensors often acquire sparse point cloud data. In addition, a point cloud's scattered form and lack of surfaces limits its utility compared to meshes. Occupancy grids are an alternative, but have limited resolution when depicting surfaces and have large memory usage. Meshes have continuous surface information and can represent objects with varying point densities. Existing surface reconstruction methods such as Ball Pivoting Algorithm (BPA) and Poisson Surface Reconstruction (PSR) interpolate from point clouds to produce meshes, but their dependence on point density causes significant performance decline with decreasing number of points. To address surface reconstruction from sparse points, we propose REIN: Recurrent Edge Inference Network. REIN is a neural network that generates meshes from point clouds by sequentially generating edges and faces. The network can produce outputs with information about the general structure of the object. REIN's sequential nature provides continuous feedback on edge generation and results in flexible mesh outputs with varying number of vertices. In this work, we demonstrate the mesh generation improvement of REIN compared to other surface reconstruction methods: BPA and PSR. Experimental results on ShapeNet and ModelNet10 show 81.5% average improvement in Chamfer Distance and 14% average improvement in Point Normal Similarity compared to Ball Pivoting Algorithm and Poisson Surface Reconstruction. Qualitatively, the generated meshes have a closer similarity to the ground truth.