Costi, I think your reasoning is right as of the importance of the multi-modal input for generalisation. As of ANN (as popular), they were explicily mentioned only in several slides in the lecture about Narrow AI and why it failed, with some hopes given to Schmidhuber's LSTM. Slides 34-36: http://research.twenkid.com/agi/2010/Narrow_AI_Review_Why_Failed_MTR.pdf
In a new course they would deserve more attention. :) Perhaps current ANN could be used for AGI if glued appropriately with other methods (not just one ANN config. on tensorflow e.g.) or maybe as Goodfellow suggested in the podcast on AI podcast of MIT - if a big enough multi-modal dataset is fed with appropriate representation (your proposal is in that direction as well). However with current NN I assume it would be massively less efficient than with a better more clever, or "selective", method. <http://research.twenkid.com/agi/2010/Narrow_AI_Review_Why_Failed_MTR.pdf> If you lacked access to the computing power of the big organizations, working their way is impossible, i.e. adding more and more resources, while using the same or just slightly adjusted simple algorithm. You have to either invent something smarter which works on the less powerful hardware available to you, or join them. "Deep Learning" as hierarchical processing at different resolutions (time including, all dimensions), starting from sensory input, interactive adjustment (but more so), RL ideas (but hierarchical), multi-modal data etc. are right directions regarding the "generality", but the way the mainstream ANN do it yet is not what I considered "AGI" either back then and now. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T5417cc95d981211e-M1bbb803753b84a8f5de3cbd8 Delivery options: https://agi.topicbox.com/groups/agi/subscription
