Incremental emergent intelligence

[Ken Kafieh]

Links?

[Ken Kafieh]

Links?

[Ted Ferguson]

Learning takes time. If you had a large, energy efficient neural network, why not give it plenty of time? 

Hardware and learning implementations should be versatile and capable. Balancing the effectiveness with the complexity of the network. You'd need to be able to design the AI and build neural network. There are many ways to solve problems using NN's both simulated or implemented. A recipe for it would be nice, but limited.

Performance Trends in AI

Training Spiking Deep Networks for Neuromorphic Hardware

How Will Artificial Intelligence Affect The Next 20, 30 And 50 Years Of Our Lives

[keghn feem]

  For a image down sample to a dot. keep or stop at point when the NN stops working at detecting accurately:  

https://en.wikipedia.org/wiki/Pixel_art_scaling_algorithms

 Then keep a ROI, region of interest. The focus. 

 The brain focus onto a small place. It will us a smaller fraction of its hardware, software or NNs: 

https://www.youtube.com/watch?v=owBz6R-Ttqc

 So everything here is RNN / LSTM pinpoint scanner or a sequential temporal NN point scanner detector.  

 Its scan movement by eye muscles.  The decision of scanning, Hilbert's Curve?:

https://www.youtube.com/watch?v=DuiryHHTrjU  

 What do know, NN do a really good job at chain codes: 

http://answers.opencv.org/question/88501/how-to-classify-chain-codes/  

http://www.cis.hut.fi/research/IA/paper/publications/bmvc97/node2.html  

http://pappubahry.com/misc/rectangles/  

http://pappubahry.com/misc/rectangles/  

 Or the NN brain can chase lines.

 This is unsupervised learning so the brain assumes that something exist and goes out and looks for it

with a simple NN.  if it fine something with this simple tiny detector it build on it and with other tiny NN. 

 Then AI rebuild the world with its brain from this focus point scanning. The brain has its internal doodly bored

and recreates drawings, complete images, and they are perfected when they match with what the focus point detector.

  Now with data record it can now work with it:  

https://groups.google.com/forum/#!topic/comp.ai.neural-nets/hNCR7PkfgmE

 

[keghn feem]

       Nice.

[keghn feem]

 Brute force autoencoder?

 I have seen layered autoencoder where the first input level is broken into squares and a little autoencoders learn sub features

such as solid patches of color edges of all type and so on then next level decides which one to let through and 

where on the final output. If the first level of tiny auto encoder does not have it then the brute force is used.

 I have also seen a scanning brute force autoencoder that does down sample the image put does a little part of 

the image at time. A little scanning auto encoder. It scans the whole image and generate an output into a output image

it had a lot of gpu muscle behind it, thow.  

[keghn feem]

A DARPA Perspective on Artificial Intelligence:   

https://www.youtube.com/watch?v=-O01G3tSYpU

[keghn feem]

Fast PixelCNN++: speedy image generation: 

https://github.com/PrajitR/fast-pixel-cnn

[keghn feem]

Google Brain’s new super fast and highly accurate AI: the Mixture of Experts Layer:  

 https://medium.com/@thoszymkowiak/google-brains-new-super-fast-and-highly-accurate-ai-the-mixture-of-experts-layer-dd3972c25663#.8bmlqaep6