Computational Intelligence - May 2016 - 57

inferior temporal cortex suggests that
external stimuli can be represented by
responses of neural populations, and
encoded memory patterns are organized
in a hierarchy structure [13], [14].
The organization of memory is
closely associated with the learning process in the nervous system. Spike-timing-dependent plasticity (STDP) and
other spike-timing based learning
schemes are thought to be involved in
the formation of neural assemblies and
associative memory [15], [16]. In
addition, different learning algorithms
using spiking neuron have been proposed to study hetero-association [17]-
[21]. However, the formulation and
organization of auto-associative and episodic memories (Fig. 1) by virtue of
temporal coding and learning remain
underexplored.
In this work, we propose a hierarchically structured spiking neural network model, called spatio-temporal
memory (STM), which is able to study
the formation of neural assemblies and
the organizing principle. In this model,
sensory information travels upwards
along the hierarchical network during
the bottom-up information processing
(Fig. 2). With a spike-timing based
learning algorithm during the storing
phase, the model maps sensory information into neural assembly activities
to form memory items. We demonstrate that auto-associative memory is
formulated via fast STDP learning (in
Layer I and II) and episodic memory
produced by slow STDP learning (in
Layer II). The main contribution of this
model lies in that memory formulation
is implemented by temporal population
coding and temporal learning, which
are missing components in existing
memory models [22], [23]. In [24], a
hierarchical model based on spiking
neurons named Cortext was proposed.
Although timing of spikes is employed
in Cortext, it mainly targets visual recognition, while temporal coding and
memory organization are not exploited. The hierarchical temporal memory
(HTM) model [25], is successful in
emulating the associative and episodic
memory functions. However, it is not

Associative Memory

a A A

Hetero-Associative Memory

memor

Auto-Associative Memory

Episodic Memory

memory

Figure 1 Retrieval of associative memory and episodic memory. Hetero-associative memory
is a mapping from one pattern to another. Auto-associative memory associates the input pattern to itself. Episodic memory is a collection of experiences in time in a serial form.

Slow STDP
ω (22)
Layer II
Interneurons

ω (21) ......

Tempotron Rule

Fast STDP
ω (11)
Layer I

Tempotron Rule

ω (10)

......

External
Theta Oscillation

Input Layer

Encoded Stimuli

Excitatory
Inhibitory

Figure 2 The architecture of the STM model. Neurons forming neural assemblies and
enhanced lateral connections are illustrated in different colors. Fast STDP and slow STDP
learning algorithms are employed to adapt the connections within layers (Layer I and Layer II,
respectively), while tempotron learning rule is applied to the connections between layers.

able to simulate the memory formulation and organization process through
complex spatio-temporal dynamics of
spiking neurons.
In summary, the main features of the
STM model include:
1) It is the first comprehensive computational model that integrates spiking neural learning and memory
formulation dynamic process. In
contrast to existing memory models

focusing only on memory functions,
the STM model provides a low-level neural circuits based substrate that
is feasible to devise neuromorphic
cognitive systems in hardware, e.g.,
neuromorphic chips.
2) As a multi-layer hierarchical structure composed of spiking neurons,
the STM model is capable of analyzing and illustrating var ious
important computational primitives

may 2016 | IEEE ComputatIonal IntEllIgEnCE magazInE

Table of Contents for the Digital Edition of Computational Intelligence - May 2016