CS @ UC Irvine · AI/ML · Neurosymbolic AI · Reasoning · Secure ML Inference
About
🧠 I'm an undergraduate researcher in Computer Science at UC Irvine ('26). I work on neuro-symbolic AI, brain‑inspired learning (HDC), multimodal models, and secure inference in neural networks (CKKS‑FHE). I'm very grateful for the opportunity to work with Prof. Mohsen Imani for my undergraduate research.
🔬 Current: BiasLab @ UCI
🏂 Hobbies: Snowboarding, music, gaming
🎵 Favorite Artists: D'Angelo, Dijon, Mkgee (Corresponding fave albums below)
🎮 Favorite Game: Cyberpunk 2077
With my cats Coconut (left)
& Kumquat (right) 🐱🐱 Click photo for more cat pics! 📸
Research
Brain-Inspired Reasoning under Homomorphic Encryption
Privacy-preserving neurosymbolic framework executing inference entirely under CKKS-FHE while preserving HDC-based reasoning robustness. Maintains >90% accuracy on encrypted graph inference with 4× reduction in bootstrapping overhead through noise-adaptive scheduling. Lead author; rebuttal completed.
FHE · HDC · Neurosymbolic AI · Privacy-Preserving ML
Using Vector Symbolic Architecture (VSA) to build generalizable world models with learned group structure. Achieved 87.5% zero-shot accuracy and 4× noise robustness. Contributed to ablation studies with world model grid.
Extending CLIP's zero-shot capabilities to event-based vision, bridging asynchronous event data with image-text representations. Achieved +19% zero-shot accuracy over prior methods with cross-modal retrieval across five modalities. Contributed attention visualization pipeline.
First universal HDC encoding framework that adapts dynamically between learning and cognitive reasoning. Achieves 95% learning accuracy with correlated encodings and 100% decoding accuracy under exclusive encodings. Quantifies correlation-separation trade-off for brain-like reasoning.
AdamsFoods Wholesale App — Full-stack JavaScript app with React client and Node.js/Express server. Features product catalog, checkout system, AWS S3 integration, and JWT authentication.
code · live website ·
AdamsFoods Inventory Management System — React frontend for back-office stock tracking with item CRUD, search/filter, and low-stock alerts.
code ·
Safe UAV Landing for U.S. Navy — Custom pose estimation & reasoning model for safe UAV landing in adverse weather. Reduces casualty rates through intelligent deck clearance detection.
Crash Anticipation — VideoMAE model achieving 100% on AP, F1, and lead-time recall with real-time inference capability. Future: neurosymbolic reasoning for proactive collision avoidance.
code ·
Safe UAV Landing — Demo
Crash Anticipation — Demo
Teaching & Service
Learning Assistant — ICS 33 (Spring 2025)
Volunteer web dev for non‑profits (Feeding Pets of the Homeless)
Robust Reasoning and Learning with Brain-Inspired Representations under Homomorphic Encryption
Privacy-Preserving AI · Neuro-Symbolic Learning · Homomorphic Encryption · HDC
IEEE Transactions on Artificial Intelligence (2026, under review)
We propose a privacy-preserving neurosymbolic framework that executes inference entirely under CKKS Fully Homomorphic Encryption (FHE) while preserving the robustness and interpretability of Hyperdimensional Computing (HDC)-based reasoning. The framework introduces distributed bootstrapping to mitigate ciphertext-noise explosion across deep layers, balancing accuracy and compute cost.
Key Results
Maintains >90% accuracy on encrypted graph inference without plaintext access
4× reduction in bootstrapping overhead through noise-adaptive scheduling
Preserves symbolic structure of neural states under encrypted operations
Robust performance across multiple graph-based reasoning tasks
Why Homomorphic Encryption?
Cloud-based AI inference poses significant privacy risks when handling sensitive data (medical records, financial information, biometric data). Traditional encryption requires decryption before processing, exposing data to potential breaches. Fully Homomorphic Encryption enables computation directly on encrypted data, ensuring privacy even in untrusted cloud environments.
FHE's security advantage: Data remains encrypted throughout the entire inference pipeline in untrusted cloud environments
Framework Overview
Our end-to-end neurosymbolic pipeline integrates HDC-based symbolic reasoning with CKKS-FHE operations. The framework employs distributed bootstrapping at strategic network layers to prevent noise accumulation while minimizing computational overhead. Graph Neural Networks operate entirely on encrypted representations, with symbolic decoding preserving interpretability.
End-to-end neuro-symbolic FHE pipeline with distributed bootstrapping and symbolic decoding
Noise-Adaptive Bootstrapping
A critical challenge in FHE-based deep learning is ciphertext noise accumulation. Without proper noise management, encrypted operations cause accuracy to collapse as noise overwhelms the signal. Our distributed bootstrapping scheme strategically refreshes ciphertexts at high-noise layers, achieving a 4× reduction in bootstrapping overhead compared to naive per-layer approaches.
Empirical analysis: Noise growth without bootstrapping leads to accuracy collapse. Our approach balances accuracy preservation with computational efficiency
Technical Innovation
SEAL-Based Implementation: Built on Microsoft SEAL library for CKKS operations
Noise Characterization: Empirical analysis of real noise distributions in encrypted GNN operations
Distributed Bootstrapping: Adaptive scheduling reduces compute cost by 75% while maintaining accuracy
Symbolic Preservation: HDC-based representations maintain interpretability under encryption
Graph Reasoning: Encrypted inference on graph-structured data for video anomaly detection
Applications
The framework enables privacy-preserving AI for sensitive domains:
Healthcare: Encrypted medical diagnosis without exposing patient data
Finance: Fraud detection on encrypted transaction graphs
Surveillance: Privacy-preserving video anomaly detection in public spaces
IoT Security: Encrypted reasoning on edge device data
My contribution: Lead author and corresponding author. Designed and conducted SEAL-based experiments to characterize real noise distributions and evaluate CKKS performance under GNN operations. Implemented noise injection mechanisms and distributed bootstrapping schemes within the graph network to achieve robust encrypted reasoning. Led rebuttal process for IEEE TAI 2026.
Geometric Priors for Generalizable World Models via VSA
VSA · World Models · Generalization
NeurIPS 2025 Workshop on Symmetry and Geometry in Neural Representations
We introduce a world modeling framework grounded in Vector Symbolic Architecture (VSA) with learnable Fourier
Holographic Reduced Representation encoders and group‑structured latent transitions via element‑wise complex
multiplication, enabling composition directly in latent space.
Key Results
87.5% zero‑shot accuracy on unseen state‑action pairs
53.6% higher accuracy on 20‑timestep rollouts vs. MLP baseline
4× robustness to noise
Structured State Representations
Our VSA-based model (FHRR) learns structured, grid-like state embeddings that preserve geometric relationships, while traditional MLP models produce unstructured representations.
FHRR (VSA): Grid-like structured embeddings preserve spatial relationshipsMLP: Unstructured embeddings with no clear geometric pattern
Superior Generalization
VSA (FHRR) demonstrates significantly better generalization on held-out data across different data holdout ratios, maintaining high accuracy even when large portions of training data are withheld.
Rollout accuracy vs. data holdout ratio: VSA maintains strong performance while MLP degrades rapidly
Robustness to Noise
The VSA approach exhibits greater robustness to Gaussian noise compared to MLP baselines, maintaining near-perfect accuracy under significant noise perturbations.
1-step accuracy vs. Gaussian noise: VSA maintains >80% accuracy even with noise std of 5, while MLP drops to ~20%
My contribution: ablation studies with the world model grid. Post‑submission: MiniGrid experiments.
We extend CLIP's zero-shot and text-alignment capabilities to the event modality, developing a robust encoder that bridges asynchronous event data with image-text representations. The framework aligns event embeddings with CLIP's visual and textual spaces, preserving zero-shot generalization and enabling cross-modal applications across five modalities (Image, Event, Text, Sound, Depth).
Our visualization pipeline extracts attention heatmaps and relevance maps from the trained encoder, illustrating how CLIP's transferred attention captures salient event-specific regions. These visualizations demonstrate that the encoder successfully learns to focus on semantically meaningful areas in event data.
Attention heatmap: Peacock (ImageNet) - Model attends to distinctive features like tail feathers and body structure
Attention heatmap: Gramophone (N-Caltech) - Event encoder focuses on characteristic horn and base components
Attention heatmap: Fight scene (UCF-Crime) - Model captures action regions for video anomaly detection
Technical Approach
The framework employs contrastive learning to align event representations with CLIP's pre-trained visual and textual embeddings. By leveraging CLIP's powerful multi-modal alignment, the encoder inherits zero-shot capabilities while learning event-specific features through asynchronous temporal patterns.
Cross-Modal Applications
Zero-Shot Classification: Direct transfer of text-based class descriptions to event data
Cross-Modal Retrieval: Query across image, event, text, sound, and depth modalities
Video Anomaly Detection: Event-based detection without task-specific training
Multi-Modal Fusion: Combine complementary information from different sensor types
My contribution: Supported the visualization pipeline by extracting attention heatmaps and relevance maps from the trained encoder, demonstrating how the model attends to salient event-specific regions across different datasets (ImageNet, N-Caltech, UCF-Crime).
Optimal Hyperdimensional Representation for Learning and Cognitive Computation
This work introduces the first universal HDC encoding framework that adapts dynamically between learning and cognitive reasoning tasks. By modulating the correlation of hypervectors through tunable kernel scales, the model optimizes both memorization capacity and decodability, bridging the gap between symbolic reasoning and pattern learning within a unified high-dimensional representation.
Key Results
Up to 95% learning accuracy with correlated encodings vs. 65% in uncorrelated baselines
100% decoding accuracy under exclusive encodings for cognitive tasks
Quantitative formulation of the correlation–separation trade-off that governs brain-like reasoning
Universal framework adaptable to both symbolic and pattern-based tasks
The Learning-Cognition Dichotomy
HDC systems face a fundamental trade-off: learning tasks benefit from correlated hypervector representations that enable generalization, while cognitive reasoning requires exclusive (orthogonal) encodings for symbolic manipulation. Traditional HDC methods optimize for one regime at the expense of the other. Our framework dynamically adapts encoding parameters to achieve optimal performance across both domains.
Learning Accuracy Analysis
We conducted comprehensive parameter sweeps to characterize how encoding dimensions (D) and kernel scales (w) affect learning performance. The heatmaps reveal distinct optimal regions: correlated encodings (smaller w values) dramatically improve learning accuracy, especially at lower dimensions where orthogonal encodings fail.
Learning accuracy (top) and separation (bottom) as functions of dimension D and kernel scale w. Correlated encodings (low w) achieve 95% accuracy, while maintaining measurable separation for downstream reasoning.
Decoding Accuracy Analysis
For cognitive tasks requiring symbolic manipulation, decoding accuracy depends on the exclusivity of hypervector encodings. Our analysis shows that higher kernel scales (larger w) produce more orthogonal representations, achieving 100% decoding accuracy. The 2D parameter space reveals the sweet spot where encodings are sufficiently exclusive for symbolic operations yet not overly sparse.
Decoding accuracy as functions of D and w (left) and wx vs wy (right). Exclusive encodings (high w) enable perfect symbolic decoding, demonstrating the cognitive reasoning capability of the framework.
Signal-to-Noise Ratio Characterization
To formalize the correlation-separation trade-off, we analyzed the signal-to-noise ratio (SNR) distributions for both same-class and different-class hypervector pairs across varying dimensionality. The histograms reveal clear separation patterns: at higher dimensions (p=100), same-class pairs maintain high correlation while different-class pairs remain decorrelated, enabling both generalization and discrimination.
SNR distributions for same-class (blue) and different-class (orange) hypervector pairs at p=10, 50, 100. Higher dimensionality provides better class separation while maintaining within-class correlation for generalization.
Technical Contribution
Universal Encoding Framework: Single parameterized system adaptable to learning and cognitive tasks
Theoretical Foundation: Mathematical formulation of the correlation-separation trade-off
Empirical Validation: Comprehensive parameter sweeps across dimensions and scales
Brain-Inspired Design: Mimics dual-process reasoning in biological neural systems
Applications
The adaptive HDC framework enables:
Hybrid AI Systems: Seamlessly combine symbolic reasoning with pattern learning
Edge Computing: Lightweight, efficient representations for resource-constrained devices
Cognitive Architectures: Brain-inspired models balancing memory and reasoning
Neuromorphic Hardware: Optimized encodings for analog compute substrates
My contribution: Conducted decoding analyses and developed learning-accuracy heatmaps to visualize how encoding parameters affect separability and generalization. Collected and analyzed the signal-to-noise ratio distributions that support the theoretical derivation of optimal kernel scales for learning and cognition.
AdamsFoods Wholesale App
Full-Stack · React · Node.js · AWS S3 · JWT Auth
A comprehensive full-stack JavaScript application built with a React client and Node.js/Express server. The application serves as a wholesale platform with robust product management and user authentication.
Key Features
Product Catalog: Complete product browsing and management system
Shopping Cart & Checkout: Fully functional cart system with checkout workflow skeleton
AWS S3 Integration: Seamless media and document workflows using signed URL uploads with intelligent caching
JWT Authentication: Secure token-based authentication with role-based access control
Admin Routes: Protected admin endpoints with role guards for sensitive operations
REST API: Well-structured REST endpoints organized under /server
A React-based frontend application (Create React App) designed for back-office inventory tracking and stock management. Deployed on Vercel for reliable, fast access.
Computer Vision · Pose Estimation · Reasoning · CUI Dataset
Through BiasLab & Prof. Mohsen Imani, I collaborated with the U.S. Navy to develop a custom computer vision system combining pose estimation and symbolic reasoning for safe autonomous UAV landing operations.
Problem Statement
Traditional UAV landing systems rely on legacy fixed-pattern optical markers, which perform poorly in adverse weather conditions, leading to increased casualty rates and mission failures. The Navy needed a more robust solution that could operate safely in challenging environments.
Solution
Pose Estimation Model: Custom-trained model for accurate UAV positioning relative to landing deck
Reasoning Module: Intelligent system that delays landing when crew members or obstacles are detected on the landing spot/deck
Weather Robustness: Significantly improved performance in bad weather conditions compared to legacy systems
Real-time Processing: Fast inference for time-critical landing decisions
Impact
The system reduces casualty rates by preventing landings when the deck is not clear, providing a crucial safety layer for naval aviation operations. Worked on CUI (Controlled Unclassified Information) dataset under appropriate clearances.
Crash Anticipation
Computer Vision · VideoMAE · Autonomous Driving · Neurosymbolic AI
A cutting-edge video analysis system for predicting vehicle crashes before they occur, with future integration of neurosymbolic reasoning for actionable collision avoidance commands.
Current Performance (VideoMAE Baseline)
100% Accuracy: Perfect scores on AP (Average Precision), F1, and lead-time recall metrics
Real-time Inference: Capable of real-time processing with an NVIDIA 5070 Ti GPU
High Reliability: Demonstrated robust crash prediction across diverse scenarios
Current Limitations
The VideoMAE model's computational cost is too high for practical deployment on mobile devices or embedded systems in vehicles. This limits real-world applicability despite excellent performance.
Next Steps: Model Compression
Exploring smaller, more efficient models such as MobileNetV2.0 that can run on resource-constrained devices while maintaining similar performance levels. Goal: Enable deployment on standard automotive hardware.
Future Direction: Neurosymbolic Reasoning
Beyond simple crash prediction, the vision is to integrate a symbolic reasoning module that can:
Provide Direct Commands: Issue specific avoidance instructions (e.g., "turn left 15°", "brake immediately", "accelerate to avoid rear collision")
Infer Vehicle Dynamics: Calculate incoming vehicle speed, trajectory, and direction in real-time
Reason About Scenarios: Use symbolic reasoning to evaluate multiple avoidance strategies and select optimal actions
Proactive Collision Avoidance: Transform from reactive warning system to proactive safety system that prevents crashes through intelligent intervention
This neurosymbolic approach combines the perceptual power of deep learning with the interpretability and logical reasoning of symbolic AI, aligning with my broader research focus on neurosymbolic reasoning systems.