Skip to content

Welcome to the Trustworthy Knowledge-Driven AI (TKAI) Lab

At the Trustworthy Knowledge Driven AI (TKAI) Lab, we combine research ideas derived from formal methods, linguistics, cognitive science, and machine learning to efficiently build intelligent systems that are trustworthy, ethical, and secure.

These models should have a few desired properties, such as robustness and interpretability, such that humans can easily understand the generated information. This means designing agents that are expected to learn desirable behavior with minimal supervision and data that are provably stable and generalize to unseen distribution. We believe these objectives can be stably achieved by designing knowledge-guided, neuro-symbolic agents.

Our Research

Neuro-Mimetic AI

Neuro-Mimetic Approaches

We design brain-inspired computational models that serve as stable, backpropagation-free alternatives to traditional deep learning.

  • Predictive Coding: Self-correcting hierarchical prediction networks that minimize local errors.
  • Local Representation Alignment: Algorithms replacing global gradients with local target representations.
  • Memory-Augmented RNNs: Stable architectures that generalize to longer sequences with finite precision.
Predictive Coding Model Structure

Key Publications

AI for Science

AI for Scientific Discovery

We leverage machine learning to accelerate discoveries and solve complex optimization challenges in physical and biological systems.

  • Enzyme-Substrate Affinity: Using protein language models and graph neural networks to predict kinetics (Kcat).
  • Physics-Informed Neural Networks: Design of predictive-coding-based PINNs to capture high-frequency oscillations.
Enzyme Kinetics & RealKcat

Key Publications

Neuro-Symbolic LLMs

Reasoning & Generalization

We bridge connectionist language models with symbolic systems to construct interpretable, robust, and provably stable agents.

  • Local RetoMaton: Task-adaptive Weighted Finite Automata (WFA) replacing global search for structured retrieval.
  • LLM Generalization: Minimum Description Length (MDL) and curvature analysis to optimize model selection.
Local RetoMaton Architecture

Key Publications

Theoretical Foundations

Theory & Physical Constraints

We study the mathematical limits of computation and information theory under realistic thermodynamic constraints.

  • Surprisal-Rényi Free Energy (SRFE): A risk-sensitive free energy interpolating forward and reverse KL.
  • Realizable Circuit Complexity: Circuit design constraints modeling heat dissipation and physical space routing.
Realizable Circuit Complexity Diagram

Key Publications

Memory-Augmented Neural Networks

Memory-Augmented Networks

We develop memory-augmented recurrent architectures that generalize to sequence lengths far beyond training limits without losing computational efficiency.

  • Neural Turing Machines: Provably stable models computing with finite-precision memory over time.
  • Sequence Generalization: Recurrent models outperforming state-of-the-art Transformers on formal language tasks.
Memory-Augmented Neural Networks Structure

Key Publications

Explore Our Research

Latest News

April 2026 Internship

Hitesh Vaidya accepted a Time-Series Foundation Models Research Intern position at GE Vernova for summer 2026.

April 2026 Presentation

Shion and Rushitha presented their works on "Bridging Predictive Coding and Minimum Description Length Principle: A Two-Part Code Framework for Deep Learning" and "A Neuro-Symbolic, Guideline-Aware Pipeline for Uterine Cancer Restaging and Treatment Recommendation from Pathology Reports" at the 2026 USF AI+X Symposium.

January 2026 Acceptance

Raul Castillo and Abdul-Malik Zekri got their work "The Alpha-Divergence Connection Between Contrastive Representation Learning and the Free Energy Principle" accepted at the 2026 Florida Undergraduate Research Conference (FURC 2026)!

December 2025 Presentation

Benjami Prada presented "Realizable Circuit Complexity: Embedding Computation in Space-Time" at the NeurIPS 2025 Workshop (What Can't Transformers Do?).

December 2025 Presentation

Raul Castillo and Abdul-Malik Zekri presented "The 'Surprisal' Between Contrastive Representation Learning and the Free Energy Principle" at the 1st Bellini College REU Symposium (BCRS 2025).

October 2025 Publication

A paper on predictive coding survey with Karl Friston and Rajesh Rao got accepted in Neural Networks journal.

September 2025 Acceptance

One paper accepted in NeurIPS 2025 workshop ("Circuit Complexity From Physical Constraints") – Congratulations Benjamin!

September 2025 Acceptance

One paper accepted in EMNLP 2025 (main) – "Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval Augmented Generation Across Learning Style"

View All News →

Join Our Lab

We are actively seeking motivated graduate and undergraduate students interested in trustworthy machine learning, neuro-symbolic AI, and cognitive models.

If you want to build intelligent agents that are robust, interpretable, and provably stable, explore our open positions and research goals.

Apply / Contact Us