428
J. Wu et al.
ent biologically inspired systems represent an important step towards an end-to-end
SNN-based automatic speech recognition system.
We note that the traditional RNN based system offers a competitive accuracy of
97.90% [42]; our proposed framework, however, is fundamentally different from
traditional deep learning approaches. It is worth noting that the network capacity and
classification accuracy of our framework can be further improved using multi-layer
SNNs.
3.3 Early Decision Making Capability
We note that the SNN-based classifier can identify temporal features within the
spatiotemporal spike pattern and generate an output spike as soon as enough discriminative evidence is accumulated. This cumulative decision-making process is
more biologically plausible, as it mimics how human makes decisions. A key benefit
of such a decision-making process is low latency. As shown in Fig. 2a, the SNN
classifier makes a decision before the whole pattern has been presented. On average,
the decision is made when only 50% of the input is presented.
Additionally, we conduct experiments on the SOM-SNN, RNN and LSTM models, whereby they are trained on the full input patterns but tested with only a partial
presentation of the input. The training label is provided to the RNN and LSTM
models at the end of each training sequence by default as it is not clear beforehand
when enough discriminative features have been accumulated. Likewise, the training
labels are provided at the end of input patterns for the SNN classifier. For testing,
we increase the duration of the test input pattern presented from 10 to 100% of the
actual duration, starting from the beginning of each pattern. As shown in Fig. 2b,
the classification accuracy as a function of the input pattern percentage increases
more rapidly for the SNN model. It achieves a satisfactory accuracy of 95.1% when
only 50% of the input pattern is presented, much higher than the 25.7% and 69.2%
accuracy achieved by the RNN and LSTM models respectively. For the RNN and
LSTM models to achieve early decision-making capability, one may require that the
models be trained with partial inputs or output labels provided at every time-step.
Therefore, SNN-based classifiers demonstrate great potential for real-time temporal
pattern classification, compared with state-of-the-art deep learning models such as
the RNN and LSTM.
3.4 Feature Representation of the SOM
To visualize the features extracted by the SOM, we plot the BMU activation sequences
and their corresponding trajectories on the SOM for a set of randomly selected
samples from class ‘bell5’, ‘bottle1’ and ‘buzzer’ in Fig. 3. We observe low intra-class
J. Wu et al.
ent biologically inspired systems represent an important step towards an end-to-end
SNN-based automatic speech recognition system.
We note that the traditional RNN based system offers a competitive accuracy of
97.90% [42]; our proposed framework, however, is fundamentally different from
traditional deep learning approaches. It is worth noting that the network capacity and
classification accuracy of our framework can be further improved using multi-layer
SNNs.
3.3 Early Decision Making Capability
We note that the SNN-based classifier can identify temporal features within the
spatiotemporal spike pattern and generate an output spike as soon as enough discriminative evidence is accumulated. This cumulative decision-making process is
more biologically plausible, as it mimics how human makes decisions. A key benefit
of such a decision-making process is low latency. As shown in Fig. 2a, the SNN
classifier makes a decision before the whole pattern has been presented. On average,
the decision is made when only 50% of the input is presented.
Additionally, we conduct experiments on the SOM-SNN, RNN and LSTM models, whereby they are trained on the full input patterns but tested with only a partial
presentation of the input. The training label is provided to the RNN and LSTM
models at the end of each training sequence by default as it is not clear beforehand
when enough discriminative features have been accumulated. Likewise, the training
labels are provided at the end of input patterns for the SNN classifier. For testing,
we increase the duration of the test input pattern presented from 10 to 100% of the
actual duration, starting from the beginning of each pattern. As shown in Fig. 2b,
the classification accuracy as a function of the input pattern percentage increases
more rapidly for the SNN model. It achieves a satisfactory accuracy of 95.1% when
only 50% of the input pattern is presented, much higher than the 25.7% and 69.2%
accuracy achieved by the RNN and LSTM models respectively. For the RNN and
LSTM models to achieve early decision-making capability, one may require that the
models be trained with partial inputs or output labels provided at every time-step.
Therefore, SNN-based classifiers demonstrate great potential for real-time temporal
pattern classification, compared with state-of-the-art deep learning models such as
the RNN and LSTM.
3.4 Feature Representation of the SOM
To visualize the features extracted by the SOM, we plot the BMU activation sequences
and their corresponding trajectories on the SOM for a set of randomly selected
samples from class ‘bell5’, ‘bottle1’ and ‘buzzer’ in Fig. 3. We observe low intra-class