Parameter on TIDIGIT dataset

[qianhuiliu]

Hello!
Could you please introduce how to set the parameters in "network.yaml" in TIDIGIT dataset?
My settings are
simulation:
Ts: 1
tSample: 64
nTimeBins: 64
nSample: 100
neuron:
type: SRMALPHA
theta: 10
tauSr: 10.0
tauRef: 1.0
scaleRef: 2 # relative to theta
tauRho: 1 # relative to theta #0.43429448190325176
scaleRho: 1
training:
error:
type: NumSpikes #ProbSpikes #NumSpikes
probSlidingWin: 20 # only valid for ProbSpikes
tgtSpikeRegion: {start: 0, stop: 64} # only valid for NumSpikes and ProbSpikes
tgtSpikeCount: {true: 20, false: 5} # only valid for NumSpikes

But it only produces ~93.6% accuracy. Could you give me some advice?

Many thanks!

[bamsumit]

The performance on TIDIGIT depends on the spike encoding. What is the spike encoding you are using?

If you are using NTIDIGITS, the config and accuracy seems fine.

[qianhuiliu]

Thank you, Bam sumit. I conduct the experiment on TIDIGIT. My encoding is using MFCC+SOM, which was introduced in [Wu 2018]. The slayer network structure is set as 484-500-500-11, which also follows the paper said.

And I change the "nSample" in "network.yaml" to 12 and get an accuracy of 95.5%, which is still 4% lower than the paper said. I have no idea how to optimize it and hope you can give me some advice. Many thanks!

[bamsumit]

@qianhuiliu are you learning axonal delay as well?

[chenjiefighting]

Hello, I would like to ask how NTIDIGITS dataset is preprocessed. Could you please provide the corresponding code? Thank you.