1 Introduction
2 Related work
2.1 Facial expression recognition
2.2 Speech emotion recognition
2.2.1 Mel frequency cepstral coefficients (MFCCs)
2.2.2 Spectral centroid
2.2.3 Pitch
2.2.4 Energy
2.2.5 Classifiers for SER
2.2.6 Deep learning
2.2.7 Bio-inspired approaches
3 Proposed approach
3.1 Introduction to spiking neural networks
-
Feedforward: In this topology, information flows in one direction with no feedback connection. These kinds of topology are usually used in SNN to model low-level sensory systems, such as vision systems. They have also been used for binding tasks such as spatio-temporal spikes or spike synchronisation (Tapson et al. 2013; Sporea and Grüning 2012; Tavanaei and Maida 2017).
-
Recurrent: Neurons interact through feedback connections, where a dynamic temporal behaviour represents the network. Although this topology is harder to compute, it can have higher computational power. Recurrent architectures are particularly useful for modelling or analysing dynamic objects. However, it is computationally more challenging to apply supervised learning on this type of architecture (Demin and Nekhaev 2018). Recurrent architectures can also be applied to investigate extensive population activities and analyse neuronal populations dynamics.
-
Supervised learning that is achieved through applying Hebbian learning. The supervision is done through a spike-based Hebbian process by reinforcing the post-synaptic neuron in order to fire at preset timing and not spike at other times. The reinforcement signal is transmitted through synaptic currents (Knudsen 1994).
-
Unsupervised learning follows the basic Hebbs law, where neurons that fire together are connected (Hebb 1962). Automatic reorganisation of connection in the Hebbian learning permits the ability of unsupervised learning with various potential applications, such as clustering or pattern recognition. Unsupervised learning with Hebbian formula enables learning of distinct patterns without using classes labels or having a specific learning goal (Hinton et al. 1999; Bohte et al. 2002; Grüning and Bohte 2014).
-
Reinforcement learning that enables learning directly from the environment where SNN includes a rewarding signal spike (Farries and Fairhall 2007).