Synaptic scaling balances learning in a spiking model of neocortex (Rowan & Neymotin 2013)

 Download zip file   Auto-launch 
Help downloading and running models
Learning in the brain requires complementary mechanisms: potentiation and activity-dependent homeostatic scaling. We introduce synaptic scaling to a biologically-realistic spiking model of neocortex which can learn changes in oscillatory rhythms using STDP, and show that scaling is necessary to balance both positive and negative changes in input from potentiation and atrophy. We discuss some of the issues that arise when considering synaptic scaling in such a model, and show that scaling regulates activity whilst allowing learning to remain unaltered.
1 . Rowan MS,Neymotin SA (2013) Synaptic Scaling Balances Learning in a Spiking Model of Neocortex Adaptive and Natural Computing Algorithms, Tomassini M, Antonioni A, Daolio F, Buesser P, ed. pp.20
Model Information (Click on a link to find other models with that property)
Model Type: Realistic Network;
Brain Region(s)/Organism: Neocortex;
Cell Type(s): Neocortex V1 L6 pyramidal corticothalamic GLU cell; Neocortex V1 L2/6 pyramidal intratelencephalic GLU cell; Neocortex V1 interneuron basket PV GABA cell; Neocortex fast spiking (FS) interneuron; Neocortex spiny stellate cell; Neocortex spiking regular (RS) neuron; Neocortex spiking low threshold (LTS) neuron; Abstract integrate-and-fire adaptive exponential (AdEx) neuron;
Gap Junctions:
Receptor(s): GabaA; AMPA; NMDA;
Transmitter(s): Gaba; Glutamate;
Simulation Environment: NEURON; Python;
Model Concept(s): Synaptic Plasticity; Long-term Synaptic Plasticity; Learning; STDP; Homeostasis;
Implementer(s): Lytton, William [bill.lytton at]; Neymotin, Sam [samn at]; Rowan, Mark [m.s.rowan at];
Search NeuronDB for information about:  Neocortex V1 L6 pyramidal corticothalamic GLU cell; Neocortex V1 L2/6 pyramidal intratelencephalic GLU cell; Neocortex V1 interneuron basket PV GABA cell; GabaA; AMPA; NMDA; Gaba; Glutamate;
autotune.hoc *
basestdp.hoc *
batch.hoc *
batch2.hoc *
checkirreg.hoc * *
col.hoc *
comppowspec.hoc *
condisconcellfig.hoc *
condisconpowfig.hoc *
declist.hoc *
decmat.hoc *
decnqs.hoc *
decvec.hoc *
default.hoc *
drline.hoc *
e2hubsdisconpow.hoc *
e2incconpow.hoc *
filtutils.hoc *
geom.hoc *
graphplug.hoc *
grvec.hoc *
init.hoc *
labels.hoc *
load.hoc *
local.hoc *
makepopspikenq.hoc *
matfftpowplug.hoc *
matpmtmplug.hoc *
matpmtmsubpopplug.hoc *
matspecplug.hoc *
network.hoc *
nload.hoc *
nqpplug.hoc *
nqs.hoc *
nqsnet.hoc *
nrnoc.hoc *
powchgtest.hoc *
python.hoc *
pywrap.hoc *
redE2.hoc *
setup.hoc *
shufmua.hoc *
simctrl.hoc *
spkts.hoc *
stats.hoc *
syncode.hoc *
vsampenplug.hoc *
xgetargs.hoc *
load_file("writedata.hoc") // Allow periodic saving of data to disk

print "Loading alz.hoc..."
// alz.hoc
// Mark Rowan, School of Computer Science, University of Birmingham, UK
// April 2012

// Code to Repeatedly delete neurons, allow synaptic compensation, and perform
// repeat+unique trials for calculating Fourier information (Crumiller et al. 2011).

// **********************************************************************
// Set buffertime in writedata.hoc to an appropriate length of time (e.g. 1600e3)
// run.hoc: SPKSZ should be at least 2500e3 elements (assuming buffertime=1600e3 and scale=1)
// run.hoc: should have prl(0,1) as the last line in the file to turn off printlist recording
// params.hoc: should have PreDur = segmentlength * (deletionstep/numcells) seconds (e.g. 160e3 s)
// **********************************************************************

// Repeat until all neurons deleted (at 1600s per segment = 160 000s = ~44hrs of data):
//   Delete deletionstep neurons (either randomly, or proportional to scale factor)
//   Allow AMPA and GABA synapses to compensate (doing it in only 1600s here, not "hours to days"!)
//   Present alternate 80s trials of 'unique' and 'repeat' stimulation for Fourier info calculation

// **** User-settable parameters ****
// Scaling params
declare("scaling", 1) // Set to 1 to enable compensatory homeostatic synaptic scaling
declare("activitytau", 100e3) // Activity sensor time constant
declare("activitybeta", 4.0e-8) // Scaling weight
declare("activitygamma", 1.0e-10) // Scaling integral controller weight
declare("scalingstart", buffertime) // Time after which to begin synaptic scaling (needs to be long enough for the activity sensors to reach approximately the real average firing rate)
declare("scaleexternal", 1) // Should we scale down the external inputs proportionally to the network deletion?
declare("scaleexternalrate", 0.25) // Vary the external scaledown constant

// Deletion params
declare("randomdelete", 1) // Set to 1 for uniform random deletion, or 0 for scaling-proportional deletion
declare("deletionstep", 5 * scale) // Delete this many cells per round

// Timing params
declare("segmentlength", buffertime) // How long is allocated to one round of deletion + compensation trials?
declare("infotriallength", 80e3) // Length (ms) of unique+repeat Fourier information trials
numcells = 470 * scale // Number of cells in the network (set in network.hoc:77 and :124)

// Recording params
declare("recording_interval", 100e3) // How many ms between scalefactor/activity/etc recordings
declare("recording_start", 5e3) // Start recording data after this time

// Define objects
objref remainingcellIDs, randgen, deletionList, auxFile, varsFile
strdef auxFileName
strdef varsFileName

proc setup() {
  // Set up list of remaining undeleted cells
  remainingcellIDs = new Vector(numcells)
  remainingcellIDs.indgen(1) // Generate elements from 0 to numcells-1

  // Initialise RNG
  randgen = new Random()

  // Set scaling parameters in INTF6

  // Create data file
  // filepath should already have been allocated in writedata.hoc
  sprint(auxFileName, "%s/%s", filepath, "aux")
  sprint(varsFileName, "%s/%s", filepath, "vars")
  auxFile = new File(auxFileName)
  varsFile = new File(varsFileName)
  header_written = 0 // Instruct write_scaling_data() to write vars file header

proc turn_on_scaling() {
  if (scaling) {
    printf("Turning on synaptic scaling\n")

proc deletionround() { local i, deleted, temp, gain
  printf("*** %d cells remaining, deleting %d\n", remainingcellIDs.size(), deletionstep)
  deletionList = new Vector() // Reset deletionList

  // Obtain list of cells to be deleted (either at uniform random, or proportional to compensation)
  deleted = 0
  while (deleted < deletionstep) {
    if (randomdelete) {
      // Delete at uniform random
      randgen.uniform(0, remainingcellIDs.size()) // Uniform-random over 0 to number of remaining cells
      temp = remainingcellIDs.x[int(randgen.repick())] // Get random cell index from remainingcellIDs
    } else {
      // TODO Delete randomly in proportion to compensation rates
      temp = 0

    if (!deletionList.contains(temp)) {
      deleted = deleted + 1
      printf("Deleting cell %d\n", temp)
      deletionList.append(temp) // Only append index if it's unique
      remainingcellIDs.remove(remainingcellIDs.indwhere("==", temp)) // Remove from list of remaining cells
  for vtr(&id,deletionList) {
    // Delete cells in deletionList
    printf("Cell %d scalefactor = %f\n", id, col.ce.o(id).get_scale_factor())
    // Write deletionlist and compensation rates of deleted cells to file
    col.ce.o(id).flag("dead",1) // Borrowed from network.hoc:dokill()
    //printf("%d\n", col.ce.o(id).get_id())

  if (scaleexternal) {
    scaledown() // Scale external input weights down as the network is deleted

  // While there are still cells to delete, queue another round of deletion
  if (remainingcellIDs.size() > deletionstep) {
    cvode.event(t + segmentlength, "deletionround()") 

proc scaledown() { local c,i,vhz,a,sy,ct,idx,sdx localobj vsy,nc
  // Reduce strength of external inputs linearly as deletion progresses.
  // Should be reasonable, as if other columns (represented by external inputs) are also
  // suffering atrophy, their output strengths (to this model's column) will be reduced too.
  vsy.append(GA) vsy.append(GA2) vsy.append(AM2) vsy.append(NM2)
  //setwmatex() // Reinitialise weights matrix
  //gain = remainingcellIDs.size() / numcells
  gain = 1 - ((1 - (remainingcellIDs.size() / numcells)) * scaleexternalrate)

  printf("External input gain: %f\n", gain)

  // For each CSTIM in lcstim
  for c = 0, lcstim.count-1 {
    i = 0 // Increment through the NetCon objects within this CSTIM
    // For each NetCon in ncl (based on col.hoc proc nsstim) -- based on col.hoc proc nsstim()
    for sdx=0,vsy.size-1 {
      sy=vsy.x(sdx) // go through synapse types
      for ct=0,CTYPi-1 if(col.numc[ct] && wmatex[ct][sy] && ratex[ct][sy]) {//go through cell types, check weights,rates of inputs 
        for idx = col.ix[ct], col.ixe[ct] {
          //printf("idx %d\n", idx)
          // For each cell of type ct (idx is id of the cell)
          // change the weights according to the new values in wmatex, scaled down by gain
          lcstim.o(c).ncl.o(i).weight(sy) = lcstim.o(c).wmatex[ct][sy] * gain
          //printf("c %d, i %d, weight %f\n", c, i, lcstim.o(c).ncl.o(i).weight(sy))
          i = i + 1 // Increment to next NetCon in the list

// TODO Check that CSTIMs are indeed alternating between 'fixed' and 'unique'
proc repeattrialcstim() { local i,j,inputseed
  // Set CSTIMs to use a sequence from a fixed seed for Fourier info calculation
  // At 80000ms per stim, this should give 10 * repeat and 10 * unique trials in 1600s
  cvode.event(t + infotriallength, "uniquetrialcstim()") // Queue next 'unique' stimulation trial

  printf("Initiate repeat trials for %d ms\n", infotriallength)
  inputseed = 1234 // Fixed seed

  for i=0,lcol.count-1 {
    for j=0,lcstim.o(i).vrse.size-1 {
      lcstim.o(i).vrse.x[j] = inputseed+j // Store random seeds for each nstim
    lcstim.o(i).initrands() // Reinitialise the external inputs' RNGs
    // TODO Does this restart CSTIM from t=0 (i.e. produce identical stim each time)
    // or does it continue from current t, thereby not producing identical stims?

proc uniquetrialcstim() { local i,j,inputseed
  // Set CSTIMs to use a sequence from current sim clock time 't' for Fourier info calculation
  // At 80000ms per stim, this should give 10 * repeat and 10 * unique trials in 1600s
  cvode.event(t + infotriallength, "repeattrialcstim()") // Queue next 'repeat' stimulation trial

  printf("Initiate unique trials for %d ms\n", infotriallength)  
  inputseed = t+1 // Seed set from clock (i.e. pseudo-randomly set) 

  for i=0,lcol.count-1 {
    for j=0,lcstim.o(i).vrse.size-1 {
      lcstim.o(i).vrse.x[j] = inputseed+j // Store random seeds for each nstim
    lcstim.o(i).initrands() // Reinitialise the external inputs' RNGs

proc write_scaling_data() { local k, id, act, trg, scl, type, dead

  if (!header_written) {
    // Write vars file header
    varsFile.printf("# ************* Runtime params *************\n")
    varsFile.printf("buffertime=%d\n", buffertime)
    varsFile.printf("numcells=%d\n", numcells)
    varsFile.printf("scaling=%d\n", scaling)
    varsFile.printf("scalingstart=%d\n", scalingstart)
    varsFile.printf("scaleexternal=%d\n", scaleexternal)
    varsFile.printf("scaleexternalrate=%f\n", scaleexternalrate)
    varsFile.printf("randomdelete=%d\n", randomdelete)
    varsFile.printf("deletionstep=%d\n", deletionstep)
    varsFile.printf("segmentlength=%d\n", segmentlength)
    varsFile.printf("infotriallength=%d\n", infotriallength)
    varsFile.printf("recording_interval=%d\n", recording_interval)
    varsFile.printf("t_start=%d\n", recording_start)
    varsFile.printf("activitytau=%e\n", activitytau)
    varsFile.printf("activitybeta=%e\n", activitybeta)
    varsFile.printf("activitygamma=%e\n", activitygamma)
    varsFile.printf("\n# Cell ID, cell type, activity sensor, target activity, scaling factor, is-dead\n")
    varsFile.printf("# Recorded every %d ms\n", recording_interval)
    header_written = 1
  // Open aux file for append

  // Record current time
  auxFile.printf("t = %f\n", t)

  // Write data to given file
  for k=0,numcells-1 {
    id = col.ce.o(k).get_id()
    act = col.ce.o(k).get_activity()
    trg = col.ce.o(k).get_target_act()
    scl = col.ce.o(k).get_scale_factor()
    type = col.ce.o(k).get_type()
    dead = col.ce.o(k).get_dead()
    auxFile.printf("%d,%d,%f,%f,%f,%d\n", id,type,act,trg,scl,dead)
    //printf("%d,%d,%f,%f,%f,%d\n", id,type,act,trg,scl,dead)

  // Close file
  // Queue next event
  cvode.event(t+recording_interval, "write_scaling_data()")

// Callback procedure to start the event queue
// NOTE: Only called once, after hoc file initialised and run() has been called
proc deletioneventqueue() {
  // Set up CSTIMs
  cvode.event(t, "uniquetrialcstim()") // Start alternating 'unique+random' trials
  cvode.event(scalingstart, "turn_on_scaling()") // Start scaling after initial delay to find activity
  cvode.event(t + segmentlength, "deletionround()") // Call deletionround() after segmentlength ms
  cvode.event(t + recording_start, "write_scaling_data()") // Start recording of scalefactors etc

declare("alzfih", new FInitializeHandler("deletioneventqueue()")) // Called as soon as INIT finishes

Loading data, please wait...