Hi, I'm Jocelyn
I’m Jocelyn Rego, a computer science PhD candidate at Drexel University, advised by Dr. Ed Kim. I am passionate about using neuroscience knowledge to advance artificial intelligence, and vice versa. My research interests include biologically inspired machine learning, generative AI, computational neuroscience, brain computer interfaces, cognitive science, and neuromorphic computing.
Evolution has guided the brain to be able to quickly process a near constant influx of input stimuli, allowing for biological agents to effectively understand and interact with their environment in order to survive. Biological intelligence displays features of robustness, generalizability, and efficiency, fundamental to its ability to construct a coherent representation of the world in which it resides. These key features of biological intelligence are essential for advanced artificially intelligent systems to exhibit but have presented a challenge to replicate. The capabilities of machine intelligence are still far from those of the brain; however, fundamental principles of biological intelligence can serve as a muse for how these characteristics of robustness and adaptability may be achieved.
This thesis centers around the synergistic relationship between machine intelligence and neuroscience: how artificially intelligent systems can be improved by borrowing features from the brain and how machine learning methods can be applied to advance neuroscience knowledge. In pursuit of studying the first side of this bidirectional relationship, we describe neuroscience inspired machine learning systems that include features present in the brain, such as sparsity, competition, feedback, and spiking communication. These neuroscience inspired systems display an inherent robustness, generalizability, and efficiency; fundamental characteristics of biological intelligence that are desirable, but remain out of reach, for artificially intelligent systems. On the other side of this mutually beneficial relationship, we explore how machine learning methods can be employed to advance neuroscience knowledge. We show how neuromorphic chips and algorithms can be used to effectively model biological neural responses. The work proposed here is largely focused on how biologically plausible algorithms can be used, both in modeling and in signal processing efforts, to further our understanding of the brain at multiple levels. This work has the potential to improve our understanding of biological intelligence, providing further guidance towards building robust, generalizable, and truly intelligent machines.
Computer vision systems have made great improvements in recent decades, with neural network models reaching high accuracy on tasks like object detection. These discriminative models, however, are susceptible to attacks and make mistakes that elucidate problems with how these models understand their inputs. The general idea behind discriminative modeling, that of directly mapping input data to target outputs, seems to be the cause of these failures. With the ultimate goal of constructing artificial intelligence systems that perceive and interact with their environment in a more comprehensive manner, a shift in focus towards generative frameworks that model the underlying structures of the world through probability distributions is necessary.
This review suggests that generative models, as opposed to their discriminative counterparts, are better equipped to generalize, adapt, and perceive the world in a rich way, more similar to biological systems. A motivating factor of this stance lies in how the brain appears to be largely generative, with the majority of computations critical to perception and cognition relying on probability distributions learned over time. Psychological and neurophysiological evidence for this generative neural framework is presented. This review covers a subset of computational methods of generative modeling, including variational autoencoders, generative adversarial networks, and sparse coding. Further, the neurally plausible locally competitive algorithm for solving sparse coding is presented as an instance of how generative models can be translated into a more biologically plausible formulation. Directions for future work include reconstructing other generative models in a more biologically plausible manner and translating such models in their entirety to neuromorphic spiking chips.
PhD in Computer Science (expected 2024), Drexel University
Thesis (proposed): "Learning from the Brain: bridging the gap between biological and machine intelligence"
Candidacy: “Sight Unseen: Generative Models and Robust Computer Vision”
Advisor: Dr. Edward Kim
MS in Computer Science (2021), Drexel University
BS in Neuroscience (2019), Villanova University
S. Nesbit, A. O’Brien, J. Rego, G. Parpart, C. Gonzalez, G. T. Kenyon, E. Kim, T. C. Stewart and Y. Watkins. “Think Fast: Time Control in Varying Paradigms of Spiking Neural Networks”, International Conference on Neuromorphic Systems (ICONS) 2022.
G. Parpart, C. Gonzalez, T. C. Stewart, E. Kim, J. Rego, A. O’Brien, S. Nesbit, G. T. Kenyon and Y. Watkins. “Dictionary Learning with Accumulator Neurons”, International Conference on Neuromorphic Systems (ICONS) 2022.
E. Kim, M. Daniali, J. Rego, G.T. Kenyon. “The Selectivity and Competition of the Mind’s Eye in Visual Perception“, arXiv preprint arXiv: 2011.11167, 2021.
J. Carter, J. Rego, D. Schwartz, V. Bhandawat, E. Kim, “Learning Spiking Neural Network Models of Drosophila Olfaction“, International Conference on Neuromorphic Systems (ICONS) 2020.
E. Kim, J. Rego, Y. Watkins, G. T. Kenyon. “Modeling Biological Immunity to Adversarial Examples”, IEEE International Conference on Computer Vision and Pattern Recognition (CVPR) 2020.
RESEARCH and LAB EXPERIENCE
Intelligent Systems Lab, HRL Laboratories (Scientist III, 2022-present)
Computer and Computational Division (CCS-3), Los Alamos National Laboratory
Spiking and Recurrent Software (SPARSE) Coding Lab, Drexel University
Word Recognition and Auditory Perception Lab, Villanova University
Temporal Perception Lab, Villanova University
FELLOWSHIPS and AWARDS
Dean's Student Leadership and Service Award
Drexel University College of Computing and Informatics, 2023
Information Science and Technology Institute (ISTI) Student Fellow
Los Alamos National Lab, Summer 2021
LEADERSHIP and ORGANIZATIONS
Doctoral Student Association (Vice President), Drexel University College of Computing and Informatics
Upsilon Pi Epsilon, Honor Society for Computing and Information Disciplines, Drexel University
Board of Academic Integrity, Villanova University
Psi Chi, Psychology and Neuroscience Honors Society, Villanova University