It took decades before multi-layer neural nets demonstrated that Rosenblatt's view on layered perceptrons was right.

True: historical accounts show that the success of multi-layer neural networks in vindicating layered perceptrons came at least two decades after the early 1960s (e.g., backpropagation popularized in 1986, or the deep learning successes of the 2000s and 2010s), and that Rosenblatt had indeed advocated multi-layer or layered perceptrons. Mostly true: the multi-decade timing holds, but Rosenblatt's advocacy of layered networks is only partly documented. Mixed: the timing holds, but the evidence contradicts the claim that Rosenblatt held that view, or vice versa. Mostly false: multi-layer nets were demonstrated successfully within roughly a decade. False: they were demonstrated within a few years, or Rosenblatt clearly held no such view.

Set before the research began.

1. E1 Frank Rosenblatt vivarium.sh (wiki)

   “Rosenblatt understood that the interesting computations required more than one adaptive layer, and chapter 13 of the 1962 book proposed heuristic rules for passing error signals back to hidden units, which he called the back-propagating error-correction procedure, together with layer-wise learning rates, bypass connections, and multimodal variants.”

2. E2 Principles of Neurodynamics · AI Evolution Atlas AI Evolution Atlas · 1962

   “The book covers more than the single-layer case: Rosenblatt considers perceptrons with several layers and cross connections.”

3. E3 Frank Rosenblatt vivarium.sh (wiki)

   “The multi-layer machines Rosenblatt could not train with the tools of the 1960s returned two decades later with backpropagation, and the perceptron is now recognized as the direct ancestor of the artificial neurons composing modern networks, from multi-layer feed-forward networks to the Transformers underlying today's large language models.”

4. E4 Frank Rosenblatt vivarium.sh (wiki)

   “The heuristics did not work reliably: the step nonlinearity inherited from the threshold neuron had no usable gradient, and training networks with more than one adaptive layer remained open until Paul Werbos derived the reverse-mode gradient method in the early 1970s and David Rumelhart, Geoffrey Hinton, and Ronald Williams popularized it in 1986.”

E1 and E2 state that Rosenblatt's 1962 book discussed perceptrons with several layers and proposed rules for passing error signals back to hidden units. E3 and E4 state that his multi-layer heuristics did not work reliably. They also state that training multi-layer networks stayed unsolved until Werbos in the early 1970s and the popularization of backpropagation in 1986, about two decades later. This meets the criteria for both Rosenblatt's advocacy and the multi-decade timing, though the evidence comes from wiki-type secondary sources.