The development of a trellis-coded modulation (TCM) based probabilistic amplitude shaping (PAS) coding scheme has the potential to greatly improve the rate of data transmissions to meet the rapidly growing data demands. Previous studies on convolutional codes have determined that there is a theoretical maximum noise threshold in the memoryless channels used for data transmission; a TCM based PAS design has the potential to reach a higher maximum threshold in comparison to low-density parity-check (LDPC) codes. In order for the receiver to determine the initial data sequence from the encoded data, we propose an implementation of the maximum likelihood soft decision Viterbi algorithm that calculates the branch metrics using a probability vector for the constellation design in order to determine the survivor path.

Spinal cord injury (SCI) is a catastrophic event that’s becoming more common yearly. In order to develop techniques to aid in spinal cord repair, we need to have a detailed understanding of the pathophysiological events that occur during the SCI. Within the body, hyaluronic acid (HA) makes up the backbone of the extracellular matrix. After SCI, the high molecular weight (HMW) HA-rich matrix is disrupted, causing low MW (LMW) fragments of HA to be released. This change causes a cascade of events in response to the varied MW of HA fragments. This research explores how endothelial cells (ECs) within the SCI environment respond differentially to HA of varying MW. We are investigating these responses by assessing the effects of HA MW on EC migration, chemotaxis, and tube formation. Our preliminary data indicates that ECs migrate faster in response to solubilized 41-65 kDa (40K) and 100-150 kDa (100K) HA than 750-1000 kDa (1M) and 10-20 kDa (10K) HA using a scratch assay. As ECs are vital to restoring proper function after SCI, this data, in combination with chemotaxis and tube formation studies, is needed to illuminate the effects of HA matrix disruption on ECs. Furthermore, this data will be instrumental in the development of HA-based biomaterials as a therapeutic strategy to repair the injured spinal cord.

As automated machines, from ATMs to self-driving cars, become an important aspect of day-to-day human existence, a machines ability to see and react to visual stimuli becomes a vital role in robotics. There are many ways to give machines senses like through Ultrasonic sensors, LiDAR or Cameras. LiDAR can be useful as it can provide 360 degree visualization; therefore it is important that LiDAR data is efficiently translated into the virtual world. This study presents efforts on reconstructing three-dimensional LiDAR data in a computer. For reconstruction, we utilized a Velodyne Lidar sensor and collected data of the real world. We, then, extracted features like location and reconstructed that in the virtual world, trying to create a solid mesh. Initial representations of the real world onto the virtual world were obtained.

The Origami Design App is a web application that can view, compile, and change the parameters of robots along interfaces. UCLA’s LEMUR originally designed this webapp as a means to make robot compiling easier for those that lack the engineering or the programming background. I integrated some of the functionalities such as being able to view necessary component files when inputting a subcomponent into a html form by using a PATH method. This is significant because a user could potentially compile a robot easier or faster than using the standard RoCo application. I also edited some of the existing user interface such as editing the style of transitioning buttons with css files to provide the user with optimal visual experience. For the future, I plan to conduct a user study to observe that the changes I have implemented have benefited the Origami Design App and prove that the changes within the user interface were successful. I also plan to implement the functionality to combine components along interfaces which will most likely reduce the time that a user needs to spend in order to design a robot.

The Clark Lab studies germ-line regeneration in pursuit of in-vitro gametogenesis, an approach being explored for people who lack or have lost the ability to produce their own viable gametes. Achieving this involves first determining the mechanisms of germline cell differentiation in prenatal life. As access to human fetal tissue during gestation is limited, the Clark lab instead uses a number of model systems, including the non-human primate rhesus macaque. This especially enables the study of female gametogenesis, a process that is largely confined to fetal gestation. Over these ten weeks, I have begun to study the fetal ovarian environment in the rhesus macaque model. I have been performing immunofluorescence analysis on rhesus ovarian tissue sections to characterize protein expression at different gestational stages, with a focus on basement membrane proteins present throughout the ovary. This has the potential to inform on the identity of the supporting somatic cells within the ovary, that could then promote recapitulation of the ovarian niche in vitro. In tandem, I have been working on a literature review of fetal ovarian development in humans and model organisms, specifically non-human primates. I am focusing on the proposed origins of the somatic compartment and associated basement membrane derivatives, and aim to confirm any findings with further immunofluorescence during my project.

RNA editing fundamentally alters the flow of information from genes to proteins. One type of editing is adenosine-to-inosine editing, read by a cell’s machinery as adenosine-to-guanine editing. A-to-I editing is catalyzed by the RNA-binding protein (RBP) adenosine deaminase acting on RNA (ADAR). HuR is another RBP involved in the stabilization of bound RNA transcripts, and it is thought to interact with ADAR. To examine the effect of ADAR-HuR interaction on the stability of RNA transcripts, HuR knockdown was performed on a population of cervical cancer (HeLa) cells using RNA interference. RNA transcripts were then extracted for sequencing, which allowed comparison of the amount of edited and unedited RNA in the transcriptome-wide scale. RNA sequences allow for the comparison of the total number of RNA transcripts, thus allowing a measurement of gene expression and RNA editing levels in the control and knockdown conditions. HuR knockdown led to lower levels of RNA editing but not lower proportions of A-to-G editing. Since the relationship between HuR and ADAR remains unclear, it is unknown whether HuR knockdown affected RNA editing levels via disruption of the interaction between HuR and ADAR. Future studies could focus on the nature of the interaction between both proteins.

Histopathological evaluations are key to diagnosing and managing ovarian cancer. Pathologists are trained to recognize visual nuclear features associated with tumor aggressiveness and clinical outcomes, such as nucleus/cytoplasm ratio, mitotic figures, and nuclear atypia. However, since most patients are not profiled at the molecular level, it is unknown if these phenotypic features are directly associated with altered molecular pathways in patient samples. In contrast, a wealth of molecular data has been generated on ovarian cancer cell lines, but a comprehensive pathology image database is lacking. Recently developed artificial intelligence and machine learning tools used to interpret image data provide new opportunities to quantify the spatial nuclear organization in ovarian cancer cell lines. Computers can precisely quantitate thousands of image features, some of which are visible while others are quantitatively incomprehensible to the human mind. We hypothesize that salient visual and subvisual features extracted from images of ovarian cancer cell lines are associated with distinct molecular profiles, such as specific gene and protein expression, metabolism, and sensitivity to specific drugs. Our multidisciplinary team will use artificial intelligence and deep learning strategies to generate a database of features that correlate with specific molecular pathways in ovarian cancer cell lines. We will then integrate the image features, molecular profiles, and clinical parameters into a multidimensional database that can be mined for features that correlate with specific genes and molecular pathways as well as clinical parameters, such as response to specific therapies.

Communications systems are crucial to modern everyday life – whether it be Wi-Fi, satellite communications, or storing and sharing documents digitally. However, imperfect communication channels can result in noise distorting transmitted data. Error correcting codes seek to identify and correct distorted data. Error correction comes at the cost of efficiency – this project’s rate-⅓ trellis encoder outputs three encoded bits per every one information bit, making it three times as inefficient as only sending the original data. By puncturing – omitting certain bits in transmission – higher efficiency can be attained, but the chance of receiving the correctly decoded information decreases. Previous literature has investigated the characteristics of a rate-⅓ 64-state 8PSK-modulated trellis encoder under puncturing and used these characteristics to develop a bit error rate (BER) union bound on the data. By running BER data simulations using C++ and the Hoffman2 Cluster and comparing them to theoretical union bound plots in MATLAB, this project has confirmed the results found in previous literature. Additionally, this project will extend the BER union bound methods to develop a union bound for the frame error rate (FER), as well as simulating the FER performance for this specific encoder and various puncturing patterns of interest. Whether in satellite transmissions, self-driving cars, streaming, Wi-Fi, memory storage hard drives, 5G, or GPS, our research has countless applications in the modern, digital world.

Traditional meat production not only contributes to environmental degradation, but also may be insufficient for growing food demands due to limitations in land and water usage. Creating cultured meat may be key to answering these problems. Palatable, flavorful meat that appeals to consumers must integrate fat cells, or adipocytes, which develop and proliferate in a physical environment-dependent manner. However, much is unknown about how adipocytes grow in 3D culture, which is necessary for future scalable production in artificial meat culturing. Here, we studied the growth of adipocytes on polyacrylamide microbeads of different stiffnesses to identify the ideal stiffness for maximizing adipocyte growth and lipid accumulation in these cells. Knowledge of how to optimize adipocyte growth will advance our fundamental knowledge of adipocyte biology, thereby enhancing the efficiency of scalable adipose tissue production for the consumer market.

In machine learning, the main objective is to learn a centralized model by exploiting the numerous data which is available from the clients. However, the clients’ data might contain personal and sensitive information, and hence, it is required to provide privacy guarantees on the clients' data. In this work, we examine differentially private training algorithms for convolutional neural networks on training practical datasets such as MNIST, ENMIST, and CIFAR10. In order to maximize testing accuracy within a fixed privacy budget, we explore the usage of novel noise functions instead of Gaussian noise in the differential privacy algorithm and transfer learning for CIFAR10 from a network trained on CIFAR100. Furthermore, we consider the problem of retaining privacy while training neural networks in a federated learning framework, where data is stored and accessed locally by the client and a central server builds and updates the neural network model. Also, separately, we examine different mechanisms for updating the global model other than the aggregate average of the local updates.