Research

The path to Artificial General Intelligence (AGI)

My current research interests focus on the path to Artificial General Intelligence (AGI). One project is looking specifically at the design and development of an "agentic" analytical reasoning systems which includes a cognitive architecture for automated simulation-driven AI self-improvement. My thesis is that the state of the art in commercial AI products can most quickly via an agent-executed workflow layer shaped by multi-level guidance and feedback loops. These loops automatically detect both product and evaluation infrastructure issues and fix them. Higher-order feedback loops support this process with methodology generalization and upleveling from solved cases to generalized capabilities. Additional support comes from agentic self-learning improvement via reflection, meta-analysis loops, and ongoing cognitive process modeling refactoring, synthesis, and refinement loops.

Theoretical Neuroscience

For a long time, before the advent of deep learning, the field of Artificial Intelligence seemed stuck in the paradigm of symbolic AI for language, and machine learning models for solving classification and value estimation (logistic regression, random forests, gradient boosting, ...), although there was starting to be some impressive progress on machine translation using stastical models (BERT, LSTM).

Given this predicament, it seemed that the only way to figure out how to create intelligence in a machine was going to be to reverse engineer the brain: figure out how populations of neurons encode signals, how circuits, and expecially the cortical microciruit, transform signals to construct representations, and how the signal flows connecting brain areas work together to route, synthesize, and abstract information into perception, an understanding of the world, planning. and action. This line of inquiry led inevitably to neuroscience.

After being inspired by what Jeff Hawkins was doing with the Redwood Neuroscience Institute, started learning theoretical neuroscience in 2008 and eventually connected with the institutes reinvented instantiation as the Redwood Center for Theoretical Neuroscience at UC Berkeley, where I joined for a few years as a Visiting Scholar, working on dynamic ciruits of representation formation in the early visual system (area V1). This produced some publications (see Google Scholar).

An enjoyment of public writing and demystifying how the brain works from what is currently known to science led to writing neurosciene answers on the Q&A website Quora (60k followers) where I became the "top writer" in neuroscience and was selected as a "Top Writer" for 2018, 2017, 2016, 2014, and 2013. See representative answers.

Over 60 Quora answers have been picked up and published by sites such as Forbes (20+), Huffington Post (10+), Slate (8+), and Business Insider (7+). See selected articles.

My core neuroscience research interest is in understanding how the brain works when viewed as an information processing system, with a particular focus on the neural circuits underlying visual perception. I am also interested in human organizations when viewed as systems, and the use of technology to implement intelligent learning behavior.

My published work has focused on biologically-realistic spiking network models of visual pattern detection and sparse code formation. Sparse coding is a method for representing information that appears to be used by the brain. Its key characteristic is that very few representation variables (neurons) are active at any given moment.

One research project is E-I Net, a neural circuit model and simulation engine written in MATLAB that learns sparse code patterns using an approach inspired by the brain's visual cortex. This work was published in The Journal of Neuroscience (abstract, PDF).

For more on my neuroscience interests and point of view, see my webpage on theoretical neuroscience research.