- From: Dave Raggett <dsr@w3.org>
- Date: Wed, 11 Jul 2018 10:41:07 +0100
- To: Owen Ambur <Owen.Ambur@verizon.net>
- Cc: public-aikr@w3.org, Jason Lind <jasonl@transformation.run>
- Message-Id: <4F69E66D-61C4-447D-BBA8-BC3E4D7466EB@w3.org>
Thanks for the linked content. My reading is that they want to move beyond the acknowledged limitations of deep neural networks, e.g. to explore opportunities for incorporating hierarchical physical models, and to address learning from incomplete, sparse and noisy data, but I wouldn’t describe that as the “physics of AI”. Deep neural networks can offer impressive image recognition capabilities, but require vast amounts of training data and lack flexibility. Many species of mammals are able to make sense of what they see soon after birth, which suggests that the mammalian vision system relies on evolution to short cut the need for vast amounts of data. It would therefore be interesting to explore how evolutionary techniques could be applied to sparse learning with neural networks given an environment that is progressively enriched over the generations with a view to incorporating hierarchical models as part of the network architecture. > On 10 Jul 2018, at 19:27, Owen Ambur <Owen.Ambur@verizon.net> wrote: > > DARPA has issued a request for research proposals on the physics of AI. The goal and objectives of the request are available in StratML format at > http://stratml.us/drybridge/index.htm#DARPAPAI <http://stratml.us/drybridge/index.htm#DARPAPAI> or, more specifically, http://stratml.us/carmel/iso/DARPAPAIwStyle.xml <http://stratml.us/carmel/iso/DARPAPAIwStyle.xml> Dave Raggett <dsr@w3.org> http://www.w3.org/People/Raggett W3C Data Activity Lead & W3C champion for the Web of things
Received on Wednesday, 11 July 2018 09:41:24 UTC