My research focuses on developing robust AI systems that generalize across domains and translate theoretical advances into practical solutions for complex real-world problems. I am passionate about interdisciplinary research at the intersection of machine learning and neuroscience, with a particular focus on systems that must operate reliably under uncertainty. My work spans from bayesian methods in reinforcement learning to direct neural control in human patients, unified by the goal of creating AI that adapts and performs consistently across diverse, unpredictable environments.

Robust and Generalizable Reinforcement Learning

My work addresses a limitation in reinforcement learning (RL): agents that excel in training environments but fail to generalize to new situations. I developed Signal-to-Noise Ratio regulated Parameter Uncertainty Networks (SNR PUN), a novel Bayesian neural network regularization technique that significantly improves generalization in RL agents. Published at CVPR 2024, this method constrains parameter uncertainty to encourage learning of robust strategies rather than environment-specific solutions. The approach demonstrated superior performance across multiple domains and successful transfer to real-world robotic systems, demonstrating its potential for practical applications.

Neural Control in Visual Neuroprosthetics

Building on these generalization principles, I have applied deep learning methods to neural activity control. Working with a cortical visual prosthesis implanted in a blind participant, we demonstrated a deep learning-based control method of evoked neural responses in the human visual cortex. The ability to artificially control neural activity is a fundamental challenge for sensory neuroprostheses.

Traditional neuroprosthetic approaches rely on a simplistic mapping between stimulation sites and neural responses, ignoring the complex interactions within cortical networks. Instead, we trained deep neural networks to learn the nonlinear relationship between stimulation patterns and neural responses. This approach significantly outperformed conventional methods while requiring lower stimulation currents and producing more consistent visual perceptions. Additionally, I have worked on using spiking neural networks (SNNs) to generate temporally-varying stimulation patterns, demonstrating that temporal dynamics can enhance prosthetic vision while reducing computational requirements.

Adaptive AI for Personalized Applications

My recent work expands into personalized AI systems, particularly for individuals with disabilities. I am developing adaptive speech recognition systems for children with speech impairments, using bayesian low rank adaptation approaches to personalize foundation models. This approach leverages uncertainty quantification to tailor models to individual users, significantly improving recognition accuracy in challenging conditions. This research is part of a broader effort to create AI that adapts to the unique needs of each user, rather than relying on average performance metrics.

Future Research Directions

My future research focuses on advancing online adaptation and personalization in AI systems, with particular emphasis on:

• Closed-loop neural interfaces: Developing real-time adaptive systems that continuously optimize based on neural feedback
• Foundation models for neurotechnology: Adapting large-scale models for specialized biomedical applications
• Uncertainty-aware personalization: Building systems that leverage uncertainty quantification for individualized adaptation

As AI becomes increasingly integrated into healthcare and assistive technologies, my goal is to develop machine learning solutions that not only perform reliably across diverse conditions but actively adapt to serve the unique needs of each individual user.