Today’s AI research process resembles intelligent design more than it resembles evolution. Novel architectures are manually discovered by researchers. Is it possible to scale up genetic algorithms that perform neural architecture search using a dedicated blockchain network, designed with appropriate incentive systems?
Intelligent Design vs. Evolution
I’m no brain expert, but my understanding is that animals (including humans) possess “world models” that help them survive and reproduce. Animal brains are composed of many small neural modules (visual, language, planning, etc.), the architecture of which have evolved to make learning “world models” easy.
One goal of AI research today is to build Artificial General Intelligence (AGI), a model or system that has general-purpose intelligence similar to that of humans. And much progress has been made — modern neural networks can handle multi-input (even multi-modal, e.g. text, image, audio, etc.) and multi-output (solving many tasks with a single network).
But the research process largely consists of manually defining a network architecture, manually curating training data sets, and then training the network from scratch. This process is like designing a single animal, letting it live an entire life, then manually designing another animal based on the results. It’s not a scalable approach. In fact, it is literally an intelligent design approach rather than an evolutionary one.
The scale of nature’s search for the best brain architectures is massive. Entire populations of animals within a species compete with entire populations of other species for survival, over eons. Each individual within a species has small variation, and individuals from different species have larger variations. Each new generation of offspring preferentially have the best brain architectures of the previous generation. This process runs over millions of individuals, thousands of species, and thousands of generations.
The computational costs to replicate anything remotely at this scale for neural architecture search for AGI are unattainable for individual researchers.
Blockchains can massively scale neural architecture search.
Blockchain networks are already using the kind of compute resources needed for such a task. Current blockchain networks run useless calculations, but what if a new blockchain network was designed to perform useful AI tasks?
Consider a blockchain network where each node contains an instance of a neural architecture. Each block might correspond to a generation. Winning nodes might get to reproduce with each other, and the node operator should receive some reward. The probability of winning could be proportional to performance on some task(s).
A user of the network could submit a task along with a reward for its completion. The tasks would require 3 components: training examples, test examples, and unlabeled examples. Getting predictions for the unlabeled examples by the best models in the world (as measured on the test set) would make it worthwhile for users to offer a reward for completion.
Neural architectures that can solve many different tasks with minimal retraining time and minimal compute resources would proliferate. Such models would earn rewards from a larger percentage of tasks submitted, and would perform well with lower electricity and hardware costs.
- How can neural architectures be standardized and hashed (into a genetic-like code) in a way that enables offspring to not only get a mix of each parent but also permit mutations of sufficient complexity to eventually lead to completely novel architectures (hopefully that converge on “world model” submodules)?
- This paper on neural architecture search with genetic algorithms might be a good start for further research.
- Can data be encrypted end to end somehow, even with a public blockchain? This would drive mass adoption as the network could be used for tasks requiring proprietary or sensitive data.
- The blockchain project Fetch.ai might be a good start for further research.