Virtual-Rat Gap Crossing

My Contribution

• Trained and diagnosed reinforcement-learning policies across 237 experimental runs
• Systematically varied 17 gap widths from 12 to 28 cm to characterize the task boundary
• Designed a correct-rejection reward and dense-to-sparse reward schedules for uncrossable-gap trials
• Added Gaussian noise to visual observations and varied starting positions; evaluations covered 6 noise levels, 3 visual conditions, and 3 starting distances
• Used controlled evaluation changes to distinguish learning failures from evaluation artifacts

Research Context

Summer Undergraduate Research Fellow in the Talmo Lab and Salk Center for AI and Research Computing, June–August 2026. Selected as 1 of 10 Salk SURF fellows.

The project shifted the question from whether the virtual rat could cross a gap to what information and reward structure supported the behavior. Gaussian visual-noise tests, starting-position changes, and uncrossable-gap conditions were used to investigate the strategy behind successful performance.

Presentation

Presented “Eliciting Deliberation Before a Gap Jump in a Virtual Rat” at the Salk SURF Capstone and Talmo Lab research meeting in August 2026.

Neural Oscillation and Social Interaction

My Contribution

• Analyzed VTA-PFC local field potential recordings from 24 mice using MATLAB and R
• Integrated coherence, phase-amplitude coupling, and Granger causality with mixed-effects modeling
• Quantified frequency-specific coupling and directional interactions between VTA and PFC
• Compared ethanol and baseline conditions in the theta band (4–12 Hz)

Methods & Tools

• Signal processing: Bandpass filtering, Hilbert transform, time-frequency decomposition
• Metrics: Power spectral density (PSD), coherence, phase-amplitude coupling (PAC), Granger causality
• Tools: MATLAB, Brainstorm, R, data visualization
• Workflow: Event-aligned epoching, automated batch processing, reproducible pipelines

Key Findings

Identified reduced theta-band VTA→PFC Granger causality under ethanol exposure while PFC→VTA connectivity remained comparatively stable. Presented at the Wesleyan Summer Experience Presentation in July 2025 and Research Pathways Showcase in January 2026.

What made this work especially interesting was that integrating behavior with neural analyses required more than stacking methods together; I had to keep revising the pipeline so the metrics remained interpretable in the context of the experiment rather than drifting into abstraction.

Reproducibility

• Data: Internal lab dataset (Melón Lab, Wesleyan University)
• Code: Available upon request / private GitHub
• Environment: MATLAB, Brainstorm, and R

Next Steps & Reflection

This project made directional communication between brain regions feel more informative than treating either region in isolation. It also showed me how strongly an interpretation depends on keeping signal processing tied to the behavioral and experimental context.

Interpretable ML for Creativity

My Contribution

• Analyzed 5,510 responses and 162,686 tokens across 3 creativity domains
• Independently developed an HPC-based SHAP pipeline using up to 500 model evaluations per response
• Tested associations between SHAP attributions and token frequency, semantic distance, length, and spelling
• Contributed analyses and figures to a manuscript in preparation

Illustrative token-importance snapshot

Interactive SHAP illustration

A novel response can feel precise without becoming too easy to predict.

Methods & Tools

• Interpretability: Token-level SHAP (SHapley Additive exPlanations)
• Models: 3 transformer-based automated originality-scoring models
• Data: 5,510 responses and 162,686 tokens across 3 creativity domains
• Tools: Python, SHAP, and HPC/Slurm
• Analysis: Associations with frequency, semantic distance, token length, and spelling

Impact

Presented “An Analysis of Automated Creativity Scoring Models Using Explainable AI” with A. Tran at the Data Insights Student Speaker Series in March 2026.

What changed in my thinking during this project was that interpretability stopped feeling like a secondary add-on; it became the main way to tell whether the model was actually learning something meaningful.

Reproducibility

• Data: Machine Psychometrics and Computational Creativity Lab dataset
• Code: Python scripts for token-level SHAP analysis and HPC job submissions
• Environment: Python and Slurm-based HPC

Next Steps & Reflection

This work highlights the critical need for interpretability in ML models applied to creative or subjective tasks. I'm particularly interested in how these methods can reveal the difference between genuine pattern recognition and dataset-specific artifacts, a distinction that matters across all ML applications in cognitive science.

COVID-19 Cough Classification

My Contribution

• Integrated 3 public respiratory-audio datasets
• Built a Python pipeline for audio-quality filtering, augmentation, and spectral feature extraction
• Developed a 430K-parameter CNN-BiGRU classifier
• Evaluated the model across 5 folds and 6,624 samples under substantial class imbalance

Methods & Tools

• Preprocessing: Audio-quality filtering, augmentation, and spectral feature extraction
• Model: CNN-BiGRU with approximately 430K parameters
• Validation: Five-fold evaluation using PR-AUC, recall, and F1
• Tools: Python and TensorFlow
• Focus area: Classification under substantial class imbalance

How the model “sees” a cough

Key Learnings

Achieved 0.926 PR-AUC, 90.6% COVID-positive recall, and 0.892 F1 across five folds and 6,624 samples.

One challenge here was learning not to trust aggregate performance at face value when the underlying data were substantially imbalanced.

Reproducibility

• Training: AI4ALL Ignite AI Accelerator (Sep 2025 to Mar 2026)
• Code: GitHub repository with version control
• Environment: Python 3.10, TensorFlow, librosa

Next Steps & Reflection

This project made me more careful about data imbalance, bias, and the difference between a promising model and a clinically meaningful one.

Neural Decoding of Movement Choices

My Contribution

• Collaborated in a four-person team to develop the decoder
• Preprocessed 567-dimensional MOp/MOs spike-train data aligned to three behavioral events
• Generated raster plots and PSTHs to characterize neural dynamics and create model inputs
• Implemented a three-stage LSTM pipeline in PyTorch

Methods & Tools

• Data: Cortical spike trains from MOp (primary motor cortex) and MOs (secondary motor cortex)
• Model: LSTM (Long Short-Term Memory) recurrent neural networks
• Optimization: Backpropagation, regularization, batch normalization
• Tools: PyTorch, Python, NumPy

Training Context

Completed as a Deep Learning Scholar at Neuromatch Academy in August 2025.

What kept me engaged was that decoding intention from temporal neural activity required thinking simultaneously about model design, data structure, and the timing of the behavior itself, rather than treating prediction as a purely engineering problem.

Overview of the LSTM decoding pipeline linking cortical spike trains to left/right movement decisions.

Raster plot animation illustrating neural population activity around movement initiation.

Reproducibility

• Data: Neuromatch Academy dataset (cortical recordings)
• Code: PyTorch implementation with training/validation pipelines
• Environment: Python 3.10, PyTorch, NumPy

Next Steps & Reflection

This project sharpened my interest in temporal neural data and in models that stay close to the structure of the behavior they are trying to decode.

Capsaicin Anxiolytic Effects

My Contribution

• Established 21-day chronic restraint stress mouse model
• Conducted behavioral assays: Elevated Plus Maze (EPM) and Open Field Test
• Performed ROS fluorescence analysis of brain tissue
• Integrated behavioral and molecular readouts to assess capsaicin effects

Methods & Tools

• Animal model: 21-day chronic restraint stress in mice
• Behavioral assays: Elevated Plus Maze (EPM), Open Field Test
• Biochemical analysis: ROS (Reactive Oxygen Species) fluorescence
• Study design: Comparison across stress, treatment, and control conditions

Key Findings

Combined behavioral assays with ROS fluorescence measurements to evaluate whether capsaicin could mitigate anxiety-like behavior and oxidative stress in a chronic-stress mouse model.

This project taught me early that careful experimental design matters most when the result is not guaranteed to be clean; the interpretation only became convincing when the behavioral and biochemical pieces could speak to each other.

Dissemination

• Presented at the Jiangsu High School Students’ Science Symposium
• Communicated findings to provincial-level audience through poster presentation
• Audience included high school students and university faculty

Presenting my work at the Jiangsu High School Students’ Science Symposium through a poster presentation.

Reproducibility

• Lab: Zhuang Lab, Nanjing University (Mar 2022 to Oct 2022)
• Analysis: Behavioral and ROS fluorescence readouts
• Protocol: Standardized stress induction and behavioral testing

Next Steps & Reflection

This early research experience introduced me to the importance of rigorous experimental design and statistical analysis in neuroscience. It sparked my interest in understanding the biological mechanisms underlying behavioral change, a theme that continues in my current computational work on neural circuit dynamics.

Energetic Constraints on the p53 Cancer Mutational Landscape

Thesis Context

• Advisor: Kelly M. Thayer
• Degree: Bachelor of Arts, expected May 2027
• Focus: Energetic constraints on the p53 cancer mutational landscape

This project brings together my interests in computational modeling and biological mechanism. The analysis is ongoing, so the site describes the research question and approach without presenting preliminary results as conclusions.

Boolean Genetic Networks

My Contribution

• Designed and implemented Boolean network simulations (Kauffman nets)
• Developed mutation sensitivity analysis across different attractor states
• Optimized memory usage and computational performance for larger local runs
• Experimented with basic parallelization strategies on local hardware

What I found compelling here was how quickly a simple model can become difficult to reason about at scale, which made computational efficiency and conceptual clarity feel like part of the same problem.

Methods & Tools

• Model: Boolean genetic regulatory networks (Kauffman networks)
• Analysis: Attractor identification, mutation sensitivity, robustness measures
• Optimization: Memory-efficient data structures and exploratory parallel processing on local machines
• Tools: Python, NumPy, computational biology modeling

Key Findings

Different attractor states showed varying degrees of sensitivity to mutations, suggesting that network topology and current state jointly determine robustness. Some attractors were remarkably stable, while others were highly vulnerable to single-node perturbations. This has implications for understanding biological robustness and evolvability.

Reproducibility

• Code: Python simulation framework with configurable network parameters
• Environment: Python 3.10, NumPy, local development environment

Next Steps & Reflection

This project taught me the importance of computational efficiency in large-scale simulations, even when running on local machines. What started as a straightforward modeling question became an exercise in thinking about optimization and potential future scaling. I'm interested in extending this approach to more realistic network topologies informed by empirical gene regulatory data, and exploring how network structure shapes evolutionary dynamics.