2 Before the Perceptron — A Short Prehistory
Depends on: nothing — start here.
2.1 Why this matters
Knowing what came before the perceptron tells you what was being inherited and what was genuinely new. The “neural” idea was an old one by 1958: McCulloch & Pitts had a formal neuron in 1943; Hebb had a learning hypothesis in 1949; Minsky had analog hardware in 1951. Rosenblatt’s contribution was the synthesis — a single system combining a real-valued neuron, a learning rule, a convergence proof, and working hardware — not the invention of the neuron itself.
2.2 The mechanism
Two parallel tracks of work, each contributing a piece that the perceptron would later unify.
2.2.1 Manual / Theoretical (1943 – 1949)
These works are manual in the sense that no machine ran them — they are mathematical and conceptual proposals about what a neuron, or a network of neurons, could be.
McCulloch & Pitts (1943) (McCulloch & Pitts, 1943) proposed the first formal neuron: an all-or-none threshold unit. They proved that networks of such units can implement any propositional logic function. This is the abstraction every later artificial neuron inherits — the perceptron’s \(\mathrm{sign}(\mathbf{w}^\top\mathbf{x}+b)\) is a direct descendant.
Turing (1948) (Turing, 1948), in an unpublished National Physical Laboratory report, sketched “unorganized machines” — randomly connected NAND-like networks that could be “organized” by external reinforcement. The report sat in obscurity until the 1960s but anticipated connectionism by a decade.
Hebb (1949) (Hebb, 1949) proposed the first biologically grounded learning hypothesis: synapses strengthen when pre- and post-synaptic neurons fire together. “Hebbian learning” is the conceptual ancestor of every weight-update rule still in use today.
2.2.2 Automatic / Computational (1951 – 1956)
The next step was turning these theories into something a machine could actually run.
Minsky & Edmonds (1951) built SNARC (M. L. Minsky, 1954) at Harvard — a 40-neuron analog network wired from surplus B-24 bomber parts, simulating Hebbian learning in hardware. First physical artificial neural network.
Farley & Clark (1954) (Farley & Clark, 1954) ran the first digital simulation of a self-organizing neural network on the MIT Lincoln Lab MTC computer. Their network could classify simple patterns after training — software neural nets were now possible.
Rochester, Holland, Habit & Duda (1956) (Rochester et al., 1956) implemented Hebb’s cell-assembly theory on an IBM 704 at IBM Research — the first systematic computational test of a neural learning hypothesis. The gap between Hebb’s prediction and what the simulation produced was itself a scientific result.
2.3 Application & impact
The prehistory’s “application” is the perceptron itself. Rosenblatt’s (Rosenblatt, 1958) system unified the two tracks into one machine that learned.
2.3.1 Inheritance map
| What Rosenblatt inherited | From |
|---|---|
| Threshold neuron | McCulloch & Pitts |
| Weight-update learning idea | Hebb |
| Digital implementation | Farley & Clark, Rochester et al. |
| Custom physical hardware | descendant of SNARC’s spirit |
2.3.2 What was genuinely new in 1958
- A learning rule with a convergence guarantee — later formalized by Novikoff (Novikoff, 1962).
- Hardware that classified real images — the Mark I Perceptron, with a \(20\times20\) photocell input array.
- A public demonstration that triggered the field’s first hype cycle — and, when its limits were exposed by Minsky & Papert (M. Minsky & Papert, 1969), the first AI winter.
→ Next: The Artificial Neuron