Bettina_K_Peter_Seminar_NeuromorphicComputing.pdf

College of Engineering Poonjar
Neuromorphic Computing for Interactive
Robotics:
A Systematic Review
Bettina K Peter (PJR20CS005)
October 8, 2023
Guided by : Anchal J Vattakunnel(Professor in CSE)
Bettina K Peter (PJR20CS005) Neuromorphic Computing for Interactive Robotics: A Systematic Review 1

Index
1 Introduction
2 Spiking Neural Network
Learning in SNN
SNN Architectures
SNN Models
SP Models
3 Neuromorphic Chips/WWWWW/simulators and Frameworks
4 SNNs In Robotics Applications
Signal Acquisition and Processing
Pattern Recognition
Speech Recognition
Motor Control
Cognition and Learning
5 Future Direction
6 Conclusion
7 References

Introduction College of Engineering Poonjar
Introduction
The adult human brain runs continuously, whether awake or sleeping, on
only about 12 watts of power whereas a hypothetical clock-based
computer running a “human-scale” brain simulation requires
approximately 12 Gigawatt of power.

Human Brain
”A single neuron in the brain is an incredibly complex machine that even
today we don’t understand.
A single ’neuron’ in a neural network is an incredibly simple mathematical
function that captures a minuscule fraction of the complexity of a
biological neuron.”
-AndrewNg

Neuromorphic Computing
To acquire more autonomy and operate in the real world, robots should:
1) perceive their environments in real-time
2) process sparse information with efficiency
3) behave under changing conditions
4) acquire self-learning ability

Definition
Neuromorphic computing (also known as brain-inspired computing) is a
multidisciplinary research paradigm that investigates large-scale processing
systems that support natural neuronal computations through spike-driven
communication. Advantages of neuromorphic computing over traditional
approaches are:
energy
efficiency
execution speed
robustness

Neural Spike

Spiking Neural Network College of Engineering Poonjar
Generations Of NNs

Application Of NNs
1st Gen NNs:
Basic Computations
Early AI Research
2nd Gen NNs:
Detection and Identification
Pattern Recognition
Function Approximation
Natural Language Processing
Image and Speech Processing
3rd Gen NNs:
Image Classification and
Object Recognition
Computer Vision and Robotics
Sensor Data Processing
Intelligent Systems with
Manipulators
Robotic Applications
Detection and Recognition
Tasks
Numerical Data Processing

Working Of SNN
It aims to mimic the behavior of natural neurons found in the
brain.
Spikes are the fundamental currency of neural computation and
communication.
Neurons release neurotransmitters during a spike, allowing them to
communicate with other neurons.
These networks take spike trains as input and produce spike trains as
output.
When a neuron’s state variable, surpasses a predefined threshold, it
sends a spike to each post-synaptic neuron.

Spiking Neural Network Learning in SNN College of Engineering Poonjar
History Of Learning
Synaptic plasticity was first proposed as a mechanism for learning and
memory by D.O Hebb in 1949.
It is the change that occurs at synapses, the junctions between neurons
that allow them to communicate.

Spiking Neural Network Learning in SNN College of Engineering Poonjar
Types Of Learning
Unsupervised Learning
Spike-timing-dependent
plasticity (STDP)
Long Term Potentiation
(LTP) - Strengthening
synapses
Long-term Depression
(LTD) - Weakening synapses
Supervised Learning
Back-propagation
SpikeProp, FreqProp and
ReSuMe
Reinforced Learning
Actor-Critic Model
STDP Modulation

Spiking Neural Network SNN Architectures College of Engineering Poonjar
SNN Architectures
Figure: Feed Forward Neural
Network(FFNN)
Figure: Hybrid Neural Network
Structures
Figure: Recurrent Neural
Network(RNN)
Figure: Hybrid Neural Network
Architectures

Spiking Neural Network SNN Models College of Engineering Poonjar
SNN Models
1)HODGKIN-HUXLEY NEURON MODEL
First bio-inspired neural model.
Developed by Sir Alan Hodgkin and Sir Andrew Huxley in 1952.
This model describes how neurons initiate and propagate action
potentials.
The key equation relates membrane potential (Vm) to the current
(Iext) and membrane capacitance (Cm).
The iconic current (Iion) is composed of three main components:
sodium current, potassium current, and small leakage current.

SNN Models
2)LEAKY INTEGRATED-AND-FIRE (LIF) NEURON MODEL
Commonly used due to its simplicity.
Neurons have a membrane potential (Vm), and a threshold (Vth) for
generating spikes.
When Vm exceeds the threshold, an action potential (spike) is
generated.
Voltage is reset after a spike.
Described by a simple differential equation:
Intrinsic properties: Ve, Rm, τm; External current: Im; Vm.

SNN Models
3)IZHIKEVICH NEURON MODEL
Combines biological plausibility of Hodgkin-Huxley with
computational efficiency of integrate-and-fire (LIF) models.
Described by a 2-D system of differential equations:
Key variables:Vm, U (membrane recovery variable), Im.
Dimensionless variables a and b adjust the neuron behavior.

Spiking Neural Network SP Models College of Engineering Poonjar
Synaptic Plasticity Models
Rate-Based Models:
Common and use average
spike count over time.
Magnitude of synaptic
plasticity is determined by the
rate of pre- and post-synaptic
firing.
Backpropagation.
Experiments involving
robot-based sensory and
motor systems.
Spike-Based Models:
Consider the precise timing of
individual spikes.
Synaptic weight changes
based on the relative timing of
pre- and post-synaptic spikes.
STDP
Neural learning mechanisms in
robotics, both in simulations
and real environments.

Neuromorphic Chips/WWWWW/simulators and Frameworks College of Engineering Poonjar
Neuromorphic Hardware
Neuromorphic hardware plays a critical role in enabling robots to
perform cognitive tasks.
Recent advancements in both neuroscience and the chip industry
have led to the development of new neuromorphic hardware for
simulating SNNs.
The following tables provide a brief overview of the key elements in
this field:

Table 1
Neuromorphic Chips: Specialized hardware designed for simulating
SNNs.
E.g: TrueNorth, Braindrop, SyNAPSE, FACETS, NeuroMem, NM500, and
SynSense.

Table 2
Simulators: Software tools used to simulate SNN behavior.
E.g: Genesis, SpikeFun, Mayavi and CSIM.

Table 3
Frameworks: Software platforms that provide tools and libraries for
developing and running SNN models.
Specific frameworks are not mentioned in the provided text.

Table 4
Robots: Robots used in socially interactive neurorobotics are discussed.
E.g: Pepper and Nao

Table 4

SNNs In Robotics Applications Signal Acquisition and Processing College of Engineering Poonjar
SNN Applications
A.Signal Acquisition and Processing in Robot Applications:
Intelligent sensors for autonomous planning and adaptation.
Combining vision and non-vision sensors for faster recognition.
Rast et al.: iCub robot and SpiNNaker system for object
identification.
Enhanced networks with behaviorally relevant STDP.
Multi-sensory processing for environment perception.
Convergence-zone model applied to multi-modal systems.
Al-Qaderi and Rad’s multi-modal system outperforms uni-modal on
Pioneer 3DX.

SNNs In Robotics Applications Pattern Recognition College of Engineering Poonjar
SNN Applications
B.Pattern Recognition Applications in Robotics:
Crucial for real-world robots and human-robot interactions.
Previous approaches needed strict supervision during training.
Fanello et al.: Improved robot vision with limited constraints.
Mansouri-Benssassi and Ye: Bio-inspired SNNs with unsupervised
STDP for facial expression recognition, outperforming other methods.
Al-Qaderi and Rad: Multi-modal system for efficient facial recognition
in real-world scenarios.
Cyr and Theriault: Robots learn and adapt behavior with different
rewarding rules.

SNNs In Robotics Applications Pattern Recognition College of Engineering Poonjar
SNN Applications
Real-time human motion recognition with SNNs and instructive
learning in three phases.
a)Teaching phase: learner visualize the target action.
b)Turn-taking phase: learner extract the nonverbal information.
c)Trial phase: learner confirms the target action

SNNs In Robotics Applications Speech Recognition College of Engineering Poonjar
SNN Applications
C.Speech Recognition Applications:
Vital for robots to understand spoken commands, especially in noisy
environments.
Embodied embedded cognition approach by Davila-Chacon, Liu, and
Wermter improves ASR systems in noisy settings.
Speech recognition can be used for neural layer training through
auditory grouping.
Spiking cerebellar model used for vestibulo-ocular reflex (VOR) tasks
with iCub robot.
Model employs adaptive real-time control loops and STDP for
generating eye motor commands during VOR tasks.

SNNs In Robotics Applications Speech Recognition College of Engineering Poonjar
SNN Applications
Sound Source Localization (SSL) used for accurate sound direction
detection and robot orientation.
SNN calculates sound signal angle, tested on iCub robot and
Soundman.
ASR with SSL significantly boosts speech recognition accuracy in
humanoid robots.
Figure: SSL Architecture

SNNs In Robotics Applications Motor Control College of Engineering Poonjar
SNN Applications
D.Motor Control Applications:
SNNs for pointing motions adaptable to various robots.
CPG and SNNs efficiently control robot gaits.
DMSNN guides robot arm motion by monitoring neuron firing.
SNNs enable soft-grasping and sEMG-based reflexes in robotic hands.
Multi-modal interface for remote robot control, aiding disabled and
elderly users.
Cognitive platforms based on SNNs model user-environment
interactions.

SNNs In Robotics Applications Motor Control College of Engineering Poonjar
SNN Applications
iQSA method integrates BCIs and SNNs for Hexapod robot
locomotion.
BCIs and evolving SNNs enable prosthetic hand control.
Applications extend to medical science, disability assistance, and
teleoperation systems.
Figure: Block diagram of prosthetic control through BCI

SNNs In Robotics Applications Cognition and Learning College of Engineering Poonjar
SNN Applications
E.Cognition and Learning Applications:
SNNs used for spatial memory, enhancing navigation and exploration.
It improves SLAM in social robots, creating accurate environmental
maps.
SNNs implement associative and working memory, enhancing
cognitive abilities.
Episodic memory enables robots to perform versatile cognitive tasks
using cognitive maps.
Emotion integrated into robot decision-making, considering intrinsic
motivation.

Future Direction College of Engineering Poonjar
Future Direction
* Humanoid Robots
* Neuromorphic Chips
* Hardware Optimization
* Personalization
* Generalized Framework
* Ethics

Conclusion College of Engineering Poonjar
Conclusion
Neuromorphic computing offers promise for human-like robotic
intelligence in terms of computation, speed, and energy efficiency.
Many articles focused on improving specific applications but lacked
full integration for more capable systems.
A universal training method and conversion mechanisms are essential
for advancing human-like social interactions in neurorobots.
Interdisciplinary collaborations, particularly between roboticists and
neuroscientists, are crucial for further development.

References College of Engineering Poonjar
References
[1] LiMin Fu. 1994. Neural Networks in Computer Intelligence.
McGraw-Hill, Inc., USA.
[2] Kumar Satish. 2004. Neural Networks : A Classroom Approach. New
Delhi: Tata McGraw-Hill.
[3] Hebb, D. O. 1949. The organization of behavior; a neuropsychological
theory. Wiley.

Bettina_K_Peter_Seminar_NeuromorphicComputing.pdf

More Related Content

Similar to Bettina_K_Peter_Seminar_NeuromorphicComputing.pdf (20)

Recently uploaded (20)

Bettina_K_Peter_Seminar_NeuromorphicComputing.pdf