Neuromorphic Computing for Smart Cities

Full text

Non-Trivial Fellowship 1
Neuromorphic Computing for Smart Cities
Matthew Chen, Iris Litiu, Christina Lu
Abstract
Excessive energy consumption poses a significant threat to the environment.
Recognizing that urban areas account for approximately three-quarters of global energy usage
(Ge, Friedrich, & Vigna, 2020), our project specifically targets reducing energy consumption in
cities.
We demonstrate the feasibility and quantitative impact of neuromorphic computing
(NC) for energy-efficient Internet of Things (IoT) applications in urban environments. We find
that this approach can significantly lower the energy consumption of infrastructures and
processes in these areas while achieving greater performance and accuracy. As a proof of
concept, we optimize a network of smart traffic lights using a spiking neural network (SNN), to
adaptively change their signaling algorithms within a traffic network based on the amount of
congestion at each intersection. We outline how event-based cameras can be effectively coupled
with asynchronous neuromorphic processors, and how processors can subsequently relay
information within a road network. In turn, this reduces emissions from vehicles by cutting
down on the time they spend in traffic. We also propose the architecture of a broad network of
neuromorphic devices, to extrapolate our traffic reduction proof of concept to other smart city
applications. The applications of NC to reduce the energy consumption of processes in urban
environments are endless.
Adaptive algorithms are only the beginning; in this paper, we outline numerous other
applications of NC to improve the capabilities of IoT within smart cities. We have chosen to
focus on traffic algorithm optimization since it demonstrates core functionality improvements in
sensor-processor and processor-network communications and it is easy to generate fake data
with which to test our prototypes. The goal of this project isn’t to resolve the excessive energy
consumption issue in just eight weeks, but to lay the groundwork for future innovations that
eventually lead to sustainable, energy-efficient smart cities.

Non-Trivial Fellowship 2
Problem Analysis: The Growing Environmental
Impact of Cities
Global Warming and Environmental Degradation
The summer of 2023 was recorded to be Earth's hottest on record, with temperatures
rising 0.23 degrees Celsius (0.41 degrees Fahrenheit) above the previous highs (US, 2023). A
result of skyrocketing temperatures is the rapid melting of Arctic sea ice, which has decreased in
volume by 30% in just the last 50 years (NASA, 2023). This loss of ice not only contributes to
rising sea levels but also disrupts the habitats of numerous species. In fact, the world is currently
losing species at a rate over a hundred times greater than at any other time in recorded human
history (Martin, 2019). Looking forward, the predicted future environmental changes are even
more drastic. Notably, one million species are at risk of extinction within the next few decades
due to the effects of climate change—a possibility that will impact ecosystems, food security, and
even human health (Martin, 2019).
Taking immediate action to address global warming is necessary. The consequences of
inaction are predicted to be severe and far-reaching, threatening not only wildlife and
ecosystems but also the well-being and survival of the human race.
City Excessive Energy Consumption
Energy consumption is the leading cause of climate change, posing a grave threat to the
planet and its inhabitants. It's the largest source of human-caused greenhouse gas emissions,
accounting for 75% of the global total (Ge, Friedrich, & Vigna, 2020). Moreover, these
emissions, primarily carbon dioxide, are the “main cause of future global warming”
(Intergovernmental Panel on Climate Change, 2021). Since energy consumption is a significant
contributor to greenhouse gas emissions, reducing our use of it is essential for mitigating the
impacts of global warming. In other words, if we don’t significantly cut back on our energy
usage, we run the risk of exacerbating climate change and facing the severe consequences that
follow.
Some may ask why humanity can’t just entirely rely on renewable sources. Completely
shifting to renewable energy isn't feasible due to resource limitations. For example, wind power,
one of the most widely used renewable resources today, requires at least 60 acres of land (0.24
square kilometers) per megawatt of energy produced (Landmark Dividend LLC, 2014). To put
this into perspective, the United States consumed over 3.5 billion megawatts (3.5 million
megawatts) in 2020 alone (GlobalData Plc, 2017). Although wind energy is just one example, it
highlights the broader issue: the amount of land, physical resources and time required to meet
current energy demands with only renewable sources is near impossible. Thus, to effectively
mitigate environmental degradation, we must reduce our energy consumption.

Non-Trivial Fellowship 3
Urban areas play a significant role in this issue, consuming 78% of the world's energy
(Nations, 2023). This share is unlikely to decrease due to growing city populations. In fact, by
2050, 70% of the world's population are predicted to live in cities, up from the 56% in 2020
(Institute for Economics & Peace, 2022). As cities become increasingly populated, their already
sky-high energy consumption levels will likewise rise. Therefore, we should prioritize reducing
energy consumption in cities and make them the focal point of our sustainability efforts.
Neuromorphic devices will be significant in this effort due to numerous software and hardware
optimizations that will be explained and verified in the following sections of the paper.
The Role of Smart Cities and the Internet of Things
One increasingly popular city-building paradigm is the so-called “smart city.” It attempts
to increase the efficiency of actions that a city must perform, like energy generation and traffic
management, via data collected from sensors scattered around the city. By making
data-informed choices about the operation of each city, smart cities allow urban areas to
optimize their behavior to best meet the needs of their residents. For example, a city may use
devices embedded within the power grid to track power demand, dynamically storing and
releasing energy based on usage throughout the day in order to alleviate network load (Baibakov
& Nikolski, 2024).
Critics of smart cities often point to how the implementation and operation of these
sensors themselves can be a waste of energy, which, in a fully integrated city, may be a
significant portion of its power draw. They also challenge the efficiency and accuracy of these
systems, asking about the significance of any optimizations that could be made. Essentially,
critics ask city planners to consider whether these devices either waste energy and resources or
generate real savings in the long run.
Similarly, the Internet of Things (IoT) refers to the grid of technology and devices that
permeates our lives and communicates with each other: smartphones, Internet-enabled
appliances, medical devices, and any other networked device. As the core technology in smart
cities, the IoT provides significant advantages in efficiency and communication. Nevertheless,
like smart cities, it consumes substantial amounts of energy in its operation, thereby
exacerbating excessive energy usage.
If implemented correctly, an expanding IoT and future smart cities could reduce our
environmental impact, benefiting both the planet and humanity. However, they also carry the
risk of further accelerating the deadly consequences of climate change.

Non-Trivial Fellowship 4
Solution Overview
Neuromorphic computing (NC) is a promising, entirely novel, computing framework that
differs dramatically on the hardware level from conventional “von Neumann” chips. Whereas
traditional devices operate on hardcoded, system-wide clocks and have discrete memory,
processing, and transportation modules, NC imitates the layout of the human brain (Schuman et
al., 2024).
The biological brain, made of neurons, is an analog, rather than digital, computing
device; power and voltage are handled not by 0s and 1s, but by a continuous spectrum of neuron
activity. Like neural networks, each neuron has up to 10,000 connections through its synapses
with neighboring neurons and sends them outputs in the form of spikes of electricity. Each
neuron also has its own memory and processing power coupled at its core, (the soma).
NC also opens the door to new kinds of neural networks, based on physical hardware
rather than digital neurons. The operational principle of Spiking Neural Networks (SNNs) on
these chips—where neurons communicate by firing spikes only when necessary—results in
sparse and efficient communication that drastically cuts down energy usage compared to
continuous data transmissions in conventional systems (IBM, 2024). This spike-based
communication is important in reducing latency and power usage, especially beneficial for
real-time applications like autonomous navigation systems and image recognition.
Neuromorphic computing has several key optimizations over traditional devices. First,
they derive significant energy savings over conventional devices: this is due to the low power
consumption of spikes, as well as a reduction in data bus energy costs (as data is stored locally
with processing). They are also able to process event-based information more accurately (such
as image recognition algorithms) by discarding redundant data, making them more versatile and
quicker than traditional computers in handling live information (Baibakov & Nikolski, 2024). By
processing data on-device without the need to transmit information back to a central data
center, these systems minimize latency, enhance privacy, and reduce bandwidth and energy
consumption, and are especially important in scenarios where communications networks are
not accessible (eg. deep sea exploration).

Non-Trivial Fellowship 5
We believe that because of these characteristics of NC, it has great potential to replace a
variety of technologies in IoT and smart cities by decreasing their power usage and increasing
their efficiency.
The following table summarizes some of our proposed key use cases of neuromorphic
IoT devices, enabling infrastructure to not only respond to changes in the environment but also
holistically predict and adapt to future conditions via energy-efficient means. Note that most of
these technologies (minus the NC component) are already prevalent in many world cities, so our
proposal aims to integrate them with neuromorphic hardware and increase their usage once
initial pilot programs are successful.
Technology
Use case
Why neuromorphic hardware?
Sensors - strain gauge,
accelerometer, fiber optic
sensors, acoustic
emission sensors
Monitor building
structural health for
preventative
maintenance and
immediate response
Easily adjustable disturbance
threshold for monitoring large
changes without continuous energy
use. Detection of deep anomalies such
as unusual vibrations, shifts, or stress
patterns in real time.
Sensors - temperature,
humidity, corrosion,
wind, barometric
pressure, air quality,
seismometers
Integrate environmental
measurements into
automated predictive
modeling and analysis
Distributed monitoring that improves
anomaly detection over time, enabling
more rapid and accurate responses.
Fault tolerance coupled with
algorithms that temporally correlate
weather events leads to a better
understanding of climate phenomena.
Geospatial monitoring -
multispectral cameras,
radar sensors, LiDAR,
thermal sensors,
magnetometers,
gravimeters
Visualization of
geographical data
Can process large and complex
datasets in realtime and extrapolate
patterns. Durability allows for the
monitoring of vast and varied
geospatial areas.
Materials with adaptive
properties (more durable
and fewer maintenance
needs)
Self-healing concrete,
power lines that reroute
around outages to not
take out the entire grid,
etc.
Energy-efficient processing of
complex data streams. For example,
neuromorphic devices can learn from
the structural behavior of the concrete
over time, improving the accuracy of
detecting when and where healing is
needed. Additionally, these devices
can detect where outages occur and
decide where to cut off power to
prevent cascading outages. Processors
can be decentralized, enabling grid
scalability and reduced decision
latency.

Non-Trivial Fellowship 6
Renewable energy
systems
Solar and wind energy
farms
Analyze sensor data to optimize
energy production, predict issues that
may lead to downtime, and do
adaptive load balancing to prevent
grid instability.
Robotics
Devices to perform
inspections and
maintenance tasks in
challenging
environments
Enhance understanding of complex
environments, real-time management
of robotics actuators, resilience under
extreme conditions, and coordination
of multiple robots working on the
same task.
Wireless communication
networks
Rapid coordination
among first responders
and the public during
disasters
Impulse radio communication
between neuromorphic devices allows
for precise timing and positioning and
is ideal for low-power and
high-precision applications
The following cycle summarizes how a network of neuromorphic devices communicate and
update their local models based on the state of the global network. Once devices are deployed
throughout the smart city, they use sensor data to run an initial training cycle. The weights of
the SNNs after the initial iteration are transmitted via impulse radio to a central processor that
runs analysis on the state of the entire network. Global network weights are used to inform
incremental adjustments of local weights until individual devices optimally respond to
environmental or infrastructural feedback.

Non-Trivial Fellowship 7
This diagram depicts how an individual neuromorphic device processes input data before
relaying it to a central processor. First, sensor input is encoded as spikes with magnitudes
proportional to the amount of change in the sensor state from a baseline state (e.g., pressure
detected by a pressure sensor relative to 0 psi, intensity of light entering a camera’s pixel relative
to no photons). Then, the spikes are inputted into an SNN pretrained on simulation data.
Depending on how the network is trained, it can decode the spikes into a numerical
representation for the current system state.
Results: Evaluating a SNN in an IoT Context
Model Overview
There exists a diverse and varied array of applications of neuromorphic IoT in smart
cities. As concrete proof of concept, we have demonstrated an adaptive traffic light algorithm
updated through a network of SNNs simulating the behavior and performance of neuromorphic
processors. Our model shows how these processors communicate with each other by adapting
the weights of their local algorithms based on the traffic distribution within the entire network.
It leverages traffic data to dynamically update the city’s traffic light signaling algorithm, creating
a more efficient transport grid. The end result of this sample model is an increase in the
efficiency and speed of city traffic, leading to time, energy, and emissions savings. We argue that
the usage of neuromorphic networks in this context generalizes to many other relevant smart
city applications, such as infrastructure health monitoring and environmental monitoring, by
demonstrating how these devices can communicate and learn from each other.
We began creating our model using a few different libraries to determine which library
provided the most accurate NC simulation while remaining feasible to implement within our

Non-Trivial Fellowship 8
timeframe. We ultimately decided to pursue Torch NN, as it had the optimal tradeoff between
accuracy and ease of integration with our adaptive traffic light algorithm.
Model
Pros
Cons
Brian2
●Easy setup and intuitive
syntax
●Descriptive training and
output visualizations
●Dynamic code generation
●No support for deployment on
neuromorphic hardware
●Not scalable to larger or more
complicated models
Lava DL
●Simulation package
supports direct deployment
onto neuromorphic devices
like the Loihi 2 chip
●Scalable to large models
and simulations
●Supports biologically
realistic neural networks
●Development package is
low-level and requires custom
SNN implementation using
package-specific Lava Process
Models
●Limited development
resources and documentation
●Relatively new framework, so
some features are not yet
complete
Torch NN
●Comprehensive
documentation
●Part of an extensively
developed framework with
pretrained models and
libraries for data loading
●Flexibility in developing
new layers and loss
functions for networks with
different use cases
●Performance tuning is
required for larger scale
applications
●Lack of native support for
spiking neurons (required
custom implementation)
Both an artificial neural network (ANN) and spiking neural network (SNN) model were
implemented to compare and contrast the efficiencies and accuracies of neuromorphic
computing architectures and of traditional von Neumann architectures when applied to IoT
devices. SNNs are inherently more energy efficient than ANNs, due to their event-driven
processing, where they only fire when they receive sufficient input.
Both models scale green light duration by traffic density to prevent congestion while
allowing the greatest amount of traffic to flow through their intersections. Both use a mean
squared error loss function and an Adam optimizer to minimize the loss function. Traffic data is
randomly generated using PyTorch’s random tensor generation function. Assuming only
two-way intersections (with roads crossing each other) for simplicity, our model outputs the
duration of time the traffic lights will be green at the NW versus SE intersections.

Non-Trivial Fellowship 9
Traffic
model
ANN
SNN
Network
●Feed-forward network with
three fully connected dense
layers
●ReLU activation function
●Layer normalization between
layers
●Dropout layer to help prevent
data overfitting
●Feedforward network with two
Leaky Integrate and Fire (LIF)
neurons and two fully connected
linear layers
●To better simulate biological
neurons, LIF neurons
accumulate input over time and
emit spikes when a certain
threshold is reached, instead of
outputting continuous values
●Uses a data loader to shuffle data
so model results are more
generalizable
Training
Cycle
1. Reset accumulated gradients
2. Get the next slice of traffic
data and update the weights
of all network layers
3. Calculate target durations
4. Calculate loss via the loss
function between the current
network output and the target
5. Update the model’s
parameters with the optimizer
based on the loss function
1. Using the same target
calculations as the ANN, the data
and target are passed into the
model to generate spike activity
and membrane potentials
2. Average spiking activity over
time is passed into the loss
function along with the target,
from which the optimizer
updates model parameters
Road Network Implementation
We implemented a virtual road network simulation to assess how well individual models
can adjust their weights according to the state of the entire network. We generated an
n-by-m-dimensional road grid, with an arbitrary number of intersections of interest randomly
chosen throughout the grid. Traffic data, representing car density, is randomly generated at each
timestamp. Individual traffic light model weights are updated every timestamp. Additionally,
the tensors representing green light durations in the NW and the SE direction at each
intersection are also updated after individual model weights are updated. To reduce congestion,
for every NW and SE street with 2 or more intersections, the relative green light durations are
scaled proportionally to the greatest difference among the green light durations in the NW and
SE directions. For example, if there are three intersections along a street, and the greatest time
delta between the green light in the NW and SE directions is 5 seconds, with the green light in
the NW direction being longer, all the other green lights in the NW direction will have 5 seconds

Non-Trivial Fellowship 10
added to their duration. Then, the durations of the NW and SE directions are renormalized to fit
within the span of a light cycle. The updated tensors are then fed back into the original network,
and this cycle of updating individual models, feeding the individual model output into the larger
feedback algorithm, and feeding the global algorithm outputs back into the individual models is
repeated over each training cycle. To visualize the network workflow:

Non-Trivial Fellowship 11
We then quantitatively evaluated both the energy performance and the accuracy of each model.
ANN SNN
Fig 1: Comparison of ANN (left) and SNN (right) performance in traffic control scenario. Two
models were trained to optimize traffic flow in a hypothetical, arbitrarily defined traffic
grid. Despite allowing the ANN to run for 100,000 training epochs, where the SNN was
only allowed 500, the SNN showed comparable, if not better performance, with both loss
functions approaching 0.001, or 0.1%. If allowed to continue the run, the ANN faces
significant diminishing returns, while the SNN is able to push accuracy lower by more
than two degrees of magnitude (range of 10^-5).
Fig 2: Power draw of the ANN and CNN network. Power collection was obtained through the
ZeusMonitor GPU library to calculate the power consumption of the ANN, approximately 330
Joules. To overcome our lack of access to neuromorphic hardware, we found the number of
synaptic events that occurred (51,119,232), and then applied chip-specific power consumption
data to the events; for the Intel Loihi chip (23.6 pJ per synaptic event, (Covi et al., 2021)), as a
typical device, it would take less than 1 J to run our entire SNN network.

Non-Trivial Fellowship 12
We suggest that these results are indicative of the performance and energy savings that would be
derived from applying SNNs and neuromorphic hardware broadly in the IoT.
Significance
Projects like KI4LSA (2022) conducted by the Fraunhofer Institute for Optronics used
high-resolution cameras and radar sensors to calculate the number of vehicles and wait times at
intersections. They leveraged this data through deep reinforcement learning to train a neural
network to calculate the optimum switching behavior for the traffic lights. In simulation, their
algorithm (summarized below) reduced traffic congestion by 10-15%.
The traffic light phase transition algorithm LI4LSA proposes can be integrated with our road
network algorithm, which generates traffic light phases for simple two-street intersections, once
we generalize our implementation to intersections at a junction of an arbitrary number of roads.
The scalability issues the LI4LSA study faced can then be mitigated through our proposal of a
lightweight network of neuromorphic processors. These processors will be coupled with a
camera and an Impulse Radio (IR) receiver and transmitter to communicate with a central
processor that updates individual intersection SNNs using the global network weight results. IR
transmissions are ideal for low-power and high-precision communications and integrate well
with the sparse data transmission of the neuromorphic processors. Therefore, this
implementation prevents much of the communication latency introduced by the LI4LSA,
including the algorithm’s computational complexity and the communication overhead for
exchanging information between intersections.
Solution Integration: Neuromorphic IoT for Smart
Cities
Next Steps in Integration and Adoption
Each device in the network will include an event-based sensor and a spike-based asynchronous
processor implementing Integrate-and-Fire neurons, for effective information storage and
recall. The event-based sensors don’t process redundant data, leading to reduced power
consumption (Caccavella et al., 2023). For example, in image processing, NC devices will only be
triggered by changes in brightness; traffic light optimization algorithms will only recalculate
network weights when cameras detect significant changes in pixel values that are correlated with
changes in congestion.

Non-Trivial Fellowship 13
Example Workflow
Visual data is streamed as the coordinates of each pixel, the timestamp, and the polarity of the
event (whether brightness has increased or decreased). To process this data, event-based
sensors are combined with asynchronous processors that share the same computational
principles as Von Neumann architectures with their clocked circuits (e.g., an event-based vision
sensor and an asynchronous neuromorphic processor).
1. Neuromorphic device receives inputs from the embedded sensor featuring a pixel array
or off-chip data.
2. Events are queued and sent to processing cores in FIFO order.
3. Each processor core executes a per-event computation sequence of convolution →
Integrate-and-Fire (IF) spiking neuron →sum pooling. To synchronize results across
neuromorphic processors, the processors can be coupled with external hardware with a
FPGA to timestamp outgoing events and do format conversions.
To train the SNNs before deployment onto neuromorphic devices, back-propagation can be used
during simulation and the trained model can be deployed onto neuromorphic hardware for the
inference phase. Another approach would be to directly train algorithms on the neuromorphic
devices through local weight update methods via local plasticity (changing the activation levels
of synapses). This can develop adaptive spiking models capable of learning from new data at the
edge, since neuromorphic chips often lack support for global learning algorithms, like
backpropagation, which rely on differentiable, continuous activation functions.
Addressing Limitations of NCs
Limitation: Strategies such as batch normalization, skip connections, dropout, and pooling are
widely utilized to enhance stability and mitigate overfitting, but are designed for
continuous-valued networks.
Solution: These issues may be addressed by developing analog techniques for SNNs:
●Batch normalization can be done via spike frequency adaptation to maintain a stable
firing range, adaptive thresholding (raising the firing threshold if too much activity), and
homeostatic plasticity to self-regulate activity by adjusting synaptic weights or neuron
parameters in response to the neuron's average firing rate, membrane potential scaling,
and layer-wise normalization.
●Skip connections allow spikes to bypass certain layers, which can facilitate learning in
deeper networks by ensuring that spikes can still propagate through the network even if
intermediate neurons are not firing.
●Dropout in SNNs translates to randomly silencing neurons or synapses during training.
This can be implemented on Loihi by adjusting the probability of spike generation or by
dynamically modulating synaptic weights to achieve a similar effect.
●Pooling can be adapted to operate on the temporal patterns of spikes. For example,
spike-timing-based pooling could be used to select the earliest spikes or to count spikes

Non-Trivial Fellowship 14
within a time window. This reduces the temporal resolution of the spike trains, similar to
how spatial pooling reduces resolution in traditional networks.
Limitation: The divergent behavior of the SNN in simulation and on NC hardware can
significantly affect the model’s precision.
Solution: The need to discretize the stream of input events in temporal bins during simulation
leads to a reduction in the number of spikes produced by the neurons compared to the on-chip
inference where each event is processed independently. If the time interval for event aggregation
used in the simulation is not sufficiently small, the on-chip inference might generate a
significantly higher number of spikes than the simulation for the same input, leading to stalling
due to limited processing core bandwidth. To adjust for this, the number of spikes generated
when the membrane potential exceeds the threshold should be directly proportional to the
membrane potential value.
Limitation: The model’s training results may not be generalizable beyond the data it was
trained on.
Solution: For the model to be able to generalize its results to other datasets, for regression
tasks such as object detection, the output spike events need to be converted into continuous
values via a linear layer within the SNN, with layer normalization to eliminate batch
dependence.
Theory of Change
The following diagram indicates how to resolve various aspects of our problem by
developing solutions that target their root causes. Each subproblem under the status quo section
has one or more primary causes, which have at least one primary solution and further
specifications for those solutions.

Non-Trivial Fellowship 15
Tractability and Neglectedness
Tractability
As with the adoption of any new technology, the rapid development and integration of
NC in IoT devices may face challenges with cost, research capability, and production. To begin
adoption, we foresee needing government subsidies to encourage adoption up to a critical
threshold. However, as the technology matures, we should see rapid increases in both the
affordability and the performance of NCs, making them increasingly commercially viable; as
precedence, similar events occurred with the adoption of EVs and government-promoted tax
benefits to both manufacturers and consumers (EPA, 2024).
Neglectedness
While there is growing interest in energy efficiency and sustainability, current efforts are
typically focused on improving specific technologies and not rethinking energy-consuming
systems as a whole. Our solution of using neuromorphic computing and IoT aims to address
energy consumption at a systemic level, completely revamping the way that the devices around
us work: a novel solution.

Non-Trivial Fellowship 16
Next Steps
The following steps must be taken to create a functional framework for integrating
neuromorphic IoT into smart cities.
1. Combine our traffic algorithm with the LI4LSA algorithm and generalize it to traffic
grids of real cities. Also integrate the algorithm with pedestrian traffic lights.
2. Transfer algorithm implementation to Lava or an equivalent library to be deployed
directly to compatible neuromorphic hardware.
3. Use the same neuromorphic device communication scheme to extrapolate our process
and algorithm to other applications, such as infrastructure health monitoring and
tracking environmental conditions like air quality, humidity, water contamination, etc.
4. Draft public policy to enable the security of the data being transferred by neuromorphic
devices. Propose an integration plan for phasing in the new architecture in major cities
that doesn’t disrupt currently implemented systems via ensuring communication and
interoperability between neuromorphic systems and existing digital infrastructure
networks. Ensure compatibility by setting international standards and protocols for
neuromorphic communication.

Non-Trivial Fellowship 17
References
Baibakov, K., & Nikolski, D. (2024). Chapter 4: Neuromorphic computation and its impact on
AI-driven systems. In D. Nikolski (Ed.), Advances in neuromorphic computing (pp.
87-103). Springer. https://link.springer.com/chapter/10.1007/978-3-031-65549-4_4
Caccavella, C., Paredes-Vallés, F., Cannici, M., & Khacef, L. (2023). Low-power event-based face
detection with asynchronous neuromorphic hardware. arXiv.
https://doi.org/10.48550/arXiv.2312.14261
Chen, J., Skatchkovsky, N., & Simeone, O. (2022). Neuromorphic integrated sensing and
communications. IEEE Wireless Communications Letters.
https://doi.org/10.48550/arXiv.2209.11891
Climate Change in Data: The Physical Science Basis. https://www.ipcc.ch/. Published 2021.
Accessed August 26, 2024.
https://www.ipcc.ch/report/ar6/wg1/resources/climate-change-in-data/
Du, Z., Chen, S., & Wang, Y. (2016). Micro-15: Hardware-software co-design of neuromorphic
processors. Retrieved from
https://courses.cs.washington.edu/courses/cse590g/16wi/du_micro15.pdf
Eshraghian, J. K., Ward, M., Neftci, E., Wang, X., Lenz, G., Dwivedi, G., Bennamoun, M., Jeong,
D. S., & Lu, W. D. (2021). Training spiking neural networks using lessons from deep
learning (arXiv preprint arXiv:2109.12894). https://arxiv.org/abs/2109.12894
Fraunhofer-Gesellschaft. (2022, February). Traffic lights controlled using artificial intelligence.
Fraunhofer-Gesellschaft.
https://www.fraunhofer.de/en/press/research-news/2022/february-2022/traffic-lights-
controlled-using-artificial-intelligence.html
Schuman, C.D., Kulkarni, S.R., Parsa, M. et al. Opportunities for neuromorphic computing
algorithms and applications. Nat Comput Sci 2, 10–19 (2022).
https://doi.org/10.1038/s43588-021-00184-y
Ge M, Friedrich J, Vigna L. 4 Charts Explain Greenhouse Gas Emissions by Countries and
Sectors. World Resources Institute. Published 2020. Accessed August 26, 2024.
https://www.wri.org/insights/4-charts-explain-greenhouse-gas-emissions-countries-an
d-sectors
Martin, Martin. UN Report: Nature’s Dangerous Decline “Unprecedented”; Species Extinction
Rates “Accelerating” - United Nations Sustainable Development. United Nations
Sustainable Development. Published May 6, 2019. Accessed August 26, 2024.

Non-Trivial Fellowship 18
https://www.un.org/sustainabledevelopment/blog/2019/05/nature-decline-unpreceden
ted-report/
Matsuo, R., Elgaradiny, A., & Corradi, F. (2024). Unsupervised classification of spike patterns
with the Loihi Neuromorphic Processor. Electronics, 13(16), 3203.
https://doi.org/10.3390/electronics13163203
Müller, A., & Sabatelli, M. (2022). Safe and psychologically pleasant traffic signal control with
reinforcement learning using action masking. arXiv. https://arxiv.org/pdf/2206.10122
Royal, M. (2024). Neuromorphic Computing Guide [GitHub repository]. GitHub.
https://github.com/mikeroyal/Neuromorphic-Computing-Guide
The Smart City Journal. (n.d.). Smart traffic light technology controlled using artificial
intelligence. The Smart City Journal.
https://www.thesmartcityjournal.com/en/smart-mobility/smart-traffic-light-technology
-controlled-using-artificial-intelligence
United States Government. (n.d.). Tax Incentives for EVs. EPA.
https://fueleconomy.gov/feg/taxcenter.shtml
NASA. (2023, October 23). Arctic Sea Ice Minimum | NASA Global Climate Change. Climate
Change: Vital Signs of the Planet.
https://climate.nasa.gov/vital-signs/arctic-sea-ice/?intent=121
Wind Turbine Lease Rates – How Valuable is Your Wind Farm Lease? | Landmark Dividend -
Landmark Dividend LLC. (2014, July 10). Landmark Dividend LLC.
https://www.landmarkdividend.com/wind-turbine-lease-rates-2/#:~:text=Quantity%20
of%20Land%20Available,turbines%20and%20other%20supporting%20infrastructure.
GlobalData Plc. (2017). The Power consumption in United States of America (2017 - 2020,
GWh). Globaldata.com.
https://www.globaldata.com/data-insights/power-and-utilities/the-power-consumption
-in-united-states-of-america-1083501/
IBM. (2024, June 27). What is neuromorphic computing?. IBM.
https://www.ibm.com/think/topics/neuromorphic-computing
Covi, E., Donati, E., Liang, X., Kappel, D., Heidari, H., Payvand, M., & Wang, W. (2021).
Adaptive extreme edge computing for wearable devices. Frontiers in Neuroscience,15.
https://doi.org/10.3389/fnins.2021.611300