[ml] [Noisebridge-discuss] Technical question: Directionally sensitive photosensors

Christoph Maier cm.hardware.software.elsewhere at gmail.com
Tue Jun 11 17:01:02 UTC 2013


On Tue, Jun 11, 2013 at 9:34 AM, Mike Schachter
<mschachter at eigenminds.com> wrote:
> Also cc'ing this discussion to the neuro and ML lists, there are might be
> people there that could help with the problem.
>
>  mike
>
> On Tue, Jun 11, 2013 at 9:25 AM, Mike Schachter <mschachter at eigenminds.com>
> wrote:
>>
>> I suppose the basic idea is to create a chain of integrate-and-fire
>> neurons.


The general idea is this:
http://www.youtube.com/watch?v=a-Vy7NZTGos

and to make it interesting in a neural network sense,
I'd want _some_ directional sensitivity in the photoreceptor
(and I need to find a solution within the next week ... LDR works with
power-hogging bipolar technology, but it's a bit bulky).

Ideally, the field of view of each device would be adjustable ... from
fisheye to narrow angle ...

>> Referring to the attached diagram, each neuron gets input from it's own
>> photosensor, and from the neuron to the left of it. You will also have to
>> make it so that the signal from the photosensor to neuron *decreases* as you
>> move right on the chain.
>>
>> The general idea is that movement to the right makes each neuron spike,
>> which sends a spike to the neighboring neuron. If the inputs are timed right
>> (which is probably the most tricky part), each successive neuron will spike
>> more and send more input to it's neighbors. The last neuron, the readout
>> neuron, will spike the most. You can look at the spike rate of the last
>> neuron, and if it's high enough, then you declare movement to the right. The
>> number of spikes of the last neuron could even be related to the velocity of
>> the movement.
>>
>> Movement to the left shouldn't elicit a single spike from the readout
>> neuron.
>>
>> Another option might be to build a Reichardt correlator:
>>
>> http://people.csail.mit.edu/rondror/Ronfiles/Research/reichardt.htm
>>
>> mike

What's a Reichardt correlator?
http://www.ini.uzh.ch/~kramer/pub/pami96.pdf

>> On Mon, Jun 10, 2013 at 11:46 PM, Christoph Maier
>> <cm.hardware.software.elsewhere at gmail.com> wrote:
>>>
>>> On Mon, Jun 10, 2013 at 5:49 PM, Mike Schachter
>>> <mschachter at eigenminds.com> wrote:
>>> > Oh ok! Well I'd be happy to help Chris with code to do
>>> > directional-selection
>>> > given data from an array of photo-sensitive inputs if that would be
>>> > useful.
>>> >
>>> >  mike
>>> >
>>> >
>>> >
>>> > On Mon, Jun 10, 2013 at 5:20 PM, Norman <pryankster at gmail.com> wrote:
>>> >>
>>> >> Mike, that is exactly how a phased array works.
>>> >>
>>> >> Norman
>>> >>
>>> >>
>>> >> On 6/10/2013 5:07 PM, Mike Schachter wrote:
>>> >>
>>> >> Can you just get an array of non-directional photosensors and do the
>>> >> direction-selectivity computationally on an arduino or something?
>>> >>
>>> >>  mike
>>>
>>> Great!
>>> Just tell me how to put code into this kind of computer ...
>>> http://pony.noisebridge.net/~cmaier/neuromorphic/BIL2011/html/img10.html
>>
>>
>


More information about the ml mailing list