A prerequisite to intelligence is the ability to find a program that explains a pattern. Programs are functions. To test a candidate program, we need to implement a "function evaluator". The problem is: all modern programming languages are sub-optimal "function evaluators", which, in the context of search, leads to massive slowdowns. To implement an optimal interpreter, we need to: 1. postpone the execution of expressions, to avoid wasted work, 2. cache the result of postponed expressions, to avoid duplicated work, 3. incrementally copy cached structures, to ensure 1 and 2 don't interfere. Haskell almost achieves this, but falls short on 3, because it is unable to incrementally copy lambdas. To solve this, we introduce the concept of a "superposition", which allows multiple versions of a term to exist simultaneously. This ensures that no work is ever wasted or duplicated, allowing us to optimally interpret (or com
François Chollet fchollet
VictorTaelin
/ optimal_evaluation_in_1_or_10_or_10_years.md
Last active
September 17, 2024 03:10
Optimal Evaluation in 1 Minute or 10 Minutes or 10 Years
##VGG19 model for Keras
This is the Keras model of the 19-layer network used by the VGG team in the ILSVRC-2014 competition.
It has been obtained by directly converting the Caffe model provived by the authors.
Details about the network architecture can be found in the following arXiv paper:
Very Deep Convolutional Networks for Large-Scale Image Recognition
K. Simonyan, A. Zisserman
##VGG16 model for Keras
This is the Keras model of the 16-layer network used by the VGG team in the ILSVRC-2014 competition.
It has been obtained by directly converting the Caffe model provived by the authors.
Details about the network architecture can be found in the following arXiv paper:
Very Deep Convolutional Networks for Large-Scale Image Recognition
K. Simonyan, A. Zisserman