Content uploaded by Arvind Bhagwath
Author content
All content in this area was uploaded by Arvind Bhagwath on Sep 25, 2024
Content may be subject to copyright.
Sustainable AI
Contributed by Arvind Bhagwath – Founder of microML.IN (hps://microml.in)
The rush to implement Arcial Intelligence (AI) in industries
has disrupted long-established sustainability plans. Few
research rms esmated that AI accelerators will
consume 2318 TWh of global power in next four years,
reaching a 1.5% share of global electricity consumpon and
raises big concern on Arcial Intelligence (AI) data centers
contribung to global warming.
For the fact, our Human brain uses just 20 was of energy,
however as per current data available in internet by research
rms, Human created AI systems currently consume a lot of
power and altogether, data centers use more electricity than
most of countries.
To reduce this high energy use and carbon emissions, using
more ecient smaller Machine learning processes and
change in current AI System architecture is key for developing
sustainable AI Systems. We can reduce the size of many AI
models by an order of magnitude if we design the AI
architecture just the way like how our Human brain works.
For the fact, memories in Human Brain aren’t stored in just
one part of the brain. Dierent types of sensory
data(memory) are stored across dierent, interconnected
brain regions. These data stored in networks may be used for
predicve modelling, decision control that normally come out
as emoonal output from Brain. As a part of evoluon, self-
learning resulng from experience occurs and each me data
would be stored within networks, which later is used to
derive conclusions from these complex set of informaon
stored in neural networks.
Quantum Memory for Sustainable AI
Systems?
Memory is the faculty of the brain by which data or
informaon is encoded, stored, and retrieved when needed.
Brain is made of neurons and a single neuron is connected to
many other neurons and the total number of neurons and
connecons is extensive. Human brain contains almost 100
billion neurons. Each neuron has on average 7,000 synapc
connecons to other neurons. It has been esmated that the
brain of a three-year-old child has about 1015 synapses.
Back in 1998, a physicist Bruce Kane from Australia had
suggested that phosphorus atoms embedded in silicon would
be the ideal way to store and manipulate quantum
informaon that later put forward a design for
a Kane quantum computer. The Kane computer is based on
an array of individual phosphorus donor atoms embedded in
a pure silicon lace. Both the nuclear spins of the donors and
the spins of the donor electrons parcipate in the
computaon.
With current deeper understanding on behaviour
of Phosphorus atoms and Quantum mechanism, possibility
of quantum processing with nuclear spins could be one of
main reason on how brain stores and reads memory needs
to be explored. As we are aware, when a memory is created,
informaon ows from the cortex, the part of the brain rich
in nerve cells, to the hippocampus, the central switching
point for memories in the brain. The informaon ows in the
opposite direcon when we retrieve a memory. Brain
memories underlie our ability to learn, to tell stories, even to
recognize each other’s. Researchers have been able to trace
memory down to the structural and even the molecular level
in recent years and this should be possible only if they show
Quantum behaviour and hence role of Phosphorus atoms
play an important role. Phosphorus has only one naturally
occurring stable isotope, ³¹P, which like ¹H has spin = ½ and
thus two discrete energy states which becomes ideal
candidate to store memory at molecular level something
represented in image below:
In the nervous system, a synapse is a structure that permits
a neuron to pass an electrical or chemical signal to another
neuron or to the target eector cell. Synapses are essenal to
the transmission of nervous impulses from one neuron to
another. As shown in image above, communicaon at
chemical synapses requires release of neurotransmiers.
When the presynapc membrane is depolarized, voltage-
gated Ca2+ channels open and allow Ca2+ to enter the cell.
The calcium entry causes synapc vesicles to fuse
(contract) and thus helps in releasing neurotransmier
molecules into the synapc cle. The neurotransmier
diuses across the synapc cle and binds to ligand-gated ion
channels in the postsynapc membrane, resulng in a
localized depolarizaon or hyperpolarizaon of the
postsynapc neuron.
Neurons are specialized to pass signals to individual target
cells, and synapses are the means by which they do so. Both
the presynapc and postsynapc sites contain extensive
arrays of molecular machinery that link the two membranes
together and carry out the signalling process/transfer of data
with release of neurotransmiers. With availability
of Phosphorus atoms in Neurotransmier molecules,
Phosphate bonds might exhibit Quantum behaviour which
Brain might be using to store memory and this mechanism
could be almost like how qubit operaons are possible by
manipulang the electric elds in Quantum computers
using Phosphorus atoms.
Using lesser variants in current Machine
Learning models?
Current centralized Machine learning LLM models generates
lot of variants which might actually not be required at all for
normal use and training one such single machine learning
(ML) model can equal the combined carbon emissions of
mulple cars. We need to focus on sustainable AI systems on
reducing the negave environmental impacts associated with
current design of arcial intelligence technologies such as
creang independent Machine Learning models on
standalone microchips instead of depending on centralized
models for every request. microML.IN focuses in developing
several such standalone modules on low cost and low
powered microcontrollers for actual AI requirement of any
such industries.
Next step, we can further reduce the size of unnecessary
dataset used to train a model to minimize energy use and
carbon emissions involved leveraging the idea on how Human
brain works.
Memory in Human Brain is oen understood as an
informaonal processing system with explicit and implicit
funconing that is made up of a sensory memory, short-term
memory and long-term memory.
Sensory memory holds informaon, derived from the senses,
less than one second aer an item is perceived. Short-term
memory is also known as working memory. Short-term
memory allows recall for a period of several seconds to a
minute without rehearsal. For example, recalling a ten-digit -
telephone number is a kind of Short-term memory. In Long-
term memory, declarave, or explicit memory is
the conscious storage and non-declarave or implicit memory
is the unconscious storage and recollecon of informaon.
Implicit memory uses past experiences to remember things
without thinking about them and it is used to perform
certain procedural tasks, such as driving, riding a bike, playing
a musical instrument etc.
Long-term memory can store much larger quanes of
informaon for potenally unlimited duraon (somemes a
whole life span). Its capacity is immeasurable. It is maintained
by more stable and permanent changes in neural connecons
widely spread throughout the brain.
We can leverage this kind of storage mechanism instead of
pushing all the data to one centralized data center that
currently consumes huge energy. The rise of generave AI
and surging GPU shipments is causing data centers to scale to
thousands of accelerators, shiing the emphasis to power as
a mission-crical problem to solve.
Role of Phosphorus in Genec or DNA Memory?
Several genes, proteins and enzymes have been extensively
researched for their associaon with Long-term memory.
Long-term memory, unlike short-term memory, is dependent
upon the synthesis of new proteins.
For instance, Rats when exposed to an intense learning event
may retain a life-long memory of the event. When such an
exposure was experimentally applied, more than 5,000
dierently methylated DNA regions appeared in
the hippocampus neuronal genome of the rats at one and at
24 hours aer training. These alteraons in methylaon
paern occurred at many genes that were down-regulated,
oen due to the formaon of new 5-methylcytosine sites in
CpG rich regions of the genome.
Genome size is the total number of the DNA base pairs in one
copy of a haploid genome. Genome size varies widely across
species. Invertebrates have small genomes; this is also
correlated to a small number of transposable elements. In
humans, the nuclear genome comprises approximately 3.2
billion nucleodes of DNA, divided into 24 linear molecules,
the shortest 50 000 000 nucleodes in length and the longest
260 000 000 nucleodes, each contained in a dierent
chromosome.
Unlike Neural memory, DNA or Genec memory is straight
forward. Genec memory is a memory present at birth that
exists in the absence of sensory experience and is
incorporated into the genome over long spans of me.
Phosphorus here too plays a very important role in building
up this memory. As we are aware, DNA and RNA structures
here are connected by phosphorous bonds and this DNA
memory is like the ash drive that is used to store biological
data almost like computers, which use 0 and 1. DNA uses A,
C, G and U/T (the ‘nucleodes’, ‘nucleosides’ or ‘bases’).
This memory got developed with several million years of
evoluon based on events recorded in DNA which is just
stored in code of four chemical bases adenine (A), guanine
(G), cytosine ©, and thymine (T).
A, C, G and U/T (the ‘nucleodes’, ‘nucleosides’ or ‘bases’)
helps to build dierent Amino Acids. Proteins are molecules
made of amino acids and a gene is a segment of a DNA
molecule that contains the instrucons needed to make
a unique protein.
Example: If you cut the leg o a salamander, it will grow back.
Similarly in roses, they have thorns to protect them from
being eaten by animals. Extra protecon funcons like Thorns
for Roses or Regeneraon is triggered by DNA from earlier
events of damage. This is how memory from these phosphate
bonds helps for survival and growth.
The idea of DNA data storage is not merely theorecal.
Sciensts have mastered to decode DNA memory that is
helping for several reverse engineering projects like
development of life saving vaccines. In vaccine producon
process, DNA code has to be uploaded to a DNA printer
which then converts the bytes to actual DNA molecules.
CRISPR DNA-eding technology is available now to record
images of a human hand into the genome. It was showcased
by Church’s group at Harvard for E. coli, which were read out
with higher than 90 percent accuracy. Researchers at the
University of Washington and Microso Research have
developed a fully automated system for wring, storing and
reading data encoded in DNA. The genec material DNA has
garnered considerable interest as a medium for digital
informaon storage because its density and durability.
Conclusion:
To meet the current demand of AI systems, businesses are
invesng billions of dollars to grow GPU capacity which is
denitely not sustainable and not green. The rush to
implement AI in business has disrupted long-established
sustainability plans with growing cost of electricity and
challenges facing energy grids. In my view, understanding
why human brain uses just 20 wa of power and performs
beer than AI systems should give new denion on how
memory is processed, dened and stored. Quantum way to
store the data and processing needs to be explored. It posits
that quantum-mechanical phenomena, such as entanglement
and superposion of Phosphorus atoms may play an
important role in storing and processing the data in much
sustainable way.