Skip to content

Latest commit

 

History

History
683 lines (487 loc) · 22.6 KB

fp_synapticInput.rst

File metadata and controls

683 lines (487 loc) · 22.6 KB

Synaptic Input for Neurons

In this task you will investigate a single neuron receiving excitatory or inhibitory input. We begin with a situation in which the neuron does not fire in response to input, by disabling the reset mechanism and then in the last section investigate firing in response to input. The goal of the task is to familiarise yourself with the analog non-idealities that arise from a physical realisation of the idealised model equations. You will characterise different sources of noise in the system and get a feeling for the influence the hardware parameters on the temporal behaviour.

Setup

Generally speaking you will need to run this cell only at the beginning or after something is broken. The easiest way to escape from a broken state, will be to select "Kernel -> Restart" in the menu above and reexecute the following two cells.

We will also setup a folder to save experimental results in

import pathlib

results_folder = pathlib.Path('results/task3')
results_folder.mkdir(parents=True, exist_ok=True)

Plotting

%matplotlib inline

import matplotlib.pyplot as plt
import numpy as np

from pynn_brainscales.brainscales2 import Population

plt.style.use("_static/matplotlibrc")

def plot_membrane_dynamics(times: np.array, voltages: np.array):
    """

    :param times: sample times
    :param voltages: voltages to plot

    """
    plt.plot(times, voltages, color='grey')
    plt.xlabel("Wall clock time [ms]")
    plt.ylabel("ADC readout [a.u.]")

def plot_population_dynamics(population: Population, segment_id=-1, ylim=None):
    """
    Plot the membrane potential of the neuron in a given population view. Only
    population views of size 1 are supported.
    :param population: Population, membrane traces and spikes are plotted for.
    :param segment_id: Index of the neo segment to be plotted. Defaults to
                       -1, encoding the last recorded segment.
    :param ylim: y-axis limits for the plot.
    """
    if len(population) != 1:
        raise ValueError("Plotting is supported for populations of size 1.")
    # Experimental results are given in the 'neo' data format
    mem_v = population.get_data("v").segments[segment_id].irregularlysampledsignals[0]
    spikes = population.get_data("spikes").segments[-1].spiketrains[0]

    plt.plot(mem_v.times, mem_v, alpha=0.1, color='black')

    # indicate the spikes as red dots
    plt.scatter(spikes, np.max(mem_v)*np.ones_like(spikes), color='red')
    plt.xlabel("Wall clock time [ms]")
    plt.ylabel("ADC readout [a.u.]")
    if ylim:
        plt.ylim(ylim)

Synaptic Parameters

import numpy as np
from pathlib import Path
import quantities as pq

import pynn_brainscales.brainscales2 as pynn

from pynn_brainscales.brainscales2 import Population
from pynn_brainscales.brainscales2.standardmodels.cells import SpikeSourceArray
from pynn_brainscales.brainscales2.standardmodels.synapses import StaticSynapse

Create a neuron and inject a single spike.

# save results of this experiment in subfolder
exercise_folder = results_folder.joinpath('synaptic_parameters')
exercise_folder.mkdir(parents=True, exist_ok=True)

Experiments

Adjust the synaptic parameters and display the membrane traces.

  • What do the different parameters control?

By setting the "weight" parameter to be negative the neuron will receive inhibitory synaptic input.

  • What differences compared to the excitatory input can you observe?

Now enable the membrane threshold and adjust it in such a way that you ellicit one spike.

from ipywidgets import interact, IntSlider
from functools import partial
IntSlider = partial(IntSlider, continuous_update=False)
import tqdm.notebook as tqdm


@interact(
    tau=IntSlider(min=1, max=1022, step=1, value=26),
    gm=IntSlider(min=1, max=1022, step=1, value=1022),
    weight=IntSlider(min=-63, max=63, step=1, value=63),
    n_runs=IntSlider(min=1, max=20, step=1, value=1),
    threshold_v_threshold=IntSlider(min=0, max=600, step=1, value=300),
)
def experiment(
    tau,
    gm,
    weight,
    n_runs,
    threshold_v_threshold,
    threshold_enable=False,
    plot=True,
    savefig=False,
    results=False
):
    synapse_type = 'inhibitory' if weight < 0 else 'excitatory'
    # parameters
    experiment_duration = 0.4  # ms (hw domain)
    pynn.setup(initial_config=calib)
    neuron = pynn.Population(1, pynn.cells.HXNeuron(
        excitatory_input_i_bias_tau=tau,
        excitatory_input_i_bias_gm=gm,
        inhibitory_input_i_bias_tau=tau,
        inhibitory_input_i_bias_gm=gm,
        # neuron to be observed: disable threshold such we can observe PSP shape
        threshold_enable=threshold_enable,
        threshold_v_threshold=threshold_v_threshold
    ))
    neuron.record(["v", "spikes"])

    # external input
    spike_times = [experiment_duration / 2]
    pop_input = pynn.Population(1,
        pynn.cells.SpikeSourceArray(spike_times=spike_times))

    synapse = pynn.standardmodels.synapses.StaticSynapse(weight=weight)
    proj = pynn.Projection(pop_input, neuron, pynn.AllToAllConnector(),
                       synapse_type=synapse, receptor_type=synapse_type)

    # run experiment
    voltages = []
    mems = []
    min_shape = 2**32

    for run in tqdm.trange(n_runs, leave=False):
        pynn.run(experiment_duration)
        mem = neuron.get_data("v").segments[-1].irregularlysampledsignals[0]
        mems.append(mem)
        voltages.append(np.array(mem))
        min_shape = min(mem.shape[0], min_shape)
        pynn.reset()

    voltages = [v[:min_shape] for v in voltages]
    voltages = np.stack(voltages)

    if plot:
        plot_membrane_dynamics(mems[-1].times[:min_shape], np.mean(voltages, axis=0))

    if savefig:
        fig.savefig(exercise_folder.joinpath('membrane_trace.png'))

    pynn.end()
    if results:
        return mems, voltages, mems[-1].times[:min_shape]

Solution

Please enter your answers to the questions above here...

Normalising the Membrane Dynamics

We can use the experiment above to determine the minimum and maximum of the membrane dynamics over a number of runs.

def determine_vmax_vmin(tau, gm, weight, v_threshold, n_runs):
    mems, voltages, _ = experiment(tau=tau, gm=gm, weight=weight, n_runs=n_runs,    threshold_enable=True, threshold_v_threshold=v_threshold, results=True, plot=False)
    v_max, v_min = np.max(voltages[:,:,0]), np.min(voltages[:,:,0])
    return v_min, v_max

v_min, v_max = determine_vmax_vmin(1, 1022, weight=63, v_threshold=300, n_runs=10)

This can be used to normalise the membrane voltage measurements. Note that this normalisation depends on the choices of hyperparameters $\tau$, $g_m$, $v_\mathrm{th}$ used to obtain it. The idea is to map $v_\mathrm{th}$ to $1$ and $v_\mathrm{reset}$ to $0$, although this will only be approximately true for all membrane traces.

def normalise_voltage(v_min, v_max):
    def normalise(v):
        delta_v = v_max - v_min
        return 1/delta_v * (v - v_min)

    return normalise

nv = normalise_voltage(v_min, v_max)

Trial-to-Trial Variations

taus = np.arange(1,1022,100)
gm = 1022
weight = 63
n_runs = 5
max_voltages = []
baselines = []

for tau in tqdm.tqdm(taus):
    m, v, _ = experiment(tau=tau, gm=gm, weight=weight, n_runs=n_runs, threshold_v_threshold=300, plot=False, results=True)
    v_baseline = v[:100].mean()
    mean = np.mean(v, axis=0)
    max_v = np.max(mean, axis=0)
    max_voltages.append(max_v[0])
    baselines.append(v_baseline)
plt.plot(taus, np.array(max_voltages) - np.array(baselines))
tau = 26
gm = 1022
weight = 63
n_runs = 200 # each run takes some time (with this value it should finish in less than 1-2 minutes)

m, v, _ = experiment(tau=tau, gm=gm, weight=weight, n_runs=n_runs, threshold_v_threshold=300, plot=False, results=True)
v_baseline = v[:100].mean()
max_v = np.max(v-v_baseline, axis=1)

def histogram(heights):
    fig, ax = plt.subplots()
    ax.set_xlabel("height of peaks [MADC]")
    ax.set_ylabel("counts")

    ax.hist(heights)

    return fig

fig = histogram(max_v)

We can also sweep weight dependence of synaptic input

tau = 26
gm = 1022
n_runs = 10
means = []
n_sub = 20
ws = np.linspace(-64,64,n_sub)
category_colors = plt.get_cmap('coolwarm')(0.5+ws/128)

max_mean_v = []

for idx in tqdm.trange(n_sub):
    m, v, times = experiment(tau=tau, gm=gm, weight=ws[idx], n_runs=n_runs, threshold_v_threshold=300, plot=False, results=True)
    v_baseline = v[:100].mean()

    for j in range(n_runs):
        plt.plot(times, v[j] - v_baseline, color=category_colors[idx], alpha=.1)

    plt.plot(times, np.mean(v, axis=0) - v_baseline, color=category_colors[idx])

    sign = -1 if ws[idx] < 0 else 1

    max_mean_v.append(sign*np.max(np.abs(np.mean(v, axis=0) - v_baseline)))

plt.xlabel("Wall clock time [ms]")
plt.ylabel("ADC readout [a.u.]")

Exercises

  • Plot the maximum of the voltages we obtained

Solution

Please enter your solution here...

.. only:: Solution

    .. code:: ipython3

        plt.plot(ws, max_mean_v)
        plt.ylabel("ADC readout [a.u.]")
        plt.xlabel("Weight")

.. only:: not Solution

    .. code:: ipython3

        ... # TODO



Fixed-Pattern Noise

Investigate the fixed pattern noise between synapses/synapse drivers.

def extract_heights(voltage, spike_times):
    '''
    Extract the PSP heights from a membrane trace where several inputs
    where injected one after another.

    :param voltage: Recorded membrane trace
    :param spike_times: Input spike times for the different inputs in ms
        (hw domain).
    '''

    # determine baseline voltage
    v_baseline = voltage[:100].mean()

    # extract for each synapse the maximum in the membrane trace and
    # substact the baseline voltage
    heights = []

    # determine how many samples are recorded between inputs
    idx_first_input = np.argmin(np.abs(voltage.times - spike_times[0] * pq.ms))
    samples_per_input = 2 * idx_first_input

    # loop over inputs
    for n_input in range(len(spike_times)):
        start_stop = (np.array([0, 1]) + n_input) * samples_per_input
        input_slice = slice(start_stop[0], start_stop[1])
        height = np.max(voltage[input_slice]) - v_baseline
        heights.append(float(height))

    return heights
# save results of this experiment in subfolder
exercise_folder = results_folder.joinpath('fixed_pattern_noise')
exercise_folder.mkdir(parents=True, exist_ok=True)

def fixed_pattern_noise_experiment():
    # parameters
    n_synapses = 100
    weight = 63
    time_between_inputs = 0.4

    pynn.setup(initial_config=calib)

    # neuron to be observed: disable threshold such we can observe PSP shape
    neuron = pynn.Population(1, pynn.cells.HXNeuron(threshold_enable=False))
    neuron.record("v")

    # external input
    # since we want to test several external synapses, we create a population
    # of external neurons which spike one after another
    spike_times = (np.arange(n_synapses) + 0.5) * time_between_inputs
    # SpikeSourceArray expects a list of spike times for each neuron in the
    # population -> reshape
    pop_input = pynn.Population(n_synapses,
        pynn.cells.SpikeSourceArray(spike_times=spike_times.reshape([-1, 1]).tolist()))

    synapse = pynn.standardmodels.synapses.StaticSynapse(weight=weight)
    proj = pynn.Projection(pop_input, neuron, pynn.AllToAllConnector(),
                           synapse_type=synapse, receptor_type='excitatory')

    # run experiment
    pynn.run(n_synapses * time_between_inputs)

    # save membrane trace
    mem = neuron.get_data("v").segments[-1].irregularlysampledsignals[0]
    np.savetxt(exercise_folder.joinpath('membrane_trace.txt'), mem.base)
    pynn.end()

    return mem, spike_times

Exercises

Extract the PSP height from the trace and plot it in a histogram. You can use the functions you implemented above.

Hints: * The function to plot histograms with matplotlib is called hist

Solution

Please enter your solution here...

.. only:: Solution

    .. code:: ipython3

        mem, spike_times = fixed_pattern_noise_experiment()
        heights = extract_heights(mem, spike_times)
        plt.hist(heights)
        plt.savefig(exercise_folder.joinpath('histogram.png'))

.. only:: not Solution

    .. code:: ipython3

        mem, spike_times = fixed_pattern_noise_experiment()
        heights = ...
        # TODO plot and save histogram

Stacking of PSPs

@interact(exc_weight=IntSlider(min=0, max=63, step=1, value=31),
          inh_weight=IntSlider(min=0, max=63, step=1, value=31),
          tau=IntSlider(min=1, max=1022, step=1, value=26),
          gm=IntSlider(min=1, max=1022, step=1, value=1022),
          isi=IntSlider(min=10, max=100, step=5, value=50),
          n_runs=IntSlider(min=1, max=10, step=1, value=1)
          )
def run_experiment(exc_weight: int, inh_weight: int, tau: float, gm: float, isi: float, n_runs: int, plot = True, return_results = False):
    '''
    Run external input demonstration on BSS-2.

    Adjust weight of projections, set input spikes and execute experiment
    on BSS-2.

    :param exc_weight: Weight of excitatory synaptic inputs, value range
        [0,63].
    :param inh_weight: Weight of inhibitory synaptic inputs, value range
        [0,63].
    :param isi: Time between synaptic inputs in microsecond (hardware
        domain)
    '''

    plt.figure()
    plt.title("Fourth experiment: External stimulation")

    pynn.setup(initial_config=calib)

    # use calibrated parameters for neuron
    stimulated_p = pynn.Population(1, pynn.cells.HXNeuron(
        excitatory_input_i_bias_tau=tau,
        excitatory_input_i_bias_gm=gm,
        inhibitory_input_i_bias_tau=tau,
        inhibitory_input_i_bias_gm=gm
    ))
    stimulated_p.record(["v", "spikes"])

    # calculate spike times
    wait_before_experiment = 0.01  # ms (hw)
    isi_ms = isi / 1000  # convert to ms
    spiketimes = np.arange(6) * isi_ms + wait_before_experiment

    # all but one input are chosen to be exciatory
    excitatory_spike = np.ones_like(spiketimes, dtype=bool)
    excitatory_spike[1] = False

    # external input
    exc_spikes = spiketimes[excitatory_spike]
    exc_stim_pop = pynn.Population(1, SpikeSourceArray(spike_times=exc_spikes))
    exc_proj = pynn.Projection(exc_stim_pop, stimulated_p,
                               pynn.AllToAllConnector(),
                               synapse_type=StaticSynapse(weight=exc_weight),
                               receptor_type="excitatory")

    inh_spikes = spiketimes[~excitatory_spike]
    inh_stim_pop = pynn.Population(1, SpikeSourceArray(spike_times=inh_spikes))
    inh_proj = pynn.Projection(inh_stim_pop, stimulated_p,
                               pynn.AllToAllConnector(),
                               synapse_type=StaticSynapse(weight=-inh_weight),
                               receptor_type="inhibitory")

    # run experiment

    # run experiment
    voltages = []
    mems = []
    min_shape = 2**32
    experiment_duration = 0.6

    for run in range(n_runs):
        pynn.run(experiment_duration)
        mem = stimulated_p.get_data("v").segments[-1].irregularlysampledsignals[0]
        if plot:
            plot_population_dynamics(stimulated_p, ylim=(100, 600))
        mems.append(mem)
        voltages.append(np.array(mem))
        min_shape = min(mem.shape[0], min_shape)
        pynn.reset()

    voltages = [v[:min_shape] for v in voltages]
    voltages = np.stack(voltages)

    if return_results:
        return voltages

Exercises

  • Adjust the time between the synaptic inputs and investigate when the neuron is firing.
  • Save a plot in which you can observe firing behaviour and save the parameters.

Solution

Document your observations here...

Behaviour under Stimulation by a Poisson Process

So far we have only looked at a single neuron receiving deterministic input. In this last exercise for this section of the lab course, we will look at one neuron receiving "stochastic" input from two sources: an inhibitory and excitatory input source emitting spikes according to a poisson process with intensity $\lambda_e,\lambda_i$. Poisson processes occur frequently in nature and have wide applications in many areas. One key point of this exercise therefore is to give you an idea how a spiking neuron running on Neuromorphic Hardware could naturally and efficiently process event data.

def poisson_spike_times(lambda0 = 100, delta_t = 0.6):
    n_spikes = np.random.poisson(lambda0 * delta_t)
    times = delta_t * np.random.uniform(0, 1, n_spikes)
    return np.sort(times)

The following code defines the experiment:

@interact(exc_weight=IntSlider(min=0, max=63, step=1, value=31),
      inh_weight=IntSlider(min=0, max=63, step=1, value=31),
      lambda_exc=IntSlider(min=1, max=200, step=1, value=50),
      lambda_inh=IntSlider(min=1, max=200, step=1, value=50),
      n_runs=IntSlider(min=1, max=10, step=1, value=1)
      )
def run_experiment(exc_weight: int, inh_weight: int, lambda_exc: float, lambda_inh:float, n_runs: int, plot = True, return_results = False):
    '''
    Run external input demonstration on BSS-2.

    Adjust weight of projections, set input spikes and execute experiment
    on BSS-2.

    :param exc_weight: Weight of excitatory synaptic inputs, value range
        [0,63].
    :param inh_weight: Weight of inhibitory synaptic inputs, value range
        [0,63].
    :param lambda_exc: excitatory poisson process intensity
    :param lambda_inh: inhibitory poisson process intensity
    '''

    plt.figure()
    plt.title("Fourth experiment: External stimulation")

    pynn.setup(initial_config=calib)

    # use calibrated parameters for neuron
    stimulated_p = pynn.Population(1, pynn.cells.HXNeuron())
    stimulated_p.record(["v", "spikes"])

    # calculate spike times
    experiment_duration = 0.6
    wait_before_experiment = 0.01  # ms (hw)

    # external input
    exc_stim_pop = pynn.Population(1, SpikeSourceArray(spike_times=[]))
    exc_proj = pynn.Projection(exc_stim_pop, stimulated_p,
                               pynn.AllToAllConnector(),
                               synapse_type=StaticSynapse(weight=exc_weight),
                               receptor_type="excitatory")

    inh_spike_source = SpikeSourceArray(spike_times=[])
    inh_stim_pop = pynn.Population(1, inh_spike_source)
    inh_proj = pynn.Projection(inh_stim_pop, stimulated_p,
                               pynn.AllToAllConnector(),
                               synapse_type=StaticSynapse(weight=-inh_weight),
                               receptor_type="inhibitory")

    # run experiment

    # run experiment
    voltages = []
    mems = []
    min_shape = 2**32


    for run in tqdm.trange(n_runs):
        inh_spikes = poisson_spike_times(lambda_inh, experiment_duration) + wait_before_experiment
        inh_stim_pop.set(spike_times = inh_spikes)
        exc_spikes = poisson_spike_times(lambda_exc, experiment_duration) + wait_before_experiment
        exc_stim_pop.set(spike_times = exc_spikes)

        pynn.run(experiment_duration)
        mem = stimulated_p.get_data("v").segments[-1].irregularlysampledsignals[0]
        if plot:
            plot_population_dynamics(stimulated_p, ylim=(100, 600))
        mems.append(mem)
        voltages.append(np.array(mem))
        min_shape = min(mem.shape[0], min_shape)
        pynn.reset()

    voltages = [v[:min_shape] for v in voltages]
    voltages = np.stack(voltages)

    if return_results:
        return voltages

Exercises

Explore the parameters offered by the experiment and describe your observations. - Consider multiple runs (n > 1), how would you characterise the behaviour over multiple runs. - What happens if the intensity of the inhibitory and excitatory input differ widely? - What happens if you change the weight of the inhibitory / excitatory input? - What quantitative ways of analysing the observed behaviour come to your mind?

Solutions

Please give your solutions here...

Analysis of Stimulation by a Poisson Process

The following gives you a way to record membrane voltage traces for given paramters over many runs (n_runs = 1000)

voltages = run_experiment(exc_weight=31, inh_weight=31, lambda_exc=50, lambda_inh=50, n_runs=1000, return_results=1)

Exercises

  • Plot a histogram of the voltages, for different parameter settings. Consider both population (over all runs) and a histograms over time for a single run.
  • How could you analyse the observed distribution further?

Solution

Please type your solution here...

Hint: Consider plotting the histogram with density=True).

.. only:: Solution

    .. code:: ipython3

        plt.hist(voltages[:,:,0].flatten(), bins=100, density=True)

.. only:: not Solution

    .. code:: ipython3

        # TODO
        plt.hist(...)