GENREG
Live training mapPhi activation map Sign in Register for updates

Radial Space research · gradient-free learning

I am building a model that learns without backpropagation.

GENREG is Payton's live research lab for a different kind of machine learning system: reusable activation programs, called codons, arranged into radial spaces and composed into higher-level decision layers.

The breakthrough is practical now: the system can produce useful classifiers without gradients. The latest confirmed record is 61.0% test accuracy: V7v, a pure depth-by-voting stack where each space votes over the previous space and the same fixed tournament reads the result, with zero fitted readout weights and no composer on top. The current frontier is finding the scaling law that keeps that gain while driving the parameter cost down and the accuracy up toward the 99% regime.

Registration is recommended if you want research updates, build notes, and first access when the scaling path is cracked.

01

How it differs from gradient models

Gradient models tune millions of weights by backpropagating loss. GENREG searches over small deterministic programs and composes the useful ones into spaces. The learned object is closer to a reusable alphabet than a dense weight matrix.

02

What started working

Pairwise contrast experts created the first real jump. The current best line uses a wide bank of evolved pair experts, then lets RS1 learn how to recombine their votes. That is the first confirmed readout above 0.60 on CIFAR-10 in this system.

03

What is still unsolved

The system is not at modern deep-learning accuracy yet, and the current high score is wider than the earlier clean record. The honest target is the scaling law: which codon families, interfaces, and composition rules keep improving without exploding the parameter count.

Current research direction

RS0 builds the primitive observations. RS1 and RS2 are being pushed into recombination layers: not just another classifier, but spaces that ask which contrast is actually in dispute, preserve the useful codons, and compose the right internal language for that decision.

What I am working toward

  • Reliable gradient-free learning that can be inspected as maps, slices, codons and pairwise experts.
  • A scaling recipe that keeps improving with more codons, more spaces and better interfaces.
  • Open first access for people who want to follow the research closely and test the system when the scaling law is found.

Clean baseline record

0.5339RS0 + supervised RS1 rank fusion · 24,854 genes · the reference line every new result is measured against.

Parameter-aware gain

0.565–0.567Relational pairwise E2 fused with the record · near the original record scale · useful evidence that the mechanism is not only brute width.

Latest confirmed record

0.6500RG03e: one depth-by-voting layer (45 pair experts of grid codons, margin + residual fitness) over the union of three Z31 pair-expert banks. Input union 0.6425, two-seed version 0.6423, single bank 0.6218. No composer, no fitted readout. Previous records 0.6456, 0.6332, 0.6118.

Phi composite activation map

This is the public layered activation overlay: 31 post-activation substrate banks embedded in one shared frame so people can see the class tree, the rotated phi families, and which activation families stay aligned or move away.

Open full map

Why I abandoned gradient thinking

I was told evolutionary AI was too slow, too inefficient, and useful only for toy problems because the search space becomes unmanageable on harder tasks. I never left evolutionary AI. What changed was that I stopped trying to solve an evolutionary problem with gradient-model assumptions.

The space was never the enemy. The progress started when I abandoned gradient thinking and treated the system on its own terms: explicit activation programs, searchable structure, composable spaces, pairwise contrast experts, and maps that show what the model can and cannot express. GENREG now works without backpropagation; the open problem is the scaling law that turns the working mechanism into a high-accuracy system.

Register for updates and first accessSign inTermsgenreg.tech · research in progress, reported honestly