Skip to content

README.md

Latest commit

 

History

History
103 lines (88 loc) · 4.92 KB

README.md

File metadata and controls

103 lines (88 loc) · 4.92 KB
Raw

Notebook Curriculum

These notebooks are a guided first tour of the package. They are meant to make assembly calculus feel concrete: a stimulus leaves a sparse trace, traces can be compared, labels can be read out, sequences can be inspected, and small structured computations can be followed step by step.

The tone should be curious and hands-on, not grandiose. Treat each notebook as a small instrument on the bench. Run it, look at the numbers, change one parameter, and ask what moved.

Read the volumes in order if you are new to the package. Volumes 01-04 are the core path. Volumes 05-14 are thematic tracks for the larger research program.

For the longer teaching and visualization roadmap, see NOTEBOOK_PEDAGOGY_ROADMAP.md.

For the full notebook environment:

uv sync --group notebooks
uv run jupyter lab

The package itself does not require the richer notebook stack. Keep reusable trace helpers in neural_assemblies.assembly_calculus.tracing, reusable visualization helpers in neural_assemblies.viz, and optional interactive tools in notebooks.

Volumes

  • volume-01-foundations: Brain setup, projection, merge, lexicons, and readout, plus parameter and pattern-completion labs. You learn how assemblies are formed, inspected, and stress-tested.
  • volume-02-memory-and-computation: sequence memory, LRI diagnostics, FSMs, PFAs, and LRI parameter sweeps. You learn how assemblies can carry state over time.
  • volume-03-language: controlled NEMO-style parsing examples that run on the maintained assembly-calculus stack. You learn how category, role, and order can be composed in a toy language setting.
  • volume-04-research-workflow: how to inspect claims, experiments, and scientific status without turning demos into evidence. You learn how to keep software demonstrations separate from scientific conclusions and how to trace an observation toward a claim.
  • volume-05-dynamics-and-statistical-mechanics: order parameters, stabilization regimes, scaling, and phase-diagram thinking.
  • volume-06-binding-composition-and-interference: binding diagnostics, source-response checks, composition reliability, and interference.
  • volume-07-control-circuits-and-runtime: fiber gating, operation schedules, runtime protocols, and controlled information flow.
  • volume-08-prediction-and-erp-signals: toy next-token prediction plus the research boundary around N400/P600-style signal claims.
  • volume-09-embodiment-and-grounding: object/action/word grounding maps, missing encoders and readouts, and future embodied tasks.
  • volume-10-biological-validation-and-connectomes: biological parameter alignment, neural-data validation levels, and mapped-connectome hypotheses.
  • volume-11-scaling-engines-and-performance: backend selection, winner policies, sparse compute, and hardware-sensitive performance work.
  • volume-12-systems-paper-and-theory: paper-claim maps, theory synthesis, limitation gates, and figure/evidence planning.
  • volume-13-programming-assemblies: transition objects, operation graphs, fiber schedules, and the neural-programming-language frame.
  • volume-14-frontiers-and-open-problems: web-scale curricula, compiler ideas, connectomes, robotics, self-assembling substrates, and open-problem triage.

Reader Path

By the end, a reader should be able to:

  • create a brain with named areas and stimuli
  • form an assembly and compare it to another assembly
  • build a tiny label-to-assembly lexicon
  • inspect a memorized sequence and understand why recall is parameter-sensitive
  • run a small projection or LRI parameter sweep without hiding the regime
  • run a deterministic and probabilistic automaton helper
  • parse a toy sentence through the maintained NEMO-style parser
  • identify parser failure surfaces before making broader language claims
  • find the evidence behind a claim before repeating it
  • map a larger research idea to the correct evidence level before turning it into a notebook, experiment, or paper claim

Standards

  • Start each notebook with the question it answers.
  • Keep parameters visible near the top.
  • Use deterministic seeds unless the notebook is specifically about stochasticity.
  • Print overlaps, trajectories, or parse dictionaries instead of only saying that something "worked."
  • Prefer trace helpers such as project_trace and merge_trace when the reader needs to see dynamics over rounds rather than only final snapshots.
  • Use animations for neurodynamical processes when a static before/after plot would hide the interesting transition.
  • State whether the notebook demonstrates a package API, an experimental behavior, or a literature result.
  • Use code comments to explain intent, parameters, and what to inspect. Avoid comments that merely repeat Python syntax.
  • End with a small "try next" prompt so readers can safely explore.

Avoid notebooks that make broad claims from a single seed, hide parameters, or mix legacy artifacts with maintained package APIs.