Griffin Holt

Projects

• Approaching the Floor Layout Problem with Genetic Algorithms and Convex Optimization, EE 364B, Convex Optimization II, Stanford University (4/3/2023 - 6/12/2023)

  In this paper we present a preliminary method of optimizing a floor layout given a set of desired proximity rela- tions between a number of rooms or cells. The unconstrained problem is non-convex; however, if a relational constraint is given for each pair of cells, then the problem becomes convex and can be solved using standard convex optimization methods. The problem then becomes how to search this combinatorially large space to restrain the original problem. We first utilize the Fruchterman-Reingold algorithm, which models the cells as point masses connected to one another by springs with spring constants that correspond to the corresponding proximity relation. The equilibrium state of this spring-mass system can then be sampled to produce relational constraints. We can further leverage this to create a population input to a genetic algorithm, which can be used to find more optimal solutions in relation to the objective function. We find in the end that this procedure gave us a solution that reaches the global optimum of our objective for our particular problem instance.

  Location

  Stanford, CA

  Collaborators

  • Andrew Zhang, Masters Student in Electrical Engineering, graduated Summer 2023, School of Engineering

  For More Information:

  • Approaching_the_Floor_Layout_Problem_with_Genetic_Algorithms_and_Convex_Optimization.pdf
  • Github Repository
• Spicing it up with with SpicyNERF: Novel View Synthesis using Thermal Images, CS 231N, Deep Learning with Computer Vision, Stanford University (4/24/2023 - 6/8/2023)

  In this report, we present SpicyNeRF, a preliminary method of applying and adapting Neural Radiance Field (NeRF) methods to generate novel thermal views of a scene. As a baseline, we separately run a variation of the vanilla NeRF model Nerfacto on a paired set of both thermal and traditional RGB images of the same scene, where each thermal photo corresponds to an RGB photo taken simultaneously from the same device. By comparing the results, we found that the first major hurdle to the NeRF pipeline was an accurate estimation of the parameters of the thermal camera used. This paper explores different methods to overcome this. We first explored parameter estimation methods using classical computer vision techniques. Specifically, we extracted the extrinsics of the RGB photos and directly mapped them to the thermal photos since they are taken from the same view. The thermal camera intrinsics were then estimated separately using COLMAP; however this was shown to be untenable. We then instead adopt a neural solution which makes the thermal camera extrinsics and intrinsics parameters to be learned during NeRF training. We show that this is a massive improvement over classical techniques. However, there still exist many limitations as the neural net implementation struggles to train on 360-degree view datasets, causing errors in the rendering. Over this, we finally present SpicyNeRF, which leverages a presupposed corresponding RGB dataset to create better initializations of the extrinsics of the thermal images for better training.

  Location

  Stanford, CA

  Collaborators

  • Andrew Zhang, Masters in Electrical Engineering, graduated Summer 2023, School of Engineering

  For More Information:

  • Spicing_it_up_with_SpicyNeRF_Novel_View_Synthesis_using_Thermal_Images.pdf
• GAN-BERT for Automated Essay Scoring, Stanford University, CS 224N, Natural Language Processing (2/14/2023 - 3/3/2023)

  Every year, millions of individuals take English language proficiency exams, such as TOEFL and IELTS, for professional and academic development. These exams are typically graded by human evaluators; automating the evaluation process can improve both efficiency and fairness of the examinations. Our approach to the Automated Essay Scoring (AES) task is to implement three variations of the GAN- BERT architecture: a feed-forward neural network generator; a BERT transformer generator; and a generator composed of a fine-tined GPT2 language model in tandem with a BERT transformer. We use a single pre-trained RoBERTa model, fine-tuned to our task and dataset, for a baseline comparison. All three GAN-BERT architectures outperformed the baseline model on the test set. The GAN-BERT models are also able to better differentiate between Low and Medium score essays, and Medium and High score essays. The GPT2-BERT generator demonstrated the most evidence of taking advantage of the competitive nature of the GAN structure to improve both generator and discriminator.

  Location

  Stanford, CA

  Collaborators

  • Theodore Asa Kanell, Masters Student in Computer Science, admitted Autumn 2020, School of Engineering

  For More Information:

  • GAN-BERT_for_Automated_Essay_Scoring.pdf
• Predicting Justin Verlander's Next Pitch with Machine Learning, Stanford University, CS 229 Machine Learning (10/5/2022 - 12/7/2022)

  Given a set of input features (including game situational data, the pitcher’s pitching statistics, the batter’s hitting statistics, and the results of previous pitches), we will attempt to classify the next pitch Verlander will throw from the set of VPT.

  Location

  Stanford, CA

  Collaborators

  • Mohamed A Owda, Masters Student in Computer Science, admitted Autumn 2020, School of Engineering

  For More Information:

  • Predicting_Justin_Verlanders_Next_Pitch_with_Machine_Learning.pdf
• Drafting the Best Baseball Team: An Integer Programming Problem, Brigham Young University, C S 412 Linear Programming (2/1/2022 - 4/10/2022)

  We aim to draft a 25-man roster out of all the available players from the 2019 season of Major League Baseball (MLB) that meets the constraints of a typical MLB baseball team and maximizes the total sum of the bW AR scores of all players on the roster. We formulate this goal as an Integer Programming problem and solve it using both a Real-Valued Relaxation approximation method and the Branch-and-Cut method. The Real-Valued Relaxation algorithm returns almost optimal solutions, differing less than 1 point in the optimal objective values returned by the Branch-and-Cut algorithm. Simulations of 162-game seasons show that the teams yielded from solving the Integer Programming problem achieve record- breaking winning percentages–.846 for a team with a large pay- roll cap of U.S.$197,683,216, .716 for a team with a small payroll cap of U.S.$28,229,108–thus demonstrating the effectiveness of both the bWAR score and the model of Integer Programming as a method for making informed drafting decisions.

  Location

  Provo, UT

  For More Information:

  • Drafting_the_Best_Baseball_Team.pdf
• Recurrent Neural Networks for Identifying Phases of the Honeybee Waggle Dance, Brigham Young University, C S 474 Deep Learning (10/1/2021 - 12/14/2021)

  In the 1920s, Karl von Frisch conducted a series of ex- periments regarding the behavior of honeybees [1]. One of Frisch’s most significant discoveries was that of the meaning and interpretation of the honeybee “waggle dance”: forager honeybees use geometry through a cyclic dance to communi- cate information about located food sources [2]–[4]. In this report, I describe my attempts to utilize Recurrent Neural Networks (RNNs) and their variants to identify two different portions of this waggle dance: the run and the return phase.

  Location

  Provo, UT

  For More Information:

  • RNNs_for_Waggle_Dance.pdf
• Music Genre Classification, Brigham Young University, C S 472 Machine Learning (7/1/2021 - 8/31/2021)

  The objective of the experiments is to develop a machine-learning model to classify the genre of a 10-second clip of a song. The possible genres are: Blues, Classical, Country, Disco, Hip-hop, Jazz, Metal, Pop, Reggae, and Rock. The dataset is pulled from the GTZAN public music reposi- tory; in total, 73 different input features were ex- tracted from 3000 music samples. Hyperparam- eter searches, feature reduction experiments, for- ward selection wrapper experiments, and additional fine-tuning were conducted to improve the classi- fication accuracies of nine different models. The top three performing models were, in order: a Gra- dient Boost ensemble (86.20%), a Random Forest ensemble (85.70%), and a Multi-layer Perceptron (85.17%). The ‘Classical‘ genre was the easiest to classify; the ‘Rock‘ genre was the most difficult to classify.

  Location

  Provo, UT

  For More Information:

  • Music_Genre_Classification.pdf
• Ant Colony Optimization: An Advanced Approach to the Traveling Salesman Problem, Brigham Young University, C S 312 Algorithm Design & Analysis (10/15/2020 - 12/14/2020)

  The Traveling Salesman Problem (TSP) is a widely studied computational problem. This paper walks through the process of designing and implementing an Ant Colony Optimization (ACO) algorithm to solve a TSP. We discuss the natural phenomenon behind ACO and how this is turned into an algorithm via pseudocode. The process of parameter selection, including a section in which each of the five variables ��, ��, ��, k, and Q, is highlighted with tests showing that these variables will produce locally optimal results (as far as our limited testing proves) when the cost function c = f(��, ��, ��, k, Q) is set to c = f(.8, 2, .2, 100, 1000). We compare ACO to both a Greedy and Branch-and-Bound approach to solving TSP and find that for problem size n ≤ 15, Branch-and-Bound performed slightly better than our ACO algorithm (82.47% of Greedy vs 85.62%). However, ACO performs significantly better (≥ 5% better) than Branch-and-Bound in problem sizes approximately 20-150. After 150, the two algorithms begin to converge to the Greedy algorithms results. Time and space complexity of both Greedy and ACO algorithms are discussed with the time complexity being O(n3) and O(wkn2) respectively, where n = number of cities, k = number of ants, and w = the rate of the ACO implementation’s convergence.

  Location

  Provo, UT

  For More Information:

  • Ant_Colony_Optimization.pdf
• Question Detection using Decision Tree Models, Brigham Young University, C S 580 Theory of Predictive Modeling (10/23/2020 - 12/15/2020)

  Location

  Provo, UT

  For More Information:

  • Question_Detection.pdf
• Training Data and its Inherent Biases, Brigham Young University, C S 404 Ethics in Computer Science (10/1/2020 - 12/12/2020)

  A large sphere of the field of artificial intelligence is dominated by the use of machine learning to produce models that reflect reality within some degree of accuracy. Biases introduced at the stages of data collection, data annotation, and data cleaning can result in the amplification of pre- existing social injustices. Artificial intelligence researchers have a duty to identify and mitigate these biases in order to promote fairness and truthfulness in their models.

  Location

  Provo, UT

  For More Information:

  • Training_Data_Biases.pdf
• Civil War in the Home, Brigham Young University, ENGL 316 Technical Communication (2/1/2020 - 4/21/2020)

  Location

  Provo, UT

  For More Information:

  • Civil_War_in_the_Home.pdf