3 ms·
U.S. Military Sees Future in Neuromorphic Computing
- deepnotderp 9y agoI've said it before and I'll say it again, neuromorphic computing is a fun side research topic, but at this point, nothing more. NOTE: All of this completely changes if neuromorphic chips would put up benchmark results, as of now, the neuromorphic flagship, "TrueNorth" takes a heavy accuracy hit on CIFAR compared to contemporary DL techniques (~5-8% IIRC), imagine how bad it will be on real world datasets! The essence of my opposition to neuromorphic computing can be approximately distilled to the following exchange: Neuromorphic architecture: "Here's this really efficient 1-bit/analog computer we made!" Programmer/AI Researcher: "Cool, does it do anything useful? Have you tried it on benchmarks?" Neuromorphic Architect: "Well, it doesn't work on benchmarks, but that's your problem, you just don't understand the genius!" AI researcher: -_- So you made a very efficient chip that doesn't do any useful computation? So an inefficient heater? Later sometime, neuromoprhic architect to media: "These people just don't understand the genius of neuromorphic, it's the future of chip design and will auto-magic away our problems!" Media: "OMG Yes, stoopid programmers inhibiting adoption"