Research project in Neurocomputing/Theoretical Neuroscience

Julian West

Project “Valrese”: Independent large research project in

Machine Learning/Neurocomputing/Theoretical Neuroscience

with a Python/TensorFlow platform

Drawing on my core expertise in CS/Math/Neurobiology, I have been carrying out a long-term independent research project in Machine Learning, with a major biologically-inspired slant.

Compared with the currently-fashionable style of ML, part of my approach is much more in line with how biological systems work – for example involving decentralization, asynchronous signals, chaotic systems, different timescales (akin to "metabotropic" changes in neurons), and evolutionary algorithms.
A mix of bottom-up and top-down approaches.
A search for the minimalistic essence of the computing capability of human neurons – minus all the extraneous elements of cell biology – also with an interest in AGI (Artificial General Intelligence.)

The current state of the project is a Mixed Platform that can handle Dual Network types:

"Traditional" ML/deep networks (using Tensorflow), plus a Genetic-Algorithm layer to help traverse the search space of hyper-parameters, initial conditions and network topologies
Pulse-driven networks in a Time Domain (similar to biological systems), using simulation algorithms I developed over the years.
Search space traversal takes place by means of Genetic Algorithms, as well as estimations of the gradients of a Gain function that may not be analytically available.

Of especial interest to me is ML to control/understand/fine-tune a dynamical system, such as the biological networks of my open-source project Life123.
Under such circumstances, the Gain function is not analytically available – and it will typically include random components from chaotic effects and/or interactions with unpredictable (simulated) environments. So, gradients can only be estimated from sample points in the search space.

Implementation: the dual platform is based on both my own independent research as well as on "traditional" ML/deep networks (gradient descent, Python, Tensorflow.) The code is at present not made public.

Technologies used: Originally prototyped in Mathematica, then developed in C#/.NET, with a WPF UI. Most recently ported to Python/TensorFlow/Flask, with a browser-based UI.

The old (C#/.NET) UI interface:

Using ML to control a simulated environment (code sample released to open source)
The integration of “traditional” ML (gradient descent, etc) with Genetic Algorithms
Systematic traversals of hyper-parameter spaces
Variable-dimensionality search-space traversal
Alternative Data Domains (for example, pulse-driven networks in a Time Domain; similar to biological systems)
Feedback from Higher-order areas to Lower-order areas (discussed in my blog entry)
Dynamically-changing network topologies
Using ML to fill in the knowledge gaps in the parameters needed for quantitative modeling of biological whole-cell systems (addressed, for a general audience, in my blog entry, and now starting to be implemented in my open-source project Life123 )

Custom visualizations were originally done with an SVG-based PHP library (very similar in spirit to D3.js) that I developed for this purpose, and are being ported to D3.js and to plotly.

Complex internal state