This page keeps the project history and motivation out of the root README while still preserving them in the repository.
It is not the strongest evidence source for capability claims. For that, use
scientific_status.md, the tests, and the indexed
research artifacts under research/.
Alif Jakir maintains Assemblies. He began this implementation work after encountering the assembly calculus and the language-organ model in MIT's Projects in the Science of Intelligence course, then continued it through research extensions and collaboration with Daniel Mitropolsky (MIT Poggio Lab).
The aim is practical: make assembly-calculus ideas runnable enough to inspect, modify, test, and use in new experiments.
The package makes assembly-calculus mechanisms concrete in code. It exposes runtime objects, compute engines, assembly operations, sequence memory, language-oriented modules, and simulation helpers.
The repository around the package keeps the broader research record:
- experiments that are not package guarantees
- older prototypes and artifacts that still matter historically
- accelerator work that depends on local hardware
- question, result, and claim tracking for ongoing science
Keeping those layers together makes the work easier to inspect. Keeping them separate in the docs prevents the package from claiming more than it has earned.
One early question was whether assembly-calculus and NEMO-style mechanisms could move beyond the usual language and toy demonstrations into visual discrimination.
That work included CIFAR-oriented experiments. The lesson is deliberately qualified:
- the experiments showed feasibility in principle
- they did not establish robust per-category formation at package quality
That is why the docs describe the image-learning work as research history, not as a supported feature.
The repo did not begin as a clean package. It accumulated through research iterations, exploratory scripts, and repeated rewrites. The structure separates that history into:
neural_assemblies/for the installable packageresearch/for active scientific worklegacy/for archived root modules, scripts, artifacts, and MATLAB code
Scaling work, including CUDA-oriented engineering, is part of the project's trajectory. The present claim is still bounded: accelerator paths exist and are tested where the environment supports them, but speedups depend on hardware, problem size, engine choice, and CUDA/PyTorch/CuPy configuration.
The bet behind the repo is that assemblies may be a useful computational substrate in their own right:
- sparse intermediate structure instead of dense hidden states everywhere
- Hebbian and local updates instead of end-to-end backprop as the only path
- explicit operations such as projection, association, merge, and recall
- interpretable units that can be inspected, reused, and composed
That is a research program, not a solved result. The package gives us machinery for testing it.
The repo is mainly for readers who want to work close to the mechanism:
- computational neuroscience students and researchers
- neuro-inspired ML researchers
- people comparing sparse assembly models with deep-learning systems
- contributors who want runnable experiments rather than only paper diagrams
It is less useful if you want a general ML framework, a hosted service, or a drop-in replacement for transformer tooling.
This repository has been developed with AI assistance. That means AI helped with coding, organization, refactoring, and documentation drafts.
It does not mean authorship or scientific judgment is vague. The research direction, curation, and responsibility for claims remain human, led here by Alif Jakir in the context described above.
Use this page for history and intent.
Use these sources for claims:
- package behavior:
neural_assemblies/tests/ - scientific boundaries: scientific_status.md
- maintained code boundaries: supported_surfaces.md
- research evidence: ../research/claims/index.json
- curated questions: ../research/core_questions/index.json