i had a crazy idea and wanted to see it work. in "Ecology, The Ascendent Perspective" Robert Ulanowicz explains how ecosystems change over time; and i think, how all networks self-optimise. the ascendent perspective has always been second nature to me: the foundation for my spiritual, intellectual, and creative thinking. i took the principles from an ecology textbook and used them to train a neural network
when chemicals ignite they produce heat, which ignites more of the substrate, producing more heat - an autocatalytic cycle. dust in a protoplanetary disk travels in cycles too, some cycles more effcient than others (better at conserving angular momentum). good cycles are autocatalytic relative to others; weak cycles are pruned. ascendency means "rising together". as ascendency in a system increases its cycles become: homogenous, specialised, effcient, complex, productive, brittle, insular, and slow-changing
ascendency explains the success of shipping containers, the poor resilience of modern crops, the economic shift toward niche startups, rising suicide rates, and everything else too. i started building a network simulator for studying socioeconomic issues, and then i asked myself "can this tool learn?"
shifting some % of each flow's weight backward increases a network's ascendency. after a few iterations square networks become perfect cycles, and larger grids develop cycles with considerable complexity
values travel through the network boosting weight transfer as they move, and
transfers are amplified when the output is correct. cycles which yield correct
answers are autocatalytic and become stronger. this network learns to minimise
mse(output, [100, 100])
an exchange (one input, many outputs) splits a value by weighted average. therefore weights are effectively constrained to values between 0 & 1, and the sum of each exchange's weight is always 1
it took me two weeks to integrate negative values, bias, activation, normalisation, and slope. this network learns to flip a signal
unfortunately... flipping signals is almost the limit of what my networks can achieve. they can approximate AND/OR but fail to learn NAND/XOR. the constraints are too severe and i'm sure my approach is fundamentally flawed (but don't know how)
| AND | in1 | in2 | target | out | OR | in1 | in2 | target | out |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 1 | 1 | 1.20 | 1 | 1 | 1 | 1.49 | ||
| 1 | -1 | -1 | -1.10 | 1 | -1 | 1 | 1.01 | ||
| -1 | 1 | -1 | -1.02 | -1 | 1 | 1 | 0.95 | ||
| -1 | -1 | -1 | -1.49 | -1 | -1 | -1 | -1.23 |
i'm unsure whether this paradigm for training neural networks could match back propogation. but if someone can make it work there will be some exciting possibilities for sure! by flipping flows these networks could restructure themselves: create memory capacitors, fine-tune recurrence, and syncronise values through time. i can't help but wonder whether task-independent autocatalysis could enable AI systems to consider original goals, possess an innate curiosity, and exhibit volition
there is no user interface. you must edit the source code to change the program (look at ./pages/index.js)
look at ./programs/README.md for a guide to the simulator's programming language
git clone git@github.com:ashtonsix/ascendency.git
cd ascendency
npm install
npm run dev