Neurosymbolic Spike Concept Learner towards Neuromorphic

General Intelligence

Ahmad Najiy Wahab, Khaled Mahbub and Abdel-Rahman Tawil

School of Computing and Digital Technology, Birmingham City University, Birmingham, U.K.

Keywords: Neuromorphic General Intelligence, Spiking Neural Networks, Functional Plasticity, Structural Plasticity,

Neurosymbolic, Representation Learning, Concept Learning.

Abstract: Current research in the area of concept learning makes use of deep learning and ensembles methods to learn

concepts. Concept learning allows us to combine heterogeneous entities in data which could collectively

identify as individual concepts. Heterogeneity and compositionality are crucial areas to explore in machine

learning as it has the potential to contribute profoundly to artificial general intelligence. We investigate the

use of spiking neural networks for concept learning. Spiking neurones inclusively model the temporal

properties as observed in biological neurones. A benefit of spike-based neurones allows for localised learning

rules that only adapts connections between relevant neurones. In this position paper, we propose a technique

allowing dynamic formation of synapse (connections) in spiking neural networks, the basis of structural

plasticity. Achieving dynamic formation of synapse allows for a unique approach to concept learning with a

malleable neural structure. We call this technique Neurosymbolic Spike-Concept Learner (NS-SCL). The

limitations of NS-SCL can be overcome with the neuromorphic computing paradigm. Furthermore,

introducing NS-SCL as a technique on neuromorphic platforms should motivate a new direction of research

towards Neuromorphic General Intelligence (NGI), a term we define to some extent.

1 INTRODUCTION

Neuromorphic computing (NC) is introducing a new

computing paradigm along with a class of processors

that differs from conventional computing to simulate

neural networks. The neural networks that NC adopts

behave as closely as possible to spike-based neurones

in biology. Initially the goal of NC was to simulate

the brain with large scale integrations of hardware-

based neurones. With the prominent advancement of

deep learning algorithms, dedicated processors or

Neural Processing Units (NPU) are being developed

to accelerate machine learning algorithms.

To mention a few NPUs and their use cases, the

Tensor Processing Unit (TPU) by Google have been

developed to accelerate deep learning algorithms for

consumers through Google’s cloud AI compute

services (Sengupta, Kubendran, Neftci, & Andreou,

2020); Vision Processing Units (VPU) have been

developed to serve as co-processors to accelerate

vision compute tasks and a few can perform image

inference tasks (Barry & Riordan, 2015; Rivas-

Gomez, Pena, Moloney, Laure, & Markidis, 2018);

Field Programmable Gate Arrays (FPGA) have been

used to conduct research on neural networks and their

various learning mechanisms (Lammie, Hamilton,

Van Schaik, & Azghadi, 2019; Perez-Peña, Cifredo-

Chacon, & Quiros-Olozabal, 2020; Rosado-Muñoz,

Bataller-Mompeán, & Guerrero-Martínez, 2012), but

these applications are demonstrated towards visual

pattern recognition (Liu & Yue, 2019).

It is important to note that majority of NPUs are

currently being developed to accelerate existing

neural network algorithms. Only a few are being

developed at the frontier of brain-inspired computing

research, that focuses on biologically plausible neural

models. Plausible neural models constitute of

neurones that neurologically functions much closer to

neurones observed in our central nervous system

(CNS). These plausible neurones are the spike-based

neurones which makes the 3

generation of neural

networks. Intel’s Loihi, IBM’s TrueNorth,

SpiNNaker and Neurogrid (Boahen, 2017; Davies et

al., 2018; Debole et al., 2019; Painkras et al., 2013)

are few examples that could be considered as NPUs

with 3

generation neural networks but distinctively

are being considered as Neuromorphic Processors.

The defining aspect of Neuromorphic Processors are

1168

Wahab, A., Mahbub, K. and Tawil, A.

Neurosymbolic Spike Concept Learner towards Neuromorphic General Intelligence.

DOI: 10.5220/0010339911681176

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 1168-1176

ISBN: 978-989-758-484-8

 2021 by SCITEPRESS – Science and Technology Publications, Lda. All r ights reserved

their scalability and massively parallel compute

potential to simulate spiking neurones at grand scales.

With neuromorphic devices on the rise, more

specialised devices/hardware is being developed to

host and accelerate various neural networks. We

acknowledge that these devices increase the

performance of such networks but, from a much

broader perspective in the domain of neuromorphic

computing, there is relatively slow advancements

being made on the algorithmic developments for what

could potentially be machine learning techniques

unique to the neuromorphic platforms.

It is highly probable, perhaps inevitable, that the

future of artificial intelligence will have mankind in

an age where general intelligence machines are more

like human beings contributing to society: reasoning

with and learning from its environment causing real

world effects. To reach such an age, cognition and

behavioural dynamics in general intelligence

machines should exhibit the likeness to behaviours of

human beings. This may be achieved by somewhat

imitating the underlying cognitive processes or

neurological processes that is observed in our central

nervous systems. Alternatively, purely speculative

developments that do not imitate nature may also lead

to general AI, this raises the questions as to which is

more favourable for the future of general AI. We may

perhaps trust more on general AI that work closer to

our biology than a more obscure forms of general AI

that we cannot relate to in processing rationality.

Artificial general intelligence (AGI) is considered

the holy grail of AI for decades and has been the core

motivation of machine learning since the birth of the

field. AGI is still anticipated to bring about

revolutionary advancement to science and

technology, with a wave of machine reasoning.

Initially, AGI is considered as intelligence expressed

by machines in contrast to natural intelligence

expressed by humans. Nevertheless, general

intelligence can also be considered loosely yet sub-

specifically as a form of intelligence that possess

cross-domain expertise.

Considering general intelligence in the context of

neuromorphic computing, it is reasonable to define a

very specific branch of artificial general intelligence

that is achieved through Neuromorphic means as a

form of Neuromorphic General Intelligence (NGI).

General intelligence can be achieved through

cognitive models consisting of various machine

learning techniques, formal methods and algorithms

like State, Operator, And Result (SOAR), Adaptive

Control Thought – Rational (ACT-R) and Learning

Intelligence Decision Agent (LIDA) (Bruckner,

Zeilinger, & Dietrich, 2012). We can differentiate

NGI as an approach of general intelligence emerging

from Neuromorphics, spike-based neural computing

platforms.

In this work we investigate the possibility of

delivering general intelligence on the neuromorphic

computing paradigm. Specifically, we propose to

investigate a method of structural and functional

plasticity for learning on a neurosymbolic spike-

based network to achieve concept learning on

neuromorphic platforms. Concept learning is a crucial

element to inaugurate generality in machine

intelligence.

Key questions this work will address towards NGI:

 How can spike-based networks achieve concept

learning?

 How will such spike-based concept learners be

employed in future neuromorphic platforms?

 Will spike-based concept learners inform any

design specification for neuromorphic hardware?

The real-world use cases – for concept learning

emphasising structural plasticity – can be in situations

where intelligent dynamical systems are required that

are extensible, flexible and function in real-time. The

world of Internet of Things (IoT) are such situations

as new sensors are constantly added to extend a

system in its wide variety of uses. Applying concept

learning in IoT situations allows us to form real-time

associations between various sensors regardless of

sensor type. The supposed intelligent dynamical

systems require algorithms that are unconstrained, we

have proposed structural plasticity for our approach

to adapt with extensibility as associations can form

with new components whilst existing associations

remains unaffected yet functional. The intelligence

aspect of our approach as a concept learner lies in the

neurosymbolic space. Concepts could consist of

entities in the symbolic space, representing

composites derived by heterogeneous data streamed

from different IoT sensors. The universality of

neurosymbolic space due to the heterogeneity enables

inter-correlation of information across different

sensors in an IoT ecosystem. For IoT sensor, data

requires spike-encoding for spike-based neural

processing. The encoding provides a standard

mechanism for processing data further promoting the

universality of symbolic space.

The rest of this paper will be presented as follows.

In section 2, we cover related works on inter-domain

knowledge, representation and concept learning. In

section 3, we briefly describe different aspects of our

algorithm and their rationale to realise our take on

concept learning with spiking neurones. In section 4,

Neurosymbolic Spike Concept Learner towards Neuromorphic General Intelligence

1169

we discuss the potential implications of our

techniques and the prospect of our approach to further

research in NGI.

2 RELATED WORK

Within the domain of computer science, machine

learning has given machines the ability to exhibit

intelligence and in cases excel humans in certain

intelligence tasks. However, modern AI in the current

academic standpoint are considered as narrow AI

(weak AI), excelling at a very specific intelligence

task at which they were purposefully trained for

(Fjelland, 2020). Narrow AI are domain specific,

requires human intervention to develop with and

utilise. Strong AI are not necessarily domain specific

because they are, as they should be, domain-neutral

(multi-domain). Strong AI should fundamentally

learn in an unsupervised general manner, thus, are

educated into domains they have been exposed to, as

opposed to being trained objectively and specifically.

With cross-domain knowledge in intelligent

systems, agents could tackle problems in a more

rational way by addressing new information with

consideration of patterns acquired from past

observations inter-correlated across domains. This

feat mirrors the inductive process facilitating

reasoning, to make conclusions based on one or

several evidence, in all its various forms, from past

observations and experiences. Concept learning is a

strategy of combining features or attributes

(fundamental pieces of information) which

collectively identify as one distinct concept. The goal

of concept learning, we assume in our context, is the

ability to capture complex patterns based on past

observations. The complexity of constituents making

complex patterns may be derived from the

heterogeneity of data. Concept learning is crucial to

achieving the domain-unconstrained learning

complimentary to general forms of intelligence.

Neural network algorithms are a sophisticated

development that has given us the techniques to

perform classifications and predictions. However,

neural network as a technique alone are obscure in the

sense that knowing the basis for which it took to reach

its conclusions are near impossible. The reasoning of

such is behind a ‘wall of matrices’ with little to no

semantic grounds. This has introduced a branch of AI

known as explainable AI (XAI) (Barredo Arrieta et

al., 2020) in which the goal is to make AI as

transparent as possible. Furthermore, a goal to make

AI that can explain the reasoning behind its own

conclusions and if not be as transparent as possible

for humans to interpret and understand. For humans,

we understand each other due to the same mental

model and social context we innately possess and

grew with. XAI aims to endow intelligent systems

with the communicative style through that shared

model and context for users to understand with

minimal comprehension barrier.

Neuro-symbolic AI is one approach that can

achieve reasoning with generality and yet with

interpretability, this approach couples the opaqueness

of old-school rule-based symbolic AI with the

obscure internal workings of neural networks.

Symbols are established to give fundamental meaning

to individual neurones in neural networks.

Compositionality through hierarchical attribution is a

symbolic method to represent concepts.

The techniques towards concept learning have

been explored in machine learning but the approaches

often employ ensemble methods – using multiple AI

algorithms to accomplish concept learning. Existing

literature in concept-learning consists of deep neural

networks coupled with natural language processing,

taking visual and question-answer pair as inputs in

order to learn concepts in a joint visual-linguistic

space. This technique is known as grounded learning,

using a joint representational space for both visual

and linguistic compositions. The shared space further

benefits as the semantic interface for users and

interactors to understand the reasonings behind AI’s

conclusions. Some of the outcomes on this line of

research have introduced Visual Concept-

Metaconcept Learner as VCML (Han, Mao, Gan,

Tenenbaum, & Wu, 2019), Neuro-symbolic Concept

Learner as NS-CL (Mao, Gan, Kohli, Tenenbaum, &

Wu, 2019) and Neuro-symbolic Visual Question

Answering as NS-VQA (Yi et al., 2018).

Further research in the area of concept

compositionality and semantic representation that

utilise spiking neural networks is the Semantic

Pointer Architecture: Unified Network (SPAUN)

(Stewart, Choo, & Eliasmith, 2012). SPAUN is

among one of the most accurate cognitive models on

the spiking neurone framework. The primary

component of SPAUN is the Semantic Pointer

Architecture (SPA) (Blouw, Solodkin, Thagard, &

Eliasmith, 2016) which have demonstrated

compositionality and symbolic induction.

Compositionality allows items to form associations

making composites, in symbolic terms, making up

concepts. Symbolic induction in this case is the

predictive process based on sequences of temporally

presented visual inputs, to be precise, the temporal

sequences of activities in the symbolic space in

correlation to the visual inputs.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

1170

For advanced concept learners; a certain set of

inputs patterns could influence and affect other set of

patterns regardless of sensory and perception

modalities. Associations can be formed from data of

disparate sources such as sights, sound and touch – in

computing this can extend to sensors beyond common

sensory modalities of biology. Symbolically a set of

description-patterns can be regarded as individual

concepts by compositionality and symbolic

induction. Conveniently concepts could consist of

description-patterns fragmented across different

modalities in a unified symbolic space.

For this position paper, we propose a structurally

unconstrained concept learner that can learn

dynamically from heterogeneous data streams,

Neuro-Symbolic Spiking Concept Learner (NS-

SCL).

3 SPIKE CONCEPT LEARNER

In artificial neural networks (ANN) and deep learning

(DL) frameworks, neural networks are substantially

non-spiking and during the training phases the models

undergo global weight changes. Spike-based

networks can be feasibly unique in this respect as

learning could involve making localised changes only

between relevant neurones. These localised changes

are mediated and determined by spike activities and

is known to play essential roles in learning as

observed in biology. Several learning rules have been

postulated due to the variety of learning dynamics

observed between neurones from various regions of

our central nervous system. We will cover the

learning rules further on in this section.

Figure 1: NS-SCL overlapping aspects.

Our proposed concept learner (NS-SCL) is based

on the spiking neural framework, but we

deterministically and dynamically structure the

network in a neuro-symbolic way based on incoming

spike-encoded data. In this work, we will exploit the

localised learning of spike-based networks and its

temporal properties to learn neuro-symbolic

constructs in an unsupervised way. We investigate

further into synaptic enhancements as derived in

neuroscience to mean functional forms of plasticity –

how synapse (connections) between neurones adapts

during neural processing and learning. We have

devised a learning mechanism for NS-SCL to achieve

experiential learning. The learning mechanism is

inspired by functional plasticity observed in the

central nervous system with regards to changes to

synapses (connections). The novelty lies in how we

incorporate many synaptic enhancements profiles,

this is achieved through synapse manifolds briefly

covered in Section 3.3.

Furthermore, we investigate the creation and

formation of synapse between neurones based on the

synaptic enhancement conditions as a form of

structural plasticity. New synapse indicates a new

association between neurones, in the symbolic space

we can form associations between items resulting in

the composition of higher symbolic constructs. The

conditions for when synaptic association should form

depending on spikes (of data). Two neurones spiking

may be spuriously correlated yet we form the synapse

but treating such synaptic association as a latent

connection with no functional effect to the network.

Latent synapse will cease to exist after a given period

if not subjected to further stimulation. However, if

further stimulation occurs the latent connections will

persist to exists and it should have a functional effect

in the network. Persistent stimulation indicates a

deterministically coordinated activity, it is then

reasonable to treat the synapse as overt since the

frequency of firings would satisfy some synaptic

enhancement conditions as it could be part of a

genuine association. We will cover briefly the

mechanism of structural plasticity within a neuro-

symbolic space in Section 3.4.

Figure 2: Essential algorithms of NS-SCL.

The essential algorithms in Figure 2 are required

to realise spike-based concept learning. The

functional plasticity aspect of the network is very

much like ANNs and DL with connections between

neurones strengthening or weakening based on the

training data given. The difference is our adoption of

Neurosymbolic Spike Concept Learner towards Neuromorphic General Intelligence

1171

Figure 3: NS-SCL framework.

the spiking neurone paradigm and so our approach to

functional plasticity would differ, an approach being

unique to the spike-based paradigm. Structural

plasticity is our approach to introduce a symbolic

space with spiking neural networks, this symbolic

space allows us to examine and make sense of the

network – this represents an unconstrained ontology-

like space that holds relationships and concepts.

Introducing a neural-symbolic space with live

learning mechanisms as concept leaners opens the

possibility of a self-learning NGI agent that can learn

through experience (with structural and functional

plasticity).

We will employ NS-SCL for concept learning in

an IoT space, as such space is rich with data of various

forms, with varying breadth and lengths. Here, NS-

SCL will generate concepts constituting patterns

present in data across several sensor streams. Figure

3 illustrates the framework of NS-SCL. At process A,

data from sensor streams are encoded into spike-

trains. At process B, we form new neurones and

synapse in symbolic space, where associations are

non-existent – will be based on spike timings in spike-

train evaluation space. At process C, we apply our

manifold algorithm to process spikes in evaluation

space – reinforcing associations where relevant in

symbolic space. In summary, the framework is a real-

time learning model with characteristics of functional

and structural neuroplasticity.

3.1 Learning with Neuroplasticity

In ANNs and DL, the core algorithm for learning is

founded on functional plasticity. Functional plasticity

refers to the changes made to weightings of

connections between neurones as training takes place.

We will not go into further details about functional

plasticity regarding ANNs and DL since the direction

of this paper is towards spike-based functional

plasticity.

Functional plasticity is derived from synaptic

enhancements in neuroscience. Synaptic

enhancements are the changes made to the

neurotransmitter release probability as observed at

the synapse between neurones. Effectively, higher

probability indicates more influence a neurone has

through such synapse in causing another neurone to

fire. Furthermore, short-lived synaptic enhancements

have been classified as paired-pulse facilitation,

synaptic augmentation and post-tetanic potentiation

(Regehr, 2012). The variation between these

classifications are the magnitude and duration for

which the synapse can influence subsequent

neurones, these duration ranges from milliseconds to

minutes. For our concept learner, NS-SCL, we

propose a novel mechanism for learning to cater all

scopes of synaptic influence (facilitation,

augmentation and potentiation) into one unified

neuro-symbolic spiking model. The method in which

we achieve all this is through a technique we call

Temporal Scope Synapse Manifolds (TSSM), further

covered in Section 3.3.

In biology, spiking activities and synaptic

enhancements functions on the 100

millisecond

timescales. With the compute performance of current

general-purpose computers, simulating spiking

activities at natural speeds is unachievable. Hence, it

is the core motivation of Neuromorphic Computing.

Nevertheless, we can exploit the mechanisms by

simulation altering the duration of spikes and

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

1172

simulation speed to a larger scale in order to

demonstrate the feasibility and potential of our

algorithm.

Structural plasticity is another form of

neuroplasticity regarding the changes to the structure

of the network. It has been observed that neural

structures are continuously growing and rearranging.

Dendrites are parts of neurones that allows

connections from/to other neurones. Dendritic spines

have been observed to appear and disappear

depending on their relevance. Spines can last for

months to few days and less (Trachtenberg et al.,

2002). It has been revealed that dendritic spines allow

for the formation of new synapse and is considered to

play a role in learning. The dendritic spine evidently

implies neural networks not only adapts by changes

to connections but by establishing new connections.

In ANNs and DL, the structure of the network is

often defined at the beginning during initialisation –

number of layers, neurones and connection

configurations. The structure of networks in ANNs

and DL do not change once initialised and remains

constant, but the weightings are subjected to

alteration during training. In spiked-based networks,

it is plausible to allow structural change since in

biology this phenomenon happens continuously and

frequently. The structural plasticity method for NS-

SCL will be briefly covered in Section 3.4.

3.2 Functional Plasticity Fundamentals

There is well-known postulate by Donald Hebb in

1949 regarding the activities of neurones during

learning (Hebb, 1949). Hebbian learning, as it is well-

known as, have led to significant advancement in

machine learning for the past decades - specifically

the area of neural networks across its generations. The

very most recent technological advancement in the

past decade that had emerged from the long-credited

postulate have led to the sophisticated deep learning

algorithms.

Hebb's rule is a learning rule regarding the

activities neurones exhibit during learning,

specifically, by persistent stimulation of a neurone

results in the rise in synaptic efficacy (influence on

subsequent neurones). With higher synaptic efficacy,

the more influence the pre-synaptic neurone has on

the post-synaptic neurones. Following the date of

Hebb's postulate, the field of neuroscience have made

further progress through new observations in the

biological process of neurones and spike activities.

Later observations have found that the timing of

neurones firing is a critical component in the process

of learning.

Synaptic plasticity is the observed process that

demonstrates that synaptic efficacy only rises when

connected neurones fires within a very short time-

window. Spike-Time-Dependant-Plasticity (STDP) is

the biological process by which neurones changes the

synaptic efficacy exclusively depending on the

timings of spikes between neurones, the inter-spike

interval (Shrestha, Ahmed, Wang, & Qiu, 2017).

Spike-Rate-Dependant-Plasticity (SRDP) is another

extension by which the persistent number of spikes

leads to more pronounced adjustments to synaptic

efficacy (He et al., 2014). SRDP uses spike averages

to temporally sum the potentials for synaptic

enhancements.

For NS-SCL, we will investigate how STDP can

be used with our discrete-time spiking neural model

for synapse formation. Spiking neural networks in

Neuromorphics embraces the learning rules adapting

with local changes to achieve synaptic plasticity (Liu

& Yue, 2019; Moraitis et al., 2017; Shrestha et al.,

2017). The localised nature of the adaptation

conveniently allows us to extend the structure of the

network without affecting the entire network

functionally. This kind of dynamic structural

characteristics is perhaps not so different to our own

central nervous systems.

3.3 Functional Plasticity with

Temporal Scope Synapse Manifolds

In order to cater for all scope of synaptic influence we

can assume having one connection between neurones

but functionally we can compute the connection with

many different synaptic enhancement profiles –

facilitation, augmentation and potentiation.

Figure 4: Temporal Scope Synapse Manifold (TSSM).

Figure 4 illustrates four synaptic enhancements (SE)

each would have different profiles – varying in

duration and magnitude of influence. All

enhancement profile shares the same vector-direction

component between neurones. Since each SE

functions at different durations, we have imposed

temporal boundaries – for spikes satisfying within,

only the corresponding synapse would be made the

subject to adaptation.

Increasing the base speed at which the network

operate could yield abnormally different result with

the same SE profile. For temporal manifolds, we can

Neurosymbolic Spike Concept Learner towards Neuromorphic General Intelligence

1173

introduce and establish our own parameters that do

not necessarily correspond to natural synaptic

enhancement properties as observed in neuroscience.

We can change to find ideal parameters optimal for

certain base network speeds. We consider defining

SE parameters as the optimisation aspect of our

algorithm.

Table 1: A crude temporal upscaling of SE profiles.

Profile

Mag. Dur. Condition

Facilitation 0.8 1 s 0 < s < 1s

Augmentation 0.2 10 s 1s < s < 10s

Potentiation 0.05 5 mins 10s < s < 5mins

Long-Term P. 0.05 1 hr 5 min < s < 1hr

Table 1 is an example of such SE profiles for

synapse manifolds. The magnitude represents the

weight of influence synapse have on post-neurone at

each spike simulation tick. The duration is how long

the influence will last on post-neurone. The condition

is what must be satisfied for the influence to take

place. Our algorithm will not treat the SE profile

parameters as definitive constants but as modifiable

parameters to fine-tune the behaviour of our spike-

symbolic network. We assume that different

applications for our algorithm will benefit from

different set of SE profiles.

3.4 Structural Plasticity for

Neurosymbolic

This aspect of NS-SCL is to allow the generation of

new neurones and synapses. We will form new

synapse between neurones when they fire satisfying

the manifold rules, and if there are no existing

synaptic connections between them. In the neuro-

symbolic space, new neurones and synapse

effectively forms new symbolic representation of an

item. Manifold rules and unconstrained structure

results in a deterministic yet dynamic behaviour as

unsupervised learning.

Initially the generated synapse will be treated as

latent synapse, causing no real functional effects in

neuro-symbolic network but subjected to functional

plasticity. Latent properties are applied to individual

SE profile. Latent synapse is our mechanism to

structurally regulate the neuro-symbolic space. We

will give generated synapses a probationary period. If

no subsequent changes are made to the synapse

within a set period, we will discard the SE profile

associated as we can conclude that the spike activities

leading to their generations were spuriously

grounded.

Latent synapse could become more progressively

enhanced through our plasticity learning mechanics;

they can be promoted to an overt state. In an overt

state, synapses are effectively active in the network.

Progressive synaptic enhancement can only occur

with well-coordinated spikes across spike-trains;

therefore, we can conclude that the synapse is well-

grounded yet not necessarily known for what cause.

However, the cause can be traced to by observing

neurones that fire depending on the input spike-trains

(data) and would fire somewhat deterministically.

NS-SCL requires each neurone to atomically

represent a concept/constructs in the neuro-symbolic

space. When a neurone fires it would indicate that a

learned-pattern or concept is relevant in present

moment which can be considered as short-term recall.

Two unrelated neurones – having no direct

connections – can form synapse by the temporal

manifold conditions. Thus, NS-SCL can form

complex hierarchical structures of which is

determined by temporal activity.

4 DISCUSSION

Concept learners such as VSCL, NS-CL and NS-

VQA are examples on some of the work aiming to

couple different types of learning spaces. These

approaches have demonstrated learning can be done

on joint visual-semantic space. NS-SCL is our

approach where the joint space is universal. We use

spike-based neural framework which allows us to

temporally encode any information into a shared

universal symbolic space – allowing for information

such as those originating from visual sources,

semantic sources, auditory sources, etc. Hence, NS-

SCL should be able to relate information from one

sensory mode to other sensory modes. This approach

is inspired by how our central nervous system handles

information from various sources.

Major constraints of our approach of concept

learning relates to compute and memory resources.

Since new synapse and neurones can be formed

dynamically, it would require a machine with

considerable processing capabilities and storage

volumes to handle large NS-SCL networks. On

general-purpose computers with the Von Neumann

architecture, the processor would need to process the

number of dynamically formed neurones in under

milliseconds along with synapse which could be

much greater in number. This is impracticable, there

is a point at which exceeding a certain number of

neurones in the NS-SCL neuro-symbolic space will

render the whole algorithm incomputable. We

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

1174

proposed upscaling the algorithm’s timing-related

mechanisms and slowing the spiking simulation

speed to avoid such scenarios allowing the processor

time to compute. Though, reducing the spiking

simulation speed allows for the system to function

with heavy loads, the feasibility depends on the use

case of the algorithm. For IoT use case, we can adapt

sensors to function slower to better accommodate the

sensor data in numbers considering the limitations of

NS-SCL algorithm on the Von Neumann computing

paradigm.

Neuromorphic Computing is a broad field and

requires contribution from many different disciplines.

The motivation of Neuromorphic Computing is to

allow for extreme parallel processing of neurones at

grand scales. Developing algorithms for NC can also

inform the design requirements for neuromorphic

processors. Adapting NS-SCL for Neuromorphic

platforms is the ideal solution as it would eliminate

the Von Neumann compute and memory constraints

that impedes neural processing. NS-SCL requires

dynamical creation of neurones and synapse. In

neuromorphic hardware we require a reserved pool of

unused neurones that can be utilised spontaneously at

runtime in addition forming latent synapse.

Further algorithmic developments should be

made in neuromorphic computing as it has the

potential to influence future developments of

neuromorphic hardware. Future improvements,

regarding concept learning on such platform, could

further reach a level of sophistication where spike-

based concept learners exhibit a degree of general

intelligence functioning in real-time. There have also

been emerging concerns as to the level of

sophistication AI could reach on the intelligence

spectrum. A valid proposition of maintaining AI is to

contain the general forms of AI within isolated

computing mediums like Neuromorphics. Thus, it is

plausible to define a specific branch of artificial

general intelligence that emphasises the

neuromorphic approaches – where intelligence is

coupled to hardware. We identify this specific branch

as Neuromorphic General Intelligence, NGI.

REFERENCES

Barredo Arrieta, A., Díaz-Rodríguez, N., Del Ser, J.,

Bennetot, A., Tabik, S., Barbado, A., … Herrera, F.

(2020). Explainable Artificial Intelligence (XAI):

Concepts, taxonomies, opportunities and challenges

toward responsible AI. Information Fusion, 58

(October 2019), 82–115.

https://doi.org/10.1016/j.inffus.2019.12.012

Barry, B., & Riordan, M. O. (2015). Always-on Vision

Processing Unit for Mobile Applications. IEEE

Computer Society, 56–66.

Blouw, P., Solodkin, E., Thagard, P., & Eliasmith, C.

(2016). Concepts as Semantic Pointers: A Framework

and Computational Model. Cognitive Science, 40(5),

1128–1162. https://doi.org/10.1111/cogs.12265

Boahen, K. (2017). A Neuromorph’s Prospectus. (April

2017).

Bruckner, D., Zeilinger, H., & Dietrich, D. (2012, May).

Cognitive automation-survey of novel artificial general

intelligence methods for the automation of human

technical environments. IEEE Transactions on

Industrial Informatics, Vol. 8, pp. 206–215.

https://doi.org/10.1109/TII.2011.2176741

Davies, M., Srinivasa, N., Lin, T.-H., Chinya, G., Cao, Y.,

Choday, H., … Wang, H. (2018). Loihi: A

Neuromorphic Manycore Processor with On-Chip

Learning. Retrieved from www.computer.org/micro

Debole, M. V., Taba, B., Amir, A., Akopyan, F.,

Andreopoulos, A., Risk, W. P., … Modha, D. S. (2019).

TrueNorth: Accelerating From Zero to 64 Million

Neurons in 10 Years. Computer, 52(5), 20–29.

https://doi.org/10.1109/MC.2019.2903009

Fjelland, R. (2020). Why general artificial intelligence will

not be realized. Humanities and Social Sciences

Communications, 7(1), 1–9.

https://doi.org/10.1057/s41599-020-0494-4

Han, C., Mao, J., Gan, C., Tenenbaum, J. B., & Wu, J.

(2019). Visual Concept-Metaconcept Learning.

(NeurIPS).

He, W., Huang, K., Ning, N., Ramanathan, K., Li, G., Jiang,

Y., … Pei, J. (2014). Enabling an integrated rate-

temporal learning scheme on memristor. Scientific

Reports, 4, 1–6. https://doi.org/10.1038/srep04755

Hebb, D. O. (1949). The Organization of Behaviour. New

York: Wiley & Sons.

Lammie, C., Hamilton, T. J., Van Schaik, A., & Azghadi,

M. R. (2019). Efficient FPGA Implementations of Pair

and Triplet-Based STDP for Neuromorphic

Architectures. IEEE Transactions on Circuits and

Systems I: Regular Papers, 66(4), 1558–1570.

https://doi.org/10.1109/TCSI.2018.2881753

Liu, D., & Yue, S. (2019). Event-driven continuous STDP

learning with deep structure for visual pattern

recognition. IEEE Transactions on Cybernetics, 49(4),

1377–1390.

https://doi.org/10.1109/TCYB.2018.2801476

Mao, J., Gan, C., Kohli, P., Tenenbaum, J. B., & Wu, J.

(2019). The neuro-symbolic concept learner:

Interpreting scenes, words, and sentences from natural

supervision. 7th International Conference on Learning

Representations, ICLR 2019, 1–28.

Moraitis, T., Sebastian, A., Boybat, I., Le Gallo, M., Tuma,

T., & Eleftheriou, E. (2017). Fatiguing STDP: Learning

from spike-timing codes in the presence of rate codes.

Proceedings of the International Joint Conference on

Neural Networks, 2017-May, 1823–1830.

https://doi.org/10.1109/IJCNN.2017.7966072

Neurosymbolic Spike Concept Learner towards Neuromorphic General Intelligence

1175

Painkras, E., Plana, L. A., Garside, J., Temple, S., Galluppi,

F., Patterson, C., … Furber, S. B. (2013). SpiNNaker:

A 1-W 18-core system-on-chip for massively-parallel

neural network simulation. IEEE Journal of Solid-State

Circuits, 48(8), 1943–1953.

https://doi.org/10.1109/JSSC.2013.2259038

Perez-Peña, F., Cifredo-Chacon, M. A., & Quiros-

Olozabal, A. (2020). Digital neuromorphic real-time

platform. Neurocomputing, 371, 91–99.

https://doi.org/10.1016/j.neucom.2019.09.004

Regehr, W. G. (2012). Short-term presynaptic plasticity.

Cold Spring Harbor Perspectives in Biology, 4(7), 1–

19. https://doi.org/10.1101/cshperspect.a005702

Rivas-Gomez, S., Pena, A. J., Moloney, D., Laure, E., &

Markidis, S. (2018). Exploring the vision processing

unit as co-processor for inference. Proceedings - 2018

IEEE 32nd International Parallel and Distributed

Processing Symposium Workshops, IPDPSW 2018,

589–598. https://doi.org/10.1109/IPDPSW.2018.00098

Rosado-Muñoz, A., Bataller-Mompeán, M., & Guerrero-

Martínez, J. (2012). FPGA implementation of Spiking

Neural Networks. In IFAC Proceedings Volumes

(IFAC-PapersOnline) (Vol. 45).

https://doi.org/10.3182/20120403-3-DE-3010.00074

Sengupta, J., Kubendran, R., Neftci, E., & Andreou, A.

(2020). High-Speed, Real-Time, Spike-Based Object

Tracking and Path Prediction on Google Edge TPU.

Proceedings - 2020 IEEE International Conference on

Artificial Intelligence Circuits and Systems, AICAS

2020, 134–135.

https://doi.org/10.1109/AICAS48895.2020.9073867

Shrestha, A., Ahmed, K., Wang, Y., & Qiu, Q. (2017).

Stable spike-timing dependent plasticity rule for

multilayer unsupervised and supervised learning.

Proceedings of the International Joint Conference on

Neural Networks, 2017-May, 1999–2006.

https://doi.org/10.1109/IJCNN.2017.7966096

Stewart, T. C., Choo, F.-X., & Eliasmith, C. (2012). Spaun:

A Perception-Cognition-Action Model Using Spiking

Neurons. Proceedings of the 34th Annual Meeting of

the Cognitive Science Society CogSci 2012, 1018–

1023. Retrieved from

http://palm.mindmodeling.org/cogsci2012/papers/018

4/paper0184.pdf

Trachtenberg, J. T., Chen, B. E., Knott, G. W., Feng, G.,

Sanes, J. R., Welker, E., & Svoboda, K. (2002). Long-

term in vivo imaging of experience-dependent synaptic

plasticity in adult cortex. Nature, 420(6917), 788–794.

https://doi.org/10.1038/nature01273

Yi, K., Torralba, A., Wu, J., Kohli, P., Gan, C., &

Tenenbaum, J. B. (2018). Neural-symbolic VQA:

Disentangling reasoning from vision and language

understanding. Advances in Neural Information

Processing Systems, 2018-Decem

(NeurIPS), 1031–

1042.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

1176