Hello everyone,
I am a PhD researcher working on the topic “Lightweight Machine Learning Methods for Rapid Analysis of Cosmological Data.”
My research objective is to develop computationally efficient machine learning methods for analysing large cosmological datasets, particularly when computational resources are limited. I am interested in datasets such as SDSS, IllustrisTNG and JWST observations.
Recently, I became interested in Jeff Hawkins’s Thousand Brains Theory and the Monty implementation developed by the Thousand Brains Project. The idea of learning structured internal models of objects, rather than merely identifying patterns in images, seems particularly interesting for scientific data analysis.
I would appreciate the community’s guidance on the following questions:
1. Potential applications to cosmology
Could Thousand Brains algorithms, particularly Monty’s sensorimotor learning and reference-frame-based representations, be adapted to analyse astronomical images and cosmological datasets? For example, could they help learn structural representations of galaxies, identify their components, or infer underlying properties from astronomical observations rather than relying exclusively on conventional image classification?
2. Research suitability and current limitations
Considering the current maturity of the Thousand Brains Project and Monty, how suitable and reliable are these methods for research applications outside their existing object-recognition benchmarks? What are their present limitations regarding scalability, computational requirements, continuous learning and generalisation?
3. Comparison with conventional machine learning
Could a Thousand Brains-inspired approach offer advantages over conventional CNNs, Transformers or other lightweight machine learning methods in terms of sample efficiency, computational cost, generalisation and interpretability? Are there any existing comparative studies or benchmarks that you would recommend?
4. Getting started
If I want to explore this research direction, what should I learn first? Which theoretical concepts, programming languages, libraries and computational tools are essential?
Could you also recommend the appropriate installation procedure, documentation, tutorials, example projects and GitHub repositories for getting started with Monty?
5. Recommended initial experiment
What would be a suitable small-scale experiment to evaluate whether Monty’s underlying principles could be useful for cosmological data analysis? Would starting with simulated 3D objects and multi-view learning be a reasonable approach before moving to actual astronomical observations?
My intention is not simply to apply an existing algorithm to astronomical images, but to investigate whether the underlying principles of Thousand Brains Theory could lead to more efficient and generalisable representations for scientific data analysis.
I would be grateful for suggestions, technical guidance, relevant publications and recommendations from researchers and contributors working on this project.
Thank you for your time and for making this interesting research and its implementation accessible to the wider research community.
Best regards,
SAJU S.G.
A PhD Research Scholar Research Area: Lightweight Machine Learning for Cosmological Data Analysis