No learning technique can succeed without a minimal amount of prior knowledge
Yann LeCun et al., November 1998 (Proceedings of the IEEE). Yann LeCun and colleagues at AT&T Labs, in their November 1998 paper on convolutional networks for reading handwritten digits and bank checks.
Gradient-Based Learning Applied to Document Recognition
More from Yann LeCun et al. · AI and science · Before ChatGPT
Who else?
What it refers to: Prior knowledge about the task, which in multilayer neural networks is built in by tailoring the architecture, as convolutional networks do for 2-D shapes.
Other voices
- “Given no domain knowledge except the game rules” Silver et al.,
- “Solely measuring skill at any given task falls short of measuring intelligence” François Chollet,
- “Simply by waiting for faster GPUs and bigger datasets” Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton,
- “These models somehow just generalize dramatically worse than people” Ilya Sutskever, Builder,
- “Learning meaning requires more than form” Bender et al.,
