1. Conundrum1885

    Neural networks


    Thought I'd ask for some assistance on this.

    My current project is to do what some have said is impossible, ie run a small AI on a RaspPi model B or B+ (512 or 1GB RAM)

    The plan here is twofold, use a 64GB microSD (about £30 locally from Airtel) and the Pi's on-board memory with some clever electronics to keep it in the critical zone where the quantum effects begin to affect adjacent cells leading to a cemi field analog and an exponential increase in complexity.

    To do this I plan to add temperature feedback and a very large magnet on either side that causes intentional leakage in the cells as well as a weak alternating field to imitate the cemi field.

    The 64GB microSD at the moment is intended to be used for neural pattern storage as it can be reformatted to FAT32 and hopefully a second 64GB used with an external adapter to act as a visual cortex by rapidly cycling through stored patterns and using the same effect to recognize images slightly differing from the stored image.

    Has anyone done anything remotely like this, perhaps some pointers would be useful?



    1. Conundrum1885

      Re: Neural networks


      Now have one of the Class 10 64GB (Integral) and a RPi B+ so can do some experimentation.

      It does look like the speed on the B+ can be changed by simple crystal replacement at the cost of breaking the HDMI output, however the workaround for this is to run it in composite mode and use an even multiple of the crystal frequency to get double the effective frame rate (ie 100Hz)

      Once booted the mode switch circuitry would slowly ramp up speed and feed back memory/CPU temperature on the GPIO lines to keep the system on the edge of stability without crashing hard, using external Peltiers so that both chips can be kept at the best operating temperature.

      For CPU I found that 7C works best from prior experience and memory maybe 55C if overclocked enough to show inter-cell leakage.

      A second Pi (probably model A) would then read back the system state and when a pattern match is detected the key found is stored and the original file then written back to external memory.

      This would also work for a lot of other applications such as pattern recognition at very high speeds.

  2. Conundrum1885

    RE. Re. Neural networks

    Update 05/09/15.

    This is intriguing.

    I have found evidence that even a lowly Pi Model A (256MB RAM) can be coaxed into basic AI behavior just by replacing its clock crystal with a VCO connected to a data pin.

    The hack requires that to initially boot it needs to be actively cooled but once this is done the clock speed can be slowly ramped until the AI code takes over and keeps itself in the active semi classical regime which occurs at about 61C at this clock speed.

    I didn't see this before because my setup wasn't precise enough but it looks like the cooling can be done using Ca$io projector peltiers with the copper plates and then additionally copper plated with some basic chemicals to form a mesa that is then attached to the Pi's chip with a 3D printed jig.

    Its fiddly but worth it, this adds to the usefulness of these already handy modules.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon