[NEST Users] Re: Building a nest simulation for mnist dataset

Hi Aditya,

this doesn't seem to be a NEST issue so much as a conceptual issue - you have a network that is doing X but you feel it should be doing Y. I suspect that STDP is not the way forwards - this is a biologically motivated synaptic plasticity model which makes no promises about being able to learn anything, and it is typically quite hard to persuade it so to do. You might want to investigate e-prop instead, which is more derieved from supervised learning principles and does make such promises.

Here is a link to some e-prop tutorials that you could take a look at to see if this approach would help you:

https://nest-simulator.readthedocs.io/en/stable/auto_examples/eprop_plastici...

All the best,

Abigail

On 30.01.25 09:55, Aditya Srivastava wrote:

I have built a nest simulation for a binary dataset where we have two classes normal and anomalous . So i tried to built a model iun such a way so that the model learns to produce no spike in case of a normal data but when it sees a anomalous data it is producing the spikes. So this model was working fine. Now I am trying to move forward with it and try it on image dataset where I am using simple MNIST data and I have just taken image for one digit and trying to build a same kind of architecture where it should produce almost no spike when it sees a digit one and oin case of other digits it should produce spikes but I am facing issues in it . The model is not able to produce enough spikes and when I am creating more exc connections then spikes are not reducing with each epochs so can you please suggest me a better approach or some suggestion for my current approach.

# loading dataset from sklearn.datasets import fetch_openml import numpy as np

def preprocess_mnist_digit(digit=1): """ Load and preprocess the MNIST dataset to only include images of a specific digit. Normalize pixel values by dividing by 4 and flatten images.

 Args:
     digit (int): The digit to filter (e.g., 1 for digit '1').
     num_samples (int): Number of samples to retain.

 Returns:
     flattened_images (array): Flattened MNIST images of the specified digit.
     labels (array): Corresponding labels.
 """
 # Load MNIST data
 mnist = fetch_openml('mnist_784', version=1)
 data, labels = mnist['data'], mnist['target'].astype(int)  # Convert labels to integers

 # Filter images that correspond to the specified digit
 digit_indices = np.where(labels == digit)[0]  # Get indices where label is equal to the specified digit

 # Extract the corresponding images and labels
 filtered_images = data.iloc[digit_indices].to_numpy(dtype=float) / 4.0  # Normalize
 filtered_labels = labels[digit_indices]

 return filtered_images, filtered_labels

# Load only images of the digit '1' images, labels = preprocess_mnist_digit(digit=1)

# building simulation

nest.ResetKernel() # Simulation parameters dt = 0.01 # Simulation time step in ms nest.SetKernelStatus({"resolution": dt})

N = 784 # Number of neurons images=images[:10] gmax = 3 neuron_params = { "tau_m": 140.0, # Membrane time constant in ms "E_L": -70.6, # Resting potential (in mV) "V_th": -40.4, # Threshold potential (in mV) "V_reset": -70.6, # Reset potential (in mV) "V_m":0.0, }

nest.CopyModel("stdp_synapse", "stdp_synapse_exc", {"weight":gmax,"alpha":1,"Wmax":gmax, "tau_plus": 20.0,"lambda": -0.0005, "mu_minus": 0, "mu_plus": 0}) nest.CopyModel("stdp_synapse", "stdp_synapse_inh", {"weight":-gmax/(N-1),"alpha":1,"Wmax":-gmax, "tau_plus": 20.0,"lambda": 0.005, "mu_minus": 0, "mu_plus": 0})

G_spike_gen = nest.Create("poisson_generator",N) G_input = nest.Create("parrot_neuron", N)

input_spike_recorder = nest.Create("spike_recorder",N)

nest.Connect(G_input, input_spike_recorder,"one_to_one")

G_hidden = nest.Create("iaf_psc_delta", N,params=neuron_params) G_output = nest.Create("iaf_psc_delta", 1,params=neuron_params)

hidden_spike_recorder = nest.Create("spike_recorder",N)

nest.Connect(G_hidden, hidden_spike_recorder,"one_to_one")

S_in = nest.Connect(G_spike_gen, G_input, "one_to_one")

# Connect input to hidden layer with excitatory STDP synapses S_exc = nest.Connect(G_input, G_hidden,"one_to_one",syn_spec = "stdp_synapse_exc")

# Retrieve the neuron IDs for G_input and G_hidden input_ids = G_input.tolist() hidden_ids = G_hidden.tolist()

# Get the existing excitatory connections and extract their source-target ID pairs exc_connections = nest.GetConnections(G_input, G_hidden) exc_connection_pairs = set(zip(exc_connections.get("source"), exc_connections.get("target")))

for i in input_ids: for h in hidden_ids: if (i, h) not in exc_connection_pairs: nc1 = nest.NodeCollection([i]) nc2 = nest.NodeCollection([h]) nest.Connect(nc1, nc2, syn_spec= "stdp_synapse_inh")

S_out = nest.Connect(G_hidden, G_output,"all_to_all",syn_spec = {"weight": 40.4/N})

# training function def train_net(images, spike_tol=0):

 output_spike_recorder = nest.Create("spike_recorder")
 output_membrane_recorder = nest.Create("voltmeter", 1)
 nest.Connect(G_output, output_spike_recorder)   # Record output layer spikes
 nest.Connect(output_membrane_recorder, G_output, "one_to_one")
 while True:
     for image in images:
         for i in range(784):
             G_spike_gen[i].rate=image[i]
         nest.Simulate(500)
     for i in range(784):
         spikes = nest.GetStatus(hidden_spike_recorder[i], "events")[0]
         # if length of spikes is not zero then print the number of spikes and the times of spikes also print the index of the input_spike_recorder
         if len(spikes["times"]) != 0:
             print("Number of spikes in hidden_spike_recorder",i,":", len(spikes["times"]))
             print("Times of spikes in hidden_spike_recorder",i,":", spikes["times"])
             print("\n")

     output_spikes_num = len(output_spike_recorder.get()["events"]["times"])
     print("# of output spikes: " + str(output_spikes_num))
     # del recorder_for_epoch
     if output_spikes_num <= spike_tol:
         print("Training finished")
         break

 return output_membrane_recorder, output_spike_recorder

v, spk = train_net(images, spike_tol = 2) ___________________________________________ _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org

-- Prof. Dr. Abigail Morrison

Institute for Advanced Simulation (IAS-6) Jülich Research Center

&

Computer Science 3 - Software Engineering RWTH Aachen

http://www.fz-juelich.de/inm/inm-6 http://www.se-rwth.de

Office: +49 2428 8097504 Fax # : +49 2461 61-9460 Pronouns: she/her

------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------ Forschungszentrum Jülich GmbH 52425 Jülich Sitz der Gesellschaft: Jülich Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Stefan Müller Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), Dr. Stephanie Bauer (stellv. Vorsitzende), Prof. Dr. Ir. Pieter Jansens ------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------