A python framework for building recurrently connected Wilson-Cowan neuronal populations, synaptic weights/ footprints, and basic stimulus features, as well as running simulations and visualizing. Developed by Sara Steele

Getting started: Install requirements: < activate virtualenv, python 2.7 >

• open a terminal / shell,

• navigate to this directory (i. e., cd ~/WC_Sequence/ and

• run pip install -r requirements.txt )

• if you want to run the tests, I included a test-runner script. Open a terminal and make it executable on your system: cmhod +x run_tests.sh

• then you can execute the script

my new FR (firing rate?) sequence detector

SFA_WC_pop_func.py seems to work from the old days.

something more closely resembling a software package is being written in the src/ folder

New branch: a_units_same_stim

the whole sensitivity to \delta f comes from B stimulus being attenuated- this only happens because we assume they are getting less inputs from B. Which means we're looking at neurons with a best frequency at A.

We will get ITI sensitivity by applying a gate on the inputs based on ITI, ie \cite{Wehr2005} figure 1a

Got the habituator to work! we're going to scale all the a1 inputs by a habituation function (it kicks in as soon as stim offsets... maybe it should be activity dependent but we can model that later, this equation is basically just the recover from adaptation for SFA I think) It COULD also be the DECAY of an active sustained NMDA powered INHIBITION, that could explain a really fast onset and a slow decay for the ITI gating long lasting inhibitory current I think...

[us 20] interval detector- integration layer 1 -Stimuli - we need B-A vs A-B, and it breaks for incr df ABInterval and BAInterval

[us 5] with feature based synaptic footprints

• Synapses:

[us 19] micheyl 2005 actually looks like a fast self adaptation and a slow feedback inhibition from the nmda currents from the integrator layer

[tests]

[us 21] 2 interval detectors - aba vs bab

[us 16] make sure i cite cheng cheng for all the contributions to this type of wc architecture - her interval detector shares some similarities, i should nail down how close or not we are

[us 17] add frequency arguments for stimulus

[us 18] balanced connection footprints? excitation == inhibition OR not ; should probably be a parameter DOG?

[us 8] the gist of the paper is they do nmda only model, cuz all the other things are way faster. if i were to do something like that i could do S and A...

where did all the NMDA go...? i think the S is just sitting around, waiting to be a source for someone else's current... but I don't know what NMDA current eq is probably in xj wang somewhere NOPE he's all spiking https://books.google.com/books?id=YkV5AgAAQBAJ&pg=PA458&lpg=PA458&dq=wilson+cowan+nmda+synaptic+dynamics&source=bl&ots=fiY4P7w4h3&sig=neYbFFeHFuz8mdvqZoaMkaeLgDo&hl=en&sa=X&sqi=2&ved=2ahUKEwiuicKC1oLeAhVQPK0KHfvRCcoQ6AEwBHoECAQQAQ#v=onepage&q=wilson%20cowan%20nmda%20synaptic%20dynamics&f=false it's a current! eq. 15.90 FOUND IT (sequence detector notes on gdocs) Wong and wang 2006 recurrent network mechanism of time integration in perceptual decisions Ref wang 2002 mean field

• note: i don't have to go straight to realistic nmda; i should start with simple ee, no? and read the papers...

[us 9] can i describe "attractors" in this system...

[us 10] eye candy... it would be fun to take xj's image of stimulus selective sustained activity (from working memory papers) and create something similar for auditory guys. add noise?

[us 11] perturbations... discrete events are a bad idea with uncertain tau (from ch 3, cumhist) but periodic perterbuations might be a swell deal, ie, if we could do 3/2 period or something like that

[us 15] add noise... ornstein uhlenbeck, i should have eq/ code in my other pub (ARP). There's something about how you normalize the variance with dt; i think it may be sqrt dt or something? dt <<< 1 < fokker plank eq...? describes... how distr of locations of a random walk vary over time... or something like that >

[mysteries of the deep]

• run build_and_then and and_then - there's some weird sh*t!

[us 25] split neurons into kwargs arrays... neural jsons... wow

[us 26] for big networks i probably want to vectorize my code and all my vars for all my units.

[us 14] -- note- this rambling turned into CurrentTrace and tonotopic network 2 replaced tonotopic network around 2018-11-08 ish I can probably make it faster if I get rid of the mutables and map them to object attributes; as long as the methods know which attribute to pull from, ie sfa current has a source of u1, attribute r and a target of u1, attribute a (or rather the sfa current update method knows to read source.r and update target.a) and source and target are the same; traces updates to read from... i guess each current needs a "read_source" method so trace can call the same thing and pull from different attributes. original speed copying values to list: it says it's only .01 to .08 sec, but it seems to be > 2. weird. will have to see how it scales. Response = namedtuple('Response', ['value']) r = Response(0)

[us 24] 14 is probably too much to rework right now but since i just started networks, maybe i can get traces to live with units in the net simulation? that should make for an easier data structure to deal OR-- a "current-trace" in the sensory sim, makes the most sense.

[us 23] return dx for each current

[us 4] and we can write arbitrary feature-selective networks sims and visualize

• it looks like we need 2 different formats to make line plots and area plots. [us 22] seaborn figures- mermaid

would it help to give Traces a Type, so we know how to plot them? This can also be done in the data frame function... easier there, closer to the product

For everything interesting we want with feature selectivity, each unit has to have its own tuning curve.

[us 13] how do i deliver same stim to different tuning efficiently? stim generator function, static method. maybe a class method?

[us 7] WC superclass

[us 6] stimuli superclass

[us 12] rpt stim maker

[us 2] For everything interesting we want with feature selectivity, each unit has to have its own tuning curve.

[us 1] I could make an input A guy take all the inputs and NMDA-inhibit itself now we have habituated inputs and a baby new network

[us 3] Which means stim should have an absolute reference, even if it's an index

• melopy.utilities.frequency_to_key!

The Terminal in my Pycharm isn't able to do anything besides commit to git; it certainly doesn't run tests from the same conda environment as alternating_renewal_process. Running tests by right clicking on the tests folder in the project window seems to work

been having trouble getting packages installed in my virtualenv... workaround has been to install with pip from python console... pip.main(['install','-flags', 'package'])

it comes down to this terminal being shit. just use git bash, source activate alternating_renewal_process and pip whatever you want and it'll show up in python console yay

branch 'modularizing': Author: Sara Steele sara.steele14@gmail.com Date: Thu Mar 19 12:54:00 2015 -0400

archived some code, going to start doing the test code for build-and-then (my disinhibitory object-oriented network)

commit fba42cd32d62edc86fa63ca1cdcc6f78b13011cc (HEAD -> with_inhibition, origin/with_inhibition) Author: Sara Steele sara.steele14@gmail.com Date: Fri Feb 6 16:34:15 2015 -0500

did that work?

commit 955e6ec5d4a3c35bbe950a3a6a629751937af837 (HEAD -> nullcline_analysis, origin/nullcline_analysis) Author: Sara Steele sara.steele14@gmail.com Date: Mon Feb 9 20:42:14 2015 -0500

starting the process of modularizing