The Neural Network vs. Rule-Based AI Paradigms and the Subsequent 'AI Winters'

Key Insight

Frank Rosenblatt, a Cornell University professor, introduced 'the Perceptron' in a 1957 research paper, proposing a device capable of human-like perception and generalization. Inspired by the brain's complex structure of interconnected neurons, discovered in the late 19th century, Rosenblatt aimed to replicate this with artificial neurons. Despite the modest demonstration of a computer learning to distinguish card markings, Rosenblatt's persuasive presentations in 1958, aided by an IBM mainframe, generated immense hype. Headlines proclaimed a 'New Navy Device Learns by Doing,' expecting it to 'walk, talk, see, write, reproduce itself, and be conscious of its existence,' while others hailed it as 'the first serious rival to the human brain ever devised.'

Though Rosenblatt's claims were often characterized as 'a lot of hype,' he pioneered the concept of a 'brain-like computer' known as a neural net, which formed the basis of machine learning—where computers improve performance through learning and trial and error, not explicit programming. However, early hardware limitations were significant; his Mark 1 contraption could only manage 400 artificial neurons, a stark contrast to the human brain's approximately 86 billion. This neural network approach faced strong opposition from figures like Marvin Minsky of MIT, a proponent of rule-based computing (symbolic AI or expert systems), who dismissed neural nets as a 'dead end' and actively campaigned against them.

Minsky famously co-authored the book 'Perceptrons' to expose neural network shortcomings, declaring it 'an idea with no future.' This influential criticism led to a substantial shift in government funding away from machine learning and towards Minsky's rule-based approach, which envisioned hand-coding knowledge and reasoning into machines. However, rule-based systems, too, proved problematic, failing to scale to real-world complexity. The subsequent disillusionment, exacerbated by the overblown claims for technologies like the autonomous Shakey robot and the simple chatbot Eliza, triggered the first 'AI winter' from 1974 to 1980 as funding dried up. A second, longer 'AI winter' hit in 1987 after another wave of 'expert systems' failed to deliver, leading researchers to actively avoid using the term 'artificial intelligence' itself for grants and academic work through the end of the 1990s.