Biomimetic Nanoscale Devices and
Architectures for Brain-Inspired Computing
and Artificial Intelligence
Rashmi Jha1 ,Alex Jones1, Sam Wenke1, Eric Herrmann1, Tony
Bailey1, Andrew Rush1, Manish Kumar2
1Department of Electrical Engineering and Computing Systems, University of
Cincinnati, Cincinnati, Ohio
[email protected], 513-556-13612Department of Mechanical, Industrial, and Materials Engineering, University
of Cincinnati, Cincinnati, Ohio
[email protected], 513-556-5311
Motivation
“Why would we want to mimic the brain
when computers can already perform
computations at an incredibly fast speeds?”
vs
Digital Computing Limitations
• Quantity of data recorded
– In general, many businesses already have more data than they
know what to do with.
– Internet of Things will create millions of new devices that will add
significantly more data to process.
• Complexity of data
– Time-series data and high-dimensionality data are computationally
expensive.
• Algorithms to interpret data
– How to we decide what information is useful?
Limitations of Currently Available
Machine Learning: Deep Neural Network
(DNN)
DNN on conventional computing architecture are
compute intensive, power hungry, need a large set
of training data , and are trained to solve just some
specific sets of problems.
Goals: Replication vs. Mimicry
• We don’t need to completely replicate the functionality of the brain,
we already have billions of those.
• Rather, we need can identify the core components of the brain can
learn and identify features, we can create new architectures that
optimize that operation.
Neuron Operation and Action Potential Firing
Synapse
A Mechanism for Learning
C. Zamarreño-Ramos et al, "On spike-timing-dependent-plasticity,
memristive devices, and building a self-learning visual cortex," Frontiers
in Neuroscience, vol. 5, pp. 26, 2011.
Sensory Signal Processing
Temperature,
odor etc.
(Effector Cells)
Central Nervous
System
Sensory Information Encoding
Sensory Neurons in Silicon
Axon-Hillock Circuit, proposed by Prof. Carver Mead,
1980’sIndiveri et. al., Frontiers in Neuroscience, 2011
Synaptic Crossbar Structure
• Ultra Low-power
• Scalable (n2 synapses per 2n neurons)
• High endurance and reliability
Neuron
Synapse
Synapse Design Parameters:
• Connective weight between neurons
• Forward propagation w/o weight change
• Modify weight based on the pre/post spike
patterns
• Prevent Overfitting
Potential Synapse Realizations
C. D. Schuman et al, "A Survey of Neuromorphic Computing and Neural Networks in
Hardware," 2017.
Memristor:
• Stores memory state in its
resistance similar to charge
in a capacitor
• Synaptic Weight =
Resistance
• 4F2 device density
• Operate @ sub-µA currents
& relatively low voltages
Synaptic Memory Device
TiN
SrTiO3
W
-
+
T. J. Bailey and R. Jha, "Characterization of transient redox
dynamics in SrTiO3 synaptic devices," in 2017.
State of the Art in Inferencing IBM’s TrueNorth
• 1 million VLSI neurons
• SRAM synapses
• Weights trained offline and
programmed onto the board
• Real-time inferencing of images
Google’s “Tensor Processing
Unit” (TPU)
• ASIC to optimize tensor
operations.
• 8-bit multiply-and-add on
signed and unsigned integers.
• Voice recognition processing
and inferencing.
Neuromorphic Computing for
Robotic Navigation in Space
Current AI Software on Curiosity:
AEGIS – Target Identifying Algorithm
OASIS - Autonomous Science Framework
Opportunity to embed artificial intelligence directly onto robots to increase the
speed of learning and/or decision making.
Robot Navigation with
Neuromimetic Crossbar Simulation
,Gianluca et. al., IEEE Transactions on Robotics 21.5 (2005): 994-1004.
Robot Navigation Results
1 2 3
4 5 6
o : start | × : target
Conclusions• We established the reasoning to mimic the brain in order to create
new computing paradigms to compensate for limitations in digital
computing.
• The core elements and mechanisms behind learning in the brain
were elaborated on and proven realizations of neuron and synaptic
elements were presented.
• We established potential for having embedded artificial intelligence
for robots based on emerging neuromorphic devices.
• A successful robotic pathfinding application was demonstrated using
unsupervised learning scheme to guide the robot using local
knowledge of obstacles.
• Our approach is projected to be energy-efficient and scalable for
implementation on robotic systems.
• Future work is targeted towards the integration of neuron and
synaptic elements on-chip to achieve unsupervised learning in
hardware.
Acknowledgement• This project is currently supported by National Science
Foundation under CAREER (Award # 1556294).
• We would like to thank Dr. Mark Ritter and his group at IBM TJ Watson Research Center.
• We would like to thank our collaborators Dr. GennadiBersuker (Sematech), Dr. David Gilmer (Sematech), Dr. Prashant Majhi (Intel), Dr. Kevin Leedy (AFRL), Dr. Marc Cahay (U. of Cincinnati), Dr. Ali Minai ( U. of Cincinnati), Dr. Swaroop Ghosh (USF), Dr. Scott Molitor (U.Toledo), Dr. Cory Merkel (AFRL), Dr. Matt Casto (AFRL), Dr. Brian Dupiax (AFRL), Mr. Clare Thiem (ARFL).
Thank You!
Questions and Suggestions?