Mark I Perceptron

What the machine was

The Mark I Perceptron was the first implementation of the perceptron algorithm, a custom-built analog computer designed for image recognition and developed under the project name "Project PARA"^[1]. Earlier perceptron work had been a simulation running inside a general-purpose computer; this machine put the same idea into physical parts — photocells for input, a plugboard for connections, potentiometers to hold adjustable weights, and electric motors to turn them during learning^[1].

Where it came from

In 1957 Frank Rosenblatt was at the Cornell Aeronautical Laboratory, where he simulated the perceptron on an IBM 704^[1]. His interest lay more in hardware than in simulation, and he obtained funding from the Information Systems Branch of the US Office of Naval Research and the Rome Air Development Center to build a custom-made analog computer, the Mark I Perceptron^[1]. The Smithsonian records that Rosenblatt of Cornell University described a neural network he dubbed a perceptron in 1957, and that by the following year he and colleagues had constructed the device; the museum gives 1958 as the date made for the object it holds^[2].

Rosenblatt's team assembled and tested the machine at the Cornell Aeronautical Laboratory in Buffalo, New York, between June 1959 and 14 December 1959^[1]. Its first public demonstration took place on 23 June 1960^[1]. The work ran under two contracts: Nonr-401(40), "Cognitive Systems Research Program", lasting from 1959 to 1970, and Nonr-2381(00), "Project PARA", lasting from 1957 to 1963^[1]. Wikipedia notes that the Office of Naval Research, rather than ARPA, funded the Perceptron project because it was considered unlikely to produce technological results in the near or medium term^[1].

How it worked

The machine was connected to a camera with 20×20 cadmium sulfide photocells, producing a 400-pixel image^[1]. In one version the network had an array of 400 photocells in a 20×20 grid as sensory units, a hidden layer of 512 association units, and an output layer of eight response units; Rosenblatt called this three-layered network the alpha-perceptron^[1].

The three layers were wired differently, and that difference is the heart of the design. S-units were connected to A-units randomly through a plugboard — a deliberate choice meant to eliminate intentional bias in the connections — and those connection weights were fixed rather than learned^[1]. A-units connected to R-units with adjustable weights encoded in potentiometers, and during learning the weights were updated by electric motors^[1]. So only the second stage could adapt: the random first stage stayed exactly as built, while the potentiometers turned as training went on. The Smithsonian describes the network as composed of three types of signal-generating units — sensory (S), association (A) and response (R)^[2].

The underlying idea is simple to state. A unit takes a set of inputs, multiplies each by a weight, adds the results, and responds when the total passes some level; learning means changing those weights so the responses come out right. The Mark I made that arithmetic into hardware a person could open up and inspect: 400 photocells feeding 512 association units, each connection either a wire in a plugboard or a setting on a potentiometer^[1].

What became of it

The machine was shipped from Cornell to the Smithsonian in 1967 under a government transfer administered by the Office of Naval Research^[1]. The Smithsonian says the Mark I Perceptron came to the museum in 1967 from Cornell University as a transfer from the Office of Naval Research^[2], and holds it as object number 1987.0819.01, with 1958 given as the date made and Buffalo, New York as the place made^[2].

Where it sits in the history

As the first implementation of the perceptron algorithm, the Mark I turned a learning rule into a working instrument built from a camera, a plugboard, potentiometers and motors^[1]. Its three-part organisation — sensory, association, response — is also what makes it easy to describe: a fixed random input stage, an adjustable middle stage, and a small output stage that produced the answer^[2]. The machine is now a museum object rather than a working device, but it is the physical reference point for what a trainable network looked like when it was built out of analogue parts instead of simulated in software^[1].