symbol manipulation
The classical AI approach based on functions with variables bound to instances, applying rules universally across a domain regardless of prior examples—the essence of programming, algebra, and logic.
As defined in 184 | Gary Marcus on Artificial Intelligence and Common Sense
Where it comes up · 1
Most important claims · 3 of 3Search all claims for “symbol manipulation” →
- Symbolic AI is the essence of computer programming and algebra: functions with variables bound to particular instances, applying universally across an entire domain regardless of whether a specific input was seen before, whereas deep learning is entirely about similarity to previously seen examples.184 | Gary Marcus on Artificial Intelligence and Common Sense
- Roughly 99.9% of current AI investment goes to deep learning while classical symbol-manipulation AI is out of favor, yet a symbolic system recently won an upset victory over deep learning systems in a NetHack competition at NeurIPS, suggesting promising approaches can be prematurely counted out.184 | Gary Marcus on Artificial Intelligence and Common Sense
- As a sociological matter, the deep learning community has aligned itself against symbol manipulation for roughly 60 years, with figures like Geoff Hinton becoming hostile to symbols (telling the EU not to fund them) despite having earlier written about combining them, though Yann LeCun's recent 'you can have your symbols if I can have my gradients' suggests possible thawing.184 | Gary Marcus on Artificial Intelligence and Common Sense