Dear Mr. Rueckauer

In your publication over TTFS https://www.zora.uzh.ch/id/eprint/168616/9/168616.pdf it is visible that the SNN with TTFS performs better or comparable to the rate based. However, if i change on the MNIST dataset the spikecode to ttfs then i have a drop of the accuracy of 10% (in the case of the fashionMNIST even 20%). Do you know why this happen? Following the config:
config = configparser.ConfigParser()
config['paths'] = {
'path_wd': WORKING_DIR,
'dataset_path': DATASET_DIR,
'filename_ann': MODEL_NAME,
'runlabel': MODEL_NAME+'_'+str(NUM_STEPS_PER_SAMPLE)
}
config['tools'] = {
'evaluate_ann': True,
'parse': True,
'normalize': True,
'simulate': True,
'convert' : True
}
config['conversion'] = {
'softmax_to_relu': True,
'spike_code': 'ttfs',
}
config['simulation'] = {
'simulator': 'INI',
'duration': NUM_STEPS_PER_SAMPLE,
'num_to_test': NUM_TEST_SAMPLES,
'batch_size': BATCH_SIZE,
'keras_backend': 'tensorflow'
}
config['output'] = {
'verbose': 2,
'plot_vars': {
'input_image',
'spiketrains',
'spikerates',
'spikecounts',
'operations',
'normalization_activations',
'activations',
'correlation',
'v_mem',
'error_t'
},
'overwrite': True
}

Write the configuration file

config_filepath = os.path.join(WORKING_DIR, 'config')
with open(config_filepath, 'w') as configfile:
config.write(configfile)

from snn_toolbox.