In 2017, I wrote the book 'The Deep Learning Playbook' deeplearningplaybook.com that described 3 pillars of research that will drive innovation: Meta-Learning, Modularity and Game Theory.
In 2017, we had a nascent understanding of these 3 and how they relate to each other. Today, we now have a much better understanding of their relationship and this will lead to accelerated innovation. #deeplearning #ai.
The overarching model is near completion and we only need to fill in the blanks.
There aren't many roadmaps out there by the more prominent DL researchers. Yoshua Bengio's may the only that has expressed one:
Jeff Clune (@jeffclune) describes one that resonates with me: . His 3 pillars are: (1) meta-learning architectures (2) meta-learning the algorithms and (3) learning environments.
@jeffclune AGI is sufficiently complex that it has to be grown and not engineered. The first step is to understand the first principles of why it needs to be grown. I don't think this has been elucidated clearly enough.
@jeffclune The combination of first principles and automated discovery is what gets you to AGI even if you don't understand how to design AGI. The problem with many is that they have incorrect first principles.
@jeffclune But where can one find the first principles? It can be found in the complexity sciences. @sfiscience
@jeffclune @sfiscience In my research in the past several years, I pursued several blind alleys. The danger however of saying that something is a blind alley is that it may not be! Many have said Neural Networks was such an alley, only to been disproved wrong decades later.
@jeffclune @sfiscience Actually, an entire generation (2 decades) later. If Deep Learning is indeed wrong, then it'll likely be another generation that'll discover AGI. I see many detractors these days. But they contribute no promising alternatives. If so, what then is the use of their pessimism?
@jeffclune @sfiscience AGI is damn hard and it has to be understood why it is damn hard (see: medium.com/intuitionmachi… ). But you are living today and not a generation in the future.
@jeffclune @sfiscience The gamble that everyone should be making is the gamble that AGI can be achieved in 5 years. That's because, despite its complexity, a short cut might be possible: medium.com/intuitionmachi…
@jeffclune @sfiscience The trillion-dollar question is... what is that short-cut to AGI? Have we already discovered that short-cut but labeled it as a blind alley? What have we overlooked?
@jeffclune @sfiscience The alternative approach is to believe that the short-cut is still a generation away to be discovered. But if this is indeed true, then is this information that is actionable for you today? It's not! That's why most Deep Learning detractors are totally and absolutely clueless.
@jeffclune @sfiscience Remind me if you are one of the clueless.
