Table 2.

Summary of investigated models on N-TIDIGITS18 dataset.

Feature type	Sensor	Task	Classifier	Accuracy (%)
MFCC		Digit	GRU RNN	97.90
Binned frames (fixed bins/sample)*	AMS1b	Digit	SVM	95.08
Constant time bins**	AMS1b	Digit	CNN	87.65
Constant time bins**	AMS1b	Digit	GRU RNN	82.82
Single events (raw data)	AMS1b	Digit	Phased LSTM	87.75
Data-driven time-binned features	AMS1b	Digit	Phased LSTM	91.25a
Constant time bins	AMS1b	Digit	GRU RNN	86.4
Exponential features	AMS1b	Digit	GRU RNN	90.9
Constant time bins	AMS1c	Digit	GRU RNN	88.6
Exponential features	AMS1c	Digit	GRU RNN	91.1
Constant time bins	AMS1b	Sequence	LSTM RNN	86.1b
Exponential features	AMS1b	Sequence	LSTM RNN	87.3b

The MFCC features are extracted from the original TIDIGITS dataset.

^aEvents from all neurons and both ears used in training.

^bLabel accuracy on sequences.

^*Abdollahi and Liu (2011).

^**Neil and Liu (2016).