I find this to be empty & impractical. If this methodology were beneficial, blue ocean businesses would be propping up like mushrooms.
I challenge anyone who's read this to propose a single blue ocean idea here
I challenge anyone who's read this to propose a single blue ocean idea here
A digital rubber band is not a "blue ocean idea".
http://articles.chicagotribune.com/1991-09-02/features/91030...
It's connecting multiple different technologies to attack a wide open market space. Think: smoking, drinking, gambling, oversleeping, getting over your ex, overspending, unhealthy eating, etc.
If that isn't blue ocean strategy then I don't know what is.
(Edit because of downvotes: I was serious, this is a blue ocean idea. For example, imagine children with their own ontology of language and experience (not culturally acquired). That's about as blue ocean as you can get. Neural networks are pretty close. Neural news at nine.)
As for neural networks, generative models, genetic algorithms, etc., they can help us optimize on known search spaces but that's about all they can do. It's a big leap from designing simulation-optimized bicycle parts to generating truly new ideas.
Did you read the recent Software 2.0 rant by the Tesla AI guy? https://medium.com/@karpathy/software-2-0-a64152b37c35
Other good sources are The Master Algorithm (ML) or The Brain That Changes Itself (neuroscience), the former surveying general capabilities and limitations of various approaches and the later surveying findings from medicine on just how plastic our sensory inputs and interpretive capacities are through recent research.
Second, he's describing a pretty radical implementation of what amounts to a new form of fuzzy constraint programming using exactly the technique I described: brute-force optimization on a very sophisticated objective function. It's fundamentally statistical, and it's still fundamentally an optimization problem. Neural networks happen to be modular compared to other optimization techniques, and they happen to do a miraculous job at capturing latent higher-order structure in these optimization problems.
The real innovation he's describing is that we now have powerful enough optimizers that it's cheaper, easier, faster, and sometimes more reliable to just fuzzy-optimize problems instead of coding bespoke solutions for them. That's pretty amazing, but suggests absolutely nothing.
Now, a self-modifying neural network might be interesting, that can monitor its own performance and decide when to retrain part of itself, mutate its own architecture, or request new training data. But we definitely are not there yet.
With all that said, there's a much more down-to-earth version of your idea that neural networks can have their own ontology and experiences: that's just an anthropomorphization the latent structure that they capture. This isn't unique to NNs -- it's something researchers have been exploiting for decades. It's only a small stretch to argue that that's what fitting a decision tree or a principal components model is. This is reflected directly in the jargon that these models "learn" a representation of the data. You can extract new representations of reality right now by running HDBSCAN on a data set of your choice. Neural networks just let us scale that up to richer, higher-dimensional problems.
Maybe that's all "experience" really is, consuming data from the world and encoding various representations of that data. But when it comes to generating truly novel ideas, instead of incremental improvements on existing ideas, I'm bearish on how far our current neural network technology can go. Maybe AlphaGo, which apparently invented new Go strategies that no human player had ever thought of, is a counterexample. Or it's the exception that proves the rule. I guess we'll see as computation power continues to improve.
Precisely my point.