Anveshana अन्वेषण · the search that goes into the unknown

A 4-neuron network that climbs LKG → PhD over real NCERT/CBSE textbook content, creating its own neurons whenever it fails an exam.

What it is learning

The corpus is chunks of actual textbook text from Class 1 to Class 12 across subjects. Every rung asks a harder question about that text: is this English or Maths? → which of five subjects? → which class of school is this from? → which of fourteen senior-school subjects?

The network is one shared trunk: 128 hashed bag-of-words inputs plus a 16-way one-hot "which rung am I sitting", one tanh hidden layer, 16 outputs. It starts at 4 neurons. Pass a rung and it moves up; fail and it splits its most-used neuron into two. At the end it sits a board exam on all 16 rungs, revises, and grows again while any rung is still under its bar.

Curriculum — 16 rungs of the syllabus

Rung	The question	Classes	Exam	Bar	Neurons	Grew	Sittings	Result

Does it beat its own earlier self?

Every saved generation, scored on the PhD exam (14 senior-school subjects) — one fixed held-out test, re-scored after each rung.

Neurons created over the run

Board exam & revision

–

board mean before revision

–

after revision

–

rungs under the bar at the end

Controlled experiment — growth vs. no growth

Same corpus, same rung order, same Adam. The control keeps its 4 neurons and is given more gradient steps than the grown model.

Seed	Grown: PhD exam	Grown: board	Grown: neurons	Grown: steps	Frozen: PhD exam	Frozen: board	Frozen: steps	Δ

Progress map — every generation × every rung

Rows are saved generations, columns are rungs. Cells are accuracy on that rung's full held-out exam. Click a cell to load it in the lab below.

darker = lower. The last row is the model after revision.

Test lab — put it on a real textbook page

green = this generation got the item right, red = wrong. Click to jump.

This generation's score on the whole rung

–

Try it yourself

Paste any paragraph. The page featurises it exactly as training did — hashed bag of words, same IDF, same weights — and shows what the network thinks it is. This model classifies text; it cannot chat, because there is no language model in it. Nothing here generates a sentence.

Persistence & the schedule

Progress lives in cbse_checkpoint.json: architecture, every weight, the curriculum log, the revision log and the board scores. A scheduled run resumes from it, trains another block, re-writes this page and saves the new checkpoint — so the model keeps improving across days without ever rebuilding from scratch.

Gradient steps in this run: