Context: Deep Learning’s Diminishing Returns
Rolf Landauer coined the phrase that information is physical. The consequences being that all computation and sophisticated or automated statistical analysis at scale has similar costs, at scale. Diminishing returns is a function of the second law of thermodynamics.
The article’s observation that “new approaches” are required is accurate. Are we actually trying to replicate human-level intellect or has that vanished into a cloud of self-propagating, self-validating commercial and academic or industrial career-existential necessity?
The most “successful” technologies in this context are measured by an ability to draw attention and to invoke the effervescent information artefacts of their own conceptual ancestry. These things are maximally-oriented towards self-replication, through us and – as with many if not most technologies – their primary focus is the direct or indirect management of the consequences and artefacts or costs of other technologies. The recursive hyper-inflation is quite real here.
In “A Thousand Brains” Jeff Hawkins references Vernon Mountcastle’s assertion that there was a universal algorithm of functional brain dynamics. As good a place as any to start, but I expect it will apply far beyond 12-watt brains.