The last invention.

  • 247 Replies
  • 20762 Views
*

ranch vermin

  • Not much time left.
  • Starship Trooper
  • *******
  • 484
  • Its nearly time!
Re: The last invention.
« Reply #240 on: November 10, 2017, 12:14:00 pm »
is that just a filter, or does it take learning to better the result?

*

korrelan

  • Trusty Member
  • *********
  • Terminator
  • *
  • 784
  • Look into my eyes! WOAH!
    • Google +
Re: The last invention.
« Reply #241 on: November 14, 2017, 11:43:54 pm »
I just read your question and thought… what’s he on about? The video is on the previous page and I’d not noticed.  I was rushing getting ready to leave for the first call of my working day, that’s not the video I thought I had posted lol. That was for Yot’s thread on outlines.

I’ll explain it here anyway… It uses two kinds of machine learning. 

The first is an adaptive convolutional filter I designed that learns the best parameters to apply to each section/ type of image. It automatically adjusts for brightness, saturation, clarity/ sharpness, resolution, etc.  Its job is to learn/ adapt the best method for extracting a template that the pixel bot (PB) can follow.

The second machine learning technique is the pixel bot.  This little guy is a simple bot with eight pixel sensors around its perimeter.  I have taught/ trained it to follow and trace outlines. 

So… the pixel bot learns to follow outlines and the adaptive convolution filter learns to extract the most information from the image.  If the PB manages to create a full out line then it notifies the filter that this is a good convolution combination for this type of image… and a box is drawn around the shape on the original image.

Notice that he outline is stabilized in the left window.  As soon as the mouse pointer enters a shape the PB converts the object into a set of vector coordinates, these are then easily centred and stabilized.

That’s how the system manages to easily extract numbers from captures etc… it learns to extract letters/ numbers from the confusion.

You can see the PB learning process here…



I was taking a break from my AGI and thinking about Yot’s outline routines.

 :)
It thunk... therefore it is!

*

ranch vermin

  • Not much time left.
  • Starship Trooper
  • *******
  • 484
  • Its nearly time!
Re: The last invention.
« Reply #242 on: November 15, 2017, 07:53:42 am »
I wouldve thought it would have involved alot of noise before its learnt,  how come its only off or an exact outline.

If I had a random convolution filter, it would be all kinds of blurs and differences.

*

ivan.moony

  • Trusty Member
  • *********
  • Terminator
  • *
  • 800
  • Can computers cry?
    • Structured Type System
Re: The last invention.
« Reply #243 on: November 15, 2017, 07:59:02 am »
Korrelan, very interesttng, I think it would be how humans detect edges (they move eye focus around an object). It seems that neural nets can do more than I thought.
If you vaporize a teardrop, you get a salt.   :flake:

*

korrelan

  • Trusty Member
  • *********
  • Terminator
  • *
  • 784
  • Look into my eyes! WOAH!
    • Google +
Re: The last invention.
« Reply #244 on: November 15, 2017, 10:28:35 am »
Quote
I wouldve thought it would have involved alot of noise before its learnt,  how come its only off or an exact outline.

I have manually trained the pixel bot (PB) to follow outlines, because I’ve used my judgement to best trace lines the PB has learnt to do the same, so this includes jagged outlines. It will run along a fairly jagged edge and produce a straight line quite well, it could do with more training though as sometimes it gets confused on certain patterns of pixels lol.

Quote
If I had a random convolution filter, it would be all kinds of blurs and differences.

The convolution filter is indeed based on the human eye and is localised around the mouse pointer.  It affects small regions individually not the whole image.  It does not use a random filter; again I have manually trained the filter to best extract the required detail.

Quote
Korrelan, very interesttng, I think it would be how humans detect edges (they move eye focus around an object)

This stems from my AGI research, I use a model of the mammalian visual cortex which learns to detect lines and gradients automatically; this is just a very simplified version used just for outlines.

Quote
It seems that neural nets can do more than I thought.

I tried to keep the project simple so I’ve not used any kind of neural net, it’s all good old fashioned look up lists.  So the PB for example just stores a list of perimeter sensor readings along with the direction I’ve told it to move.  The PB can then easily and quickly find the closest match in the list and move in the relevant direction.

The system is basically just reproducing my skills at following an outline.

I still disagree that using outlines is a good method for recognising objects, both scale and rotational invariance can be accounted for but occlusion cannot.  Feature detection is still by far the best method.

 :)
It thunk... therefore it is!

*

korrelan

  • Trusty Member
  • *********
  • Terminator
  • *
  • 784
  • Look into my eyes! WOAH!
    • Google +
Re: The last invention.
« Reply #245 on: November 15, 2017, 12:14:18 pm »
This early low resolution test example shows the use of a three layer neural net based on the mammalian visual cortex.  The four angled coloured lines in the centre of the windows shows the basic colours and orientations the system is going to learn.  The left window shows the current learned orientation map, the right window is the image its learning from.



The first half of the vid shows the system being subjected to just diagonal lines and the orientation map automatically learns to recognise these lines.  You can see the self organising distribution in the neural layer only using the two diagonal colours.

At 0:17 you can see the output from the LGN running through the map and only the diagonal lines are detected.

At 0:30 a picture of faces is loaded and again because the system was trained on just diagonal lines it only detects diagonal lines in the face image.

At 0:50 I start re-training the system on the faces.  You can see the neural plasticity of my system at work as the orientation map slowly evolves to incorporate the horizontal and vertical lines it’s experiencing in the face image.

At 1:23 I stop the training, the self organising nature of the system has built a map that best fits/ represents the input data.  The black dots on the left window represent the locations of output layer pyramid cells; each cell is surrounded by a receptive field tuned to recognise a particular feature in the image.

The rest of the vid just shows how the system is now interpreting the face image with all four orientations learned.  Obviously my current system uses hundreds of orientations and gradient combinations to detect features in the visual LGN output. 

If the system was only subjected to diagonal lines again it would very slowly forget how to recognise the vert/ horiz lines.  This plasticity effect falls off as the neurons mature so the system eventually reaches a balanced representation of the experienced input, with a patch of neurons able to detect every facet of the incoming data.

The advantage of this approach is that each patch of neurons in the left window will only fire when a particular pattern of input is supplied.  The neuron patches or image facets never move and so can be easily linked, searched, etc to figure out what the system is looking at.

This gives an idea of how my whole AGI system functions; it is very closely modelled on the mammalian cortex and is basically able to adapt and learn anything its sensors encounter.  It automatically extracts and organises relevant information from the data streams.

It takes a human foetus nine months to generate its V1 orientation/ gradient maps this is my system producing and equivalent map in a matter of seconds from visual experiences… it learns extremely quickly.



And if you were to zoom into the neuron sheets/ maps you would see the individual neurons and synapse.



 :)
« Last Edit: November 18, 2017, 11:17:10 am by korrelan »
It thunk... therefore it is!

*

keghn

  • Trusty Member
  • ********
  • Replicant
  • *
  • 689
Re: The last invention.
« Reply #246 on: November 15, 2017, 02:54:38 pm »
 Unsupervised truth.
 I have had a thought of using GANs to do transmissions and self learning at the same time.
 There is problem in neuroscience of what is the meaning of "Same" or "equal" is, at the level of a neuron.
 GAN are make up of two NNs. The detector and the re generator NN, that regenerate what is being  detected.
 IF you had a  alternating rows of detector NN {DeNN) and re generator NN (ReNN). The information could be past along
in a daisy chain fashion.  Like so:
DeNN, ReNN,  DeNN, ReNN, DeNN, ReNN, DeNN, ReNN, DeNN, ReNN.............................
 
 These NNs here only detect and regenerate the color of one pixel. So it will not be too slow.

 The first DeNN detect the color form the real world. Then the next ReNN generates it. And the next DeNN learns to detect it.
 And this keep going on and on so the brain has many true reference of this color. So when doing edge detection or blob
detection, the colors form two pixel can be compared to see if they are the same or different.


 

*

ranch vermin

  • Not much time left.
  • Starship Trooper
  • *******
  • 484
  • Its nearly time!
Re: The last invention.
« Reply #247 on: November 15, 2017, 03:24:52 pm »
woah now i understand,  thanks for that.

looks like a cool alternative to normal edge detection cause ur getting the angle as well,  but i guess the filter could too,  but its cool that youve adapted machine memory to do it.    they are general purpose indeed.

To keghn,   yes - if you had one neural network generating images, and the other video trained network telling it yes or no,  it will start to generate the diagonal lines - but i think that gets more interesting if the concepts are more general and abstract, then the generation might be less restricted and generate some more interesting images.

 


Users Online

31 Guests, 3 Users
Users active in past 15 minutes:
Freddy, ivan.moony
[Administrator]
[Trusty Member]

Most Online Today: 54. Most Online Ever: 208 (August 27, 2008, 09:36:30 am)

Articles