Short answers, pulled from the story.
Frank Rosenblatt was an American psychologist born on the 11th of July, 1928, who invented the Perceptron, widely recognized as a forerunner to artificial intelligence. He is sometimes called the father of deep learning for his pioneering work on artificial neural networks, culminating in the Mark I Perceptron in 1960, the first computer that could learn new skills by trial and error.
The Mark I Perceptron was a hardware machine completed in 1960 at Cornell Aeronautical Laboratory in Buffalo, New York. It could learn, recognize letters, and solve complex problems using a neural network that simulated human thought processes. The machine is currently housed in the Smithsonian Institution in Washington, D.C.
Marvin Minsky and Seymour Papert published Perceptrons in 1969, which proved that perceptrons with restricted inputs had limited capabilities. The book was widely and incorrectly cited as proof that perceptrons in general were fatally flawed, dampening neural network research for years. Rosenblatt had separately proved that unrestricted elementary perceptrons could solve any classification problem, a result that does not contradict Minsky and Papert's findings.
Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms was published by Spartan Books in 1962. It summarized Rosenblatt's perceptron research in four parts, covering historical approaches to brain modeling, three-layer perceptrons, multi-layer and cross-coupled perceptrons, and back-coupled perceptrons. The cross-coupled perceptron machines Rosenblatt described are now known as Hopfield networks.
Frank Rosenblatt died on the 11th of July, 1971, his 43rd birthday, in a boating accident in Chesapeake Bay. He was eulogized on the floor of the House of Representatives, with former Senator Eugene McCarthy among those who spoke.
Tobermory was a scaled-up perceptron machine built between 1961 and 1967, designed for speech recognition. It occupied an entire room and contained a four-layer neural network with 12,000 weights implemented in toroidal magnetic cores. By the time it was completed, however, simulating neural networks on standard digital computers had become faster than running purpose-built perceptron hardware.