袁洋_RLChina-on-the-power-of-foundation-model_watermark.pdf

1、Intelligence is Positioning:Intelligence is Positioning:On the Power of Foundation ModelsYang YuanWhat is intelligence?Intelligence is compression What compression?Zip files intelligence Intelligence is embedding/representation Supervised:labels Self-supervised:what?Intelligence is positioning A for

2、est,not a treeSimple example:similarity graph on imagesWhat is intelligence here?A:I can predict similarity of each pair of images!B:I can position each image!Simple example:similarity graph on images Two approaches:1.Supervised learning:(input,output)(1,2)=similarity of(1,2)2.SimCLR:function of two

3、 objects(1),(2)=similarity of(1,2)What is the difference?is given,a simple metric function!Positioning:compute(),automatically get similarity between and all other trees!Definition of similarity graph Where are the labels?Self-supervised learning Labels are from itself?Wrong!Labels are from prior kn

4、owledge!What operations makes an image similar to it?SimCLR For any two images,Two points in pixel space Two points in positioning space similarity directly computed by metric is our neural network A function that maps images to positions!is our“ideal space”Why?Any two points have similarities Why?F

5、ormal analysis Given objects 1,Including all augmented images,finite Augmentation pair(,)defines a similarity edge in ,=Prob(,sampled together),are similar semantically However,and are not similar in pixel space(large distance)Question:Can we find an ideal space,such that semantic similarity is capt

6、ured naturally?=()Various solutions!Today:Reproducing Kernel Hilbert Space.Reproducing Kernel Hilbert Space Given,consider:,such that ,=,Inner product in RKHS,is the kernel function in can have infinite dimension,we do not need to compute explicitly Similarity between,defined in Well defined,well sh