While researching neuroevolutionary algorithms, neural networks performed well on a variety of tasks, creating complex systems. But when I needed something more complex, trying to create a GenAI for text or a simple translator between two languages, I ran into serious problems. After almost a year of trying, it seems to me that neither NEAT nor its variations (HyperNEAT, CoDeepNEAT, etc.) can handle this, but I'm not sure. So is it possible to create generative models or at least translators using neuroevolutionary algorithms?
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
No, this is not possible. Or more precisely, it is not practical.
Evolutionary algorithms are a very general optimisation process that perform semi-random and local searches in a parameter space (the "genome") that must be converted into some expression (the "phenotype"), evaluated and selected.
This process, and any evolutionary algorithm based on it, is conceptually simple and very robust, but it is also very slow when faced with complex searches. When evaluation is slow, or when solutions require large numbers of tuned parameters, using an evolution-inspired algorithm becomes impossibly slow.
LLMs, which typically have billions of paramaters, and where training consists of processing billions of input/output pairs, are large scale in both aspects (number of parameters and amount of evaluation), and are many orders of magnitude out of reach of gradient-free learning algorithms like NEAT.
Evolution built human brains in the real world, but the underlying mechanics and scale of the system that allowed this is very different from anything we can perform on a computer. We don't fully understand it, but we can say that something like the NEAT model is a very different kind of system. For example, mammalian genomes don't have very many tunable parameters (maybe 100,000? There are an estimated 20,000 protein-coding genes for example in the human genome*), but the phenotype expression and evaluation are beyond anything related to current computation limits - it's not really computationally feasible to simulate a single cell on a computer to the level where a true-to-life evolution could occur. And that's before we consider parallel processing involving massive numbers of indvividuals over millions of years.
When we can design a system based on function approximation, are able to describe example inputs and outputs, and can take advantage of gradient-based algorithms, this makes LLM training feasible. There is no known practical gradient-free method to training models at the scale of LLMs.
* Equating a gene expressed in DNA to a "tunable parameter" is probably wrong, but it's not important. We cannot replicate biological systems in that much detail on a computer anyway.
I would also add that successfully processing the genetics and physical development of a simulated human would lead to a "blank slate" system that in the real world takes years to learn useful language skills. We don't understand which layers of the system can be shortcut and don't need to be simulated down to the atomic level ...
