Elisa Bu Sha, Kenny Daici, Grace DeCostanza, Aayush Gandhi, Vincent Jiang, Ian Lee, Emily Perez, Olivia Quiroga, Rithik Rajasekar, Simon Rodriguez, Bill Zhang, Kelly Zukowski
Advisor: Dr. Graham Cousens
Assistant: Molly Thompson
Abstract: The process of neuronal information transfer involves the propagation of electric signals known as action potentials, or spike trains, across nerve fibers in the central nervous system. Under controlled experimental conditions, it has been observed that even when presented with identical stimuli, there are a high level of trial-to-trial variability in single-unit (single neuron) spike trains. Even when comparing functionally similar neurons, the response reliability can differ considerably. It was hypothesized that the visual pathway neurons closer to photoreceptors would display greater response reliability and vice versa. Furthermore, it was hypothesized that neurons from subjects under anesthesia would also display greater response reliability compared to non-anesthetized neurons. This meta-analysis examines single-unit recordings from different species across several major sensory-processing regions of the brain. The data sets were analyzed through the calculation of the Fano factor in MATLAB 2020a and statistical analysis tests. It was determined that there was a correlation between anesthesia and response reliability, while the results for the relationship between the level of processing and response reliability were more inconclusive. These conclusions serve to further knowledge of spike train variability and its influencing factors.
Abbreviations: ISIs, interspike intervals; CLM, caudolateral mesopallium; CMM, caudomedial mesopallium; LGN, lateral geniculate nucleus; MS, medial septum; NCM, caudomedial nidopallium; PFC, prefrontal cortex; SI, primary somatosensory cortex; VI, primary visual cortex; V2, secondary visual cortex; VP, ventral pallidum
Keywords: action potentials, Fano factor, interspike intervals, single-unit recordings, spike trains, variability
INVESTIGATION OF BIAS IN MACHINE LEARNING THROUGH FACIAL RECOGNITION AND SKIN LESION CLASSIFICATION
Yuvanshu Agarwal, Chirag Furia, Ishan Hemmige, Kenneth Huang, Anqi Li, Eva Schiller, Adam Sher, Simon Sun, Lana Van Note, Edward Wang, Helen Xie, Angelina Xu
Advisor: Dr. Minjoon Kouh
Assistant: David Van Dongen
Abstract: Machine learning is a discipline of artificial intelligence concerned with the design, creation, and implementation of computer software that can learn autonomously. The purpose of this project was to evaluate both the versatility of artificial neural networks (ANNs) and their vulnerability to biases across multiple datasets. Three algorithms were constructed: a Convolutional Neural Network (CNN) using Keras and TensorFlow, a Multilayer Perceptron (MLP) classifier using neural net Scikit-Learn, and a Simplistic Nearest Centroid classifier for comparison. These algorithms were subsequently tested on three datasets: a homemade facial recognition dataset with images from NJGSS scholars, the UTKFace dataset of facial images, and the University of Edinburgh’s Dermofit skin lesion image dataset. Each dataset was tested with at least one of the neural networks, as well as the nearest centroid classifier. We used these algorithms and datasets to develop systems to recognize NJGSS scholars’ faces, classify UTKFace images according to gender and race, and classify Dermofit skin lesion images according to lesion type. These recognition and classification systems were subsequently used to investigate ethical issues in machine learning, in particular the bias created by the use of unbalanced datasets.
Cognitive Illusions: A Comparative Study of Their Effects on New Jersey Governor’s School Scholars and the General Public
Brooke Chapple, Lila DiMasi, Sara Elkilany, Annabelle Jin, Zachary Kim, Emma Morin, Ujjayi Pamidigantam, Krishna Parikh, Margaret Worchel, Yi-Shirley Xie, Krystal Yearis, Caroline Zhao
Advisor: Dr. Patrick Dolan
Assistant: Abby Pedroso
Abstract: Cognitive illusions are common thinking errors that can arise when the brain takes shortcuts, especially under time constraints or other stressful situations. The primary purpose of this study is to investigate whether or not high-performing math and science high school students in the New Jersey Governor’s School in the Sciences (NJGSS) program show differences in susceptibility to cognitive illusions in comparison to the general public. In this study, a survey was distributed to two distinct groups of participants: NJGSS scholars as well as users of Amazon’s Mechanical Turk (MTurk) platform. The survey was created to look like a personality test in an attempt to eliminate bias in the participants and distract from its true purpose. Pertinent responses relating to the 16 cognitive illusions studies were then analyzed. The results demonstrated the universality of cognitive illusions: both groups of individuals were inclined to make “unconscious inferences” that affected how they interpreted the tasks they were presented with. Although the NJGSS students are high achieving, they were just as susceptible to the illusions as the MTurk participants overall. Error analysis of the comparative study, in regards to sample size and the lack of requirements for the MTurk participants, can allow for more controlled experimentation in succeeding research. Future work includes studies on the influence of demographics, such as gender, age, and education level, on responsiveness to cognitive illusions.
Anthony DiMaggio, Jessica Dong, Deepak Gopalakrishnan, Julia Granato, Bryant Har, Rohan Kulkarni, Michelle Liu, Craig Mulhern, Jr., Arianna Otoo, Ishika Patel, Ansh Sharma, Ian Viegas
Advisor: Dr. Daniel Kaplan
Assistant: Matheus Macena de Carvalho
Abstract: The purpose of this work is to provide an overview of quantum computing and to describe and test some critical aspects of this technology with examples. All materials are based on the IBM Qiskit tool to develop quantum computing circuits. Quantum computers can be used to solve optimization problems much more efficiently than classical computers. Currently, real quantum computers experience decoherence errors. A linear relationship between the number of quantum gates and error was determined by varying the amount of CNOT and Hadamard gates separately. As of now, approaches to ameliorate these errors remain restricted due to the limited number of qubits on real quantum computers. With the VQE approach, a real-life application of quantum computers is being developed to solve a ground state analysis of simple molecules.
COMPUTATIONAL DRUG DISCOVERY FOR COVID-19 THROUGH IDENTIFICATION OF CANDIDATE FDA-APPROVED SMALL MOLECULES TARGETING 12 SARS-CoV-2 PROTEINS
Daniel Bourland, Neil Kadian, Aravind Krishnan, Anthony Montes de Oca, Camille Quaye, Vijay Ramu, Mannut Singh, Allie Tay, Jonathan Tenenbaum, Daniel Wise, Rachel Yan, Noah Youssef
Advisor: Dr. Adam Cassano
Assistant: Katie Revelas
Abstract: The greater spread and severity of the COVID-19 pandemic relative to previous outbreaks is underscored by SARS-CoV-2’s increased virulence and capacity for airborne transmission. Several structural and functional proteins responsible for host cell entry, expression of viral proteins, and post-replication exit are critical for the viral replication cycle. We performed a novel in silico drug discovery study analyzing binding interactions between 2,637 FDA-approved small molecule drugs and 12 of 24 total SARS-CoV-2 proteins. We used a combination of computationally predicted structures by C-I-TASSER as well as crystal structures. Docking was performed with the software SeeSAR. In addition to automatically predicted binding sites, we manually defined pockets not detected by SeeSAR. We selected ligands possessing low KD values, binding affinities lower than 100 nM, oral or IV routes of administration, and favorable lipophilic ligand efficiencies. Computational screens testing each drug in 2 to 10 poses revealed multiple promising ligands binding to several SARS-CoV-2 proteins with affinities below 100 nM; among these, cimetidine, deferoxamine, aliskiren, oxetacaine, unoprostone, and labetalol were the most promising, either due to an exceptionally high affinity or binding across multiple proteins. In addition, we performed molecular modifications of cimetidine and copanlisib to improve binding affinity or prevent adverse side effects. By the high-throughput evaluation of binding interactions with 12 SARS-CoV-2 proteins, we identified several promising ligands that can be advanced to in vitro and in vivo studies with the goal of developing novel therapies for COVID-19.