Frank Rosenblatt

• 1. Introduction

  ›

  Frank Rosenblatt died on the 11th of July, 1971 - his 43rd birthday - in a boating accident on Chesapeake Bay. It was the same date he was born, in 1928, in New Rochelle, New York. That strange symmetry frames a life that was itself full of unlikely convergences: a psychologist who built computers, a neuroscientist who studied rat memory, an astronomer who searched for signals from other worlds. And at the center of all of it, a machine that could learn. What did Rosenblatt actually build, and why did it take decades for the world to understand what he had done? Those are the questions this story tries to answer.

• 2. The Making of a Mind

  ›

  The Bronx High School of Science sent Rosenblatt to Cornell University in 1946, and Cornell never really let him go. He earned his A.B. in 1950 and his Ph.D. in 1956, and the work he did for that doctorate was already a sign of things to come. He built his own computer - the Electronic Profile Analyzing Computer, or EPAC - specifically to handle the number-crunching required for his research in psychometrics, the science of measuring human psychological traits. The data he was working with came from a paid survey of more than 200 Cornell undergraduates, covering 600 items per respondent. The total volume of calculations ran to 2.5 million arithmetic operations. Processing that would have taken 15 minutes by other means took EPAC just 2 seconds. Even then, he needed an IBM CPC alongside his custom machine to finish the job. What that PhD project revealed was not just a talented psychologist but a man who, when the available tools were not good enough, simply built better ones.

• 3. The Perceptron

  ›

  At Cornell Aeronautical Laboratory in Buffalo, where Rosenblatt rose from research psychologist to head of the cognitive systems section, he began work on an idea that would define his legacy. In 1957, he ran the first simulations of his perceptron on an IBM 704 computer. When a triangle was held before the perceptron's eye, the machine would pick up the image and route it through a random succession of lines to response units, where the image was registered. It was not a demonstration of brute-force calculation. It was a demonstration of learning. By 1960, that concept had become hardware: the Mark I Perceptron, a physical machine that could learn new skills by trial and error through a type of neural network designed to simulate human thought. The New York Times covered it under the headline "New Navy Device Learns By Doing". The New Yorker took notice as well. International recognition followed. The Mark I now sits in the Smithsonian Institution in Washington, D.C., recognized as a forerunner to artificial intelligence.

• 4. Theorems and Controversies

  ›

  Rosenblatt's theoretical work on perceptrons was mathematically rigorous. He proved four main theorems. The first established that elementary perceptrons - with no discrepancies in the training set and a sufficient number of independent A-elements - can solve any classification problem. The fourth proved that the learning algorithm converges when a perceptron can solve the problem in question. He also worked extensively on generalization: the ability of a trained model to recognize a pattern even under translation, rotation, or other transformations. His collaborator H. D. Block contributed centrally to this work. In 1969, however, Marvin Minsky and Seymour Papert published a book called Perceptrons that many read as an undoing of Rosenblatt's achievement. Minsky and Papert focused on perceptrons with restrictions - a bounded number of connections, or a limited receptive field - and showed that such constrained versions could not solve certain problems, including the connectivity of input images. The book was widely cited as proof that perceptrons as a whole were fatally limited. That reading was wrong. Rosenblatt had proved the power of unrestricted perceptrons; Minsky and Papert had demonstrated limits on restricted ones. The two findings do not contradict each other. But the misreading stuck, and research on neural networks cooled for years.

• 5. Principles of Neurodynamics

  ›

  Rosenblatt gathered his perceptron research into a book published by Spartan Books in 1962: Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. It had appeared earlier as an unclassified Defense Technical Information Center report, numbered 1196-G-8, dated the 15th of March, 1961. The book is organized in four parts. The first surveys historical approaches to brain modeling and lays out the foundational concepts of perceptrons. The second covers three-layer series-coupled perceptrons. The third addresses multi-layer and cross-coupled perceptrons. The fourth examines back-coupled perceptrons and poses open questions for future research. Rosenblatt drew explicitly on the work of McCulloch and Pitts for his neuron model, and acknowledged theoretical debts to Hebb, Hayek, Lashley, Ashby, Minsky, MacKay, and von Neumann. He used the book as the text for an interdisciplinary course at Cornell called Theory of Brain Mechanisms, taught to students from both the Engineering and Liberal Arts colleges. The cross-coupled perceptron machines he described in part three are what researchers today call Hopfield networks.

• 6. Tobermory and the Limits of Hardware

  ›

  Between 1961 and 1967, Rosenblatt's team built Tobermory, a scaled-up perceptron machine focused specifically on speech recognition. It filled an entire room. Tobermory was a four-layer neural network implemented with 12,000 weights, each stored in a toroidal magnetic core. George Nagy completed a PhD under Rosenblatt in 1962, primarily for work on the machine. By the time Tobermory was finished, though, the landscape had shifted. Digital computers had become fast enough that simulating neural networks on general-purpose hardware was now quicker than running dedicated perceptron machines. A purpose-built machine that took years and an entire room to construct had been overtaken by the computers it was designed to outpace.

• 7. Memory, Rats, and Other Questions

  ›

  From 1963 onward, Rosenblatt turned his attention to memory itself - specifically, to the question of how a human memory could persist over a lifetime. He designed neural network models of memory and analyzed them mathematically, though he never arrived at a convincing simulation experiment. Toward the end of the 1960s, inspired by James V. McConnell's experiments with memory transfer in planarians, Rosenblatt moved to a different approach entirely: biology. Working within the Cornell Department of Entomology, he ran experiments in which rats were trained on discrimination tasks - navigating a Y-maze, operating a two-lever Skinner box - and then had their brains extracted. Those extracts, along with their antibodies, were injected into untrained rats, who were then tested on the same tasks to see whether any behavior had transferred. Rosenblatt's conclusion, reached after several years of investigation, was that the early reports of large transfer effects were wrong. Any genuine transfer, if it existed at all, was very small. He also supervised PhD students who worked on the role of DNA in memory storage.

• 8. Beyond the Lab

  ›

  Rosenblatt built a private observatory on a hilltop behind his house in Brooktondale, about 6 miles east of Ithaca. From there he studied photometry and developed a technique for detecting low-level laser signals against intense non-coherent light backgrounds. He also joined the search for extraterrestrial intelligence, conducting that search from his own backyard. Outside science, he was active in liberal politics, working for Eugene McCarthy's 1968 presidential campaign in both New Hampshire and California, and joining protests against the Vietnam War in Washington. When he died in 1971, he was eulogized on the floor of the House of Representatives, with McCarthy among those who spoke. The Institute of Electrical and Electronics Engineers today presents an annual award in his name - the IEEE Frank Rosenblatt Award - honoring contributions to the field he helped create.