Frederick Hayes III, MIDS

Synthetic Dasein: Neuromodulated Plasticity

Modeled the "metabolic cost" of cognition in dynamic GridWorld environments. By utilizing phase-dependent gating of Spike-Timing-Dependent Plasticity (STDP), I successfully handed over agent control to a Spiking Neural Network (SNN), reducing compute overhead from 1800 to 90 units per episode.

Key Focus Areas:

• Spiking Neural Networks & Continuous-time Architectures
• Metabolic Loss Functions & Learning Dynamics
• Oscillatory Coordination

View the Repository on GitHub

NeuroBeacon: Calibrated RL for Cognitive Augmentation

Developed a Reinforcement Learning agent specifically designed for cognitive augmentation in human-in-the-loop systems. Instead of optimizing for raw accuracy, this project prioritized calibration (ECE), ensuring the system could transparently flag uncertainty and "know when it doesn't know."

Key Focus Areas:

• Deep Reinforcement Learning (DQN/PPO)
• AI Calibration (Expected Calibration Error)
• Transparent Decision-Making Architecture

View UCB Capstone Showcase Page

Audience-Adaptive AI Agents (RAG & LangGraph)

Built a dual-audience Retrieval-Augmented Generation system using LangGraph for dynamic sub-question decomposition. To rigorously evaluate interaction quality and mitigate hallucinated generalizations, I developed the "Fred Score," a composite metric balancing fluency, semantic match, and retrieval alignment.

Key Focus Areas:

• Agentic Orchestration & Multi-Step Reasoning
• Custom Evaluation Metrics (BLEURT, Hallucination Detection)
• Audience-Aware Prompting

Read the Final Paper Here

When Spikes Remember: Compositionality in Local Learning

Investigated Spiking Neural Networks as a biologically plausible alternative to standard Transformers. Demonstrated that local plasticity rules (STDP) can form compositional memories without the need for global backpropagation, addressing the biological implausibility of standard Deep Learning error signals.

Writing & Publications

I frequently write about the intersection of AI mechanism and philosophy on my column, Let's Talk About the Work.

• "When AI Dreams in Waves" (2025): An analysis of Continuous Thought Machines and emergent neural timing.
• "One Model, Many Voices: Measured" (2025): A technical analysis of audience-adaptive agents and the proposal of the "Fred Score" (Officially adopted into the UC Berkeley graduate Generative AI curriculum).