Research Overview#
I am a researcher at the Thousand Brains Project , where I work on reverse-engineering neocortical computation to develop more biologically-plausible artificial intelligence systems. My research focuses on implementing cortical column architectures and sensorimotor learning algorithms in Monty , our brain-inspired AI framework named after Vernon Mountcastle.
My work is grounded in the belief that intelligence emerges from sensorimotor interaction with the world, and that current deep learning approaches, based on outdated neuroscience, miss fundamental aspects of biological computation. I am particularly interested in how cortical columns—the proposed universal computational units of the neocortex—process information across different modalities and how they communicate to form coherent representations of the world.
Academic Journey#
My path to brain-inspired AI reflects an evolution from reductionist to systems-level thinking:
Microfluidics and Tool Development (UC Berkeley, BS 2015)
Initially driven by a desire to understand the human body through mathematics and biology, I recognized that advancing our tools was essential for studying cells—the fundamental units of life. My work in droplet microfluidics contributed to the technological foundation that enabled Drop-seq, revolutionizing single-cell genomics by massively parallelizing what was previously limited to analyzing handful of cells at a time.
Single-Cell Genomics (Stanford University, Research Assistant)
Working in Professor Stephen Quake’s laboratory, I transitioned from wet lab to computational approaches, gaining hands-on experience with cutting-edge single-cell RNA sequencing technologies that would inform my later work on cellular diversity and function.
Computational Neuroscience and Deep Learning (University of Michigan, PhD in EECS)
My doctoral work culminated in MorphNet
, a generative deep learning framework for predicting three-dimensional neuron morphology from single-cell gene expression data. This work addressed the labor-intensive nature of neuron tracing while exploring the relationship between molecular identity and structural form. Through this research, I developed a deeper appreciation for network-level phenomena and recognized that understanding intelligence requires moving beyond single-cell analysis to systems-level thinking.
Research Philosophy#
My approach to intelligence research is guided by several core principles:
Thoughts as Movement: One of my most profound insights from working with the Thousand Brains Theory is that thoughts can be conceptualized as a type of movement. This perspective, combined with my current investigation of grid cells in the entorhinal cortex, suggests that abstract reasoning may be grounded in the same spatial and motor mechanisms we use for physical navigation.
Unifying Scales: I envision bridging micro-, meso-, and macro-scale neuroscience—from single neuron mechanisms through neural circuits to whole brain regions. While current work in Monty operates at a high level of abstraction, I aim to develop theories that seamlessly connect insights from molecular biology to fMRI-scale phenomena.
Engineering Rigor in Science: Having programmed in Python for 14 years, I bring software engineering excellence to research environments. I am passionate about reproducibility in science and implement industry-standard practices including test-driven development, continuous integration, and comprehensive version control—approaches that remain uncommon in academic research but are essential for advancing the field.
Technical Contributions#
MorphNet : A generative model using GANs and diffusion models to predict neuron morphology from gene expression, validated against morphological properties including Sholl profiles.
tbp.floppy : A Python package for counting floating-point operations in NumPy computations, demonstrating deep familiarity with Python internals and commitment to computational efficiency.
Monty Framework: Currently implementing compositional object recognition and working toward unifying learning and inference algorithms—a critical step toward continuous adaptation that distinguishes biological intelligence from current AI systems.
Vision for Intelligence Research#
I believe we are approaching AGI from the wrong direction. Rather than generalizing through massive data ingestion, biological systems demonstrate remarkable few-shot learning through active exploration. Monty already recognizes objects through brief sensorimotor interaction without pretraining—a fundamentally different paradigm from current deep learning.
My ultimate goal is to create systems that learn and adapt continuously, much like biological organisms. Success in unifying learning and inference would enable truly adaptive AI systems and robots capable of coexisting naturally with humans in dynamic environments.
Technical Expertise#
- Programming: Deep Python expertise (14 years), with particular strength in scientific computing and software architecture
- Machine Learning: PyTorch, generative models (GANs, diffusion models), biologically-inspired algorithms
- Neuroscience Tools: Single-cell RNA-seq analysis, morphological analysis, connectomics
- Software Engineering: Test-driven development, CI/CD, MLOps, reproducible research practices
- Frameworks: NumPy (internals-level knowledge), scientific Python ecosystem, computer vision
Looking Forward#
The next frontier in intelligence research lies not in scaling existing approaches but in understanding the principles that allow biological systems to learn efficiently from limited sensorimotor experience. By combining insights from cortical computation, spatial cognition, and software engineering excellence, I aim to contribute to a new generation of AI systems that truly understand and interact with the world as biological organisms do.
“Understanding intelligence requires moving beyond reductionism to embrace the beautiful complexity of systems that learn through interaction with their world.”