Stefan
Shakeri
Sponsor: Marc Facciotti, Ph.D. Biomedical Engineering E. coli, S. cerevisiae, and CHO cells currently dominate the production of therapeutic proteins. Microalgae, such as Chlamydomonas reinhardtii, present a promising low-cost, stable, and ecologically sustainable alternative platform for mammalian therapeutic protein production. However, the rapid and stable insertion and expression of transgenic DNA remains a significant challenge to using C. reinhardtii for heterologous protein expression. Our research team has correspondingly focused our recent efforts along two main thrusts: exploring different methods for transforming C. reinhardtii and improving plasmids for protein expression in this host. We found that glass bead transformation is an accessible alternative to the more commonly used electroporation and requires no specialized equipment or expensive reagents. We will present current results and data describing refinements to our transformation protocols. In addition, we present on the design and construction of new protein expression vectors that improve on existing options by enhancing customization via standardized assembly and removing selection markers whose use may pose health hazards. We envision that, together, our transformation protocols and standardized vectors can be used to engineer new strains of C. reinhardtii for the more sustainable production of therapeutic proteins. Analysis of Model Collapse in a Random Forest Machine Learning Algorithm
Abstract profile. Full document pending author claim.
Authors:
Stefan Shakeri
Date Created:
Not specified
Course Title:
Professor:
Not specified
About Paper:
Model collapse is the phenomenon in which artificial intelligence (AI) and machine learning (ML) models deteriorate in performance when trained on self-generated data over generations. AI and ML models will lose the ability to account for variance in the training data set, resulting in the loss of rare events and a convergence on certain outputs in the model predictions. With the increasing prevalence of AI-generated data on the internet (AI-generated text, images, etc.), knowing which models are more vulnerable to model collapse is necessary to better understand which models to use in what scenario when the possibility of AI-generated data is present. Significant research has been conducted on the effects of model collapse on generative adversarial networks (GANs) and large language models (LLMs); however, research on the effects of model collapse on non-generative models is less frequent. This study performs model collapse of a random forest ML model by training it across 20 generations on two separate datasets with self-generated data from the random forest model being used on each subsequent generation. Across all experiments conducted, the random forest model averaged a 12.15% reduction in accuracy through the 20 generations of training on self-generated data. UC Davis 36 th Annual Undergraduate Research, Scholarship and Creative Activities Conference 252 Quantifying AI Exposure in Stock Valuation: Developing an AI Factor with NVIDIA as a Proxy Aditi Shankar
Source:
UC Davis / Engr Computer Science / 2025
Topics:
No topics listed
Co-authors:
Stefan Shakeri