yeshwanthg
Nov 15, 2024
Graduate / SOP review for masters in data science at CMU [2]
It's interesting that I hadn't always considered pursuing a master's degree, as I wasn't fully aware of its
importance early on. My interest in data science, in fact, came about almost by accident when my mother
casually mentioned it during my eleventh grade. She brought it up after an exciting discussion at her work-
place about emerging fields, which ended with the conclusion that data science was gaining huge attention
and could be the next major leap in computer science. I didn't think about it at the moment, but as I started
college and continued coming across the term, my curiosity increased. During a seminar in my second year,
a speaker brought up the point that data analysis enables us to forecast the results of our actions. I connected
with this idea. I have had to make tough choices throughout my life, some of which had terrible results. The
possibility of predicting the outcome of my choices felt almost inconceivable, knowing the future based on
the steps I take. This question kept nagging me: is this truly possible? Since then, I have actively pursued
research in data science, co-authoring 9 high-impact research papers and contributing to 3 patents. This very
question is the reason I would like to pursue a master's degree in data science, which would allow me to
come closer to answering it.
During my Information Security Analysis and Audit class, I met Dr. Sahaya, whose thoughts on research
and its potential to drive solutions inspired me. I asked her about predicting outcomes through data science,
and she confirmed its feasibility and highlighted the challenges, especially handling vast datasets securely.
This led us to focus on securing federated learning infrastructures for large-scale data collaboration with-
out compromising privacy. I developed a Secure Federated Learning (SFL) system using digital twin and
blockchain technologies to ensure data security and prevent model poisoning which resulted in an accuracy
of 97%. Later published in IEEE Access, our work titled 'Fortifying Federated Learning in IIoT: Leveraging
Blockchain and Digital Twin Innovations for Enhanced Security and Resilience' marked a crucial milestone
in advancing secure predictive models.
Building on my research and growing interest in data science, I secured an opportunity at highperformr.ai,
where I navigated the complexities of applying theoretical knowledge in real-world scenarios. A key chal-
lenge arose when our site traffic surged from 1,000 to 7,000 daily visitors, leading to unsustainable costs
from frequent ChatGPT API calls. This experience required both technical resource management and strate-
gic thinking. I developed a solution by fine-tuning BERT, GPT, and LLAMA models using PEFT and LoRA
techniques, integrating these models into both our free tools and our social CRM app. The app helped users
manage social media accounts by drafting personalized posts through AI. The ChatGPT API's limitations in
handling multiple interactions at once conflicted with our goal of crafting posts tailored to each user's voice.
By building our own model, we generated a signature for users based on their posting history, aligning it with
content from similar but higher-quality creators to produce more refined, personalized outputs. This journey
instilled in me a deeper understanding of balancing technical and operational needs while fostering creativity
in AI applications. The sense of accomplishment I felt as I saw the content being generated and our efforts
bearing fruit planted the seed for my entrepreneurial aspirations. At CMU, I aim to further nurture this vision,
equipping myself with the skills to build innovative solutions and take meaningful strides toward founding
my own startup
After my internship at Highperformr.ai, I saw an opportunity to make a societal impact when I learned
that local government school children near our campus did not know how to operate a computer, as they
revealed during a conversation. With Dr. Sahaya's support, I took action by organizing a computer science
camp at Kumizhi Government School, teaching the basics of computer use. During the camp, one boy stood
out, not for his participation, but for his isolation. Learning that he was autistic, I was emotionally affected
and began researching ways to help children like him, especially those with Savant Syndrome, a condition
where certain autistic individuals possess extraordinary cognitive abilities. I quickly realized a sprawling
gap, while there were methods to identify if a child is autistic, there was no effective way to assess their
potential, particularly by focusing solely on the child's input. This gap was further complicated by a divide
between technological and medical assessments. Motivated to address this, I developed the idea of a multi-
modal cognitive assessment system that integrates facial expression analysis, response initiation time, and
traditional grading techniques. This system, powered by a deep learning model, not only predicts cognitive
scores but also facilitates the detection of Savant Syndrome, resulting in my first patent titled A System and
Method Facilitating Cognitive Assessment in Users. This project not only fulfilled my desire to make a
societal impact but also marked an important milestone in my journey as a researcher.
Building on this, I began questioning why separate models were necessary for different data types. Is
it impossible for a single ML model to learn from different data types? This question served as the basis
for the following research that I conducted under the guidance of Dr. Vatchala. I worked on developing a
model with shared and modality-specific layers to better correlate diverse data while preserving their unique
attributes. This led to the development of a multimodal biometric authentication system integrating facial,
vocal, and signature data, achieving an accuracy of 94.65%. This work, titled "Multi-modal Biometric Au-
thentication: Leveraging Shared Layer Architectures for Enhanced Security", has been submitted to IEEE
Access. I am eager to further this research in the MultiComp Lab at CMU, where the mission of under-
standing the interdependence between human verbal, visual, and vocal behaviors aligns well with my goal
of enhancing multimodal algorithms to accurately integrate distinct data types for improved security and
interaction insights.
CMU's pioneering research in federated learning and secure AI aligns perfectly with my goals in fair
and privacy-preserving machine learning. I am particularly excited about the opportunity to work with
Dr. Virginia Smith at CyLab, where her research on 'Fair Federated Learning via Bounded Group Loss'
closely aligns with my work in Secure Federated Learning (SFL). Her approach to achieving fairness through
bounded group loss mirrors my own efforts to ensure privacy and robustness in federated settings, especially
when managing diverse, non-IID data across devices. CyLab's commitment to advancing security through its
four core themes is incredibly inspiring, and I would be honored to contribute specifically to the "Secure by
Design Systems" theme, exploring how fairness and security can complement each other in federated learn-
ing. This environment would be ideal for bettering my understanding of secure, fairness-oriented models and
refining my skills in applying privacy-preserving techniques to real-world federated applications.
Additionally, I am drawn to the work of Dr. Ameet Talwalkar, whose research on 'Expanding the Reach
of Federated Learning by Reducing Client Resource Requirements' addresses key limitations in federated
learning by making it accessible to a wider range of users. His work on reducing client-side resource de-
mands aligns well with my interest in creating scalable and inclusive federated systems. Dr. Talwalkar's
collaboration with Dr. Smith, especially in their joint work on 'Weight sharing for hyperparameter opti-
mization in federated learning', offers a unique opportunity for me to learn about hyperparameter tuning in
federated learning. I would be thrilled to contribute my expertise in SFL to their collaborative research, ex-
ploring ways to integrate fairness, efficiency, and scalability across federated learning applications. Working
with both Dr. Talwalkar and Dr. Smith at CMU would be an invaluable experience, enabling me to develop
effective and ethical federated systems.
A fundamental question has driven my journey in data science: can we predict the outcomes of our
actions using data? This curiosity has guided my academic and research endeavors, from developing Secure
Federated Learning systems to filing a patent on cognitive assessment for autistic children. Each step has been
crucial to my journey toward answering this question. The Masters in Computational Data Science program
at CMU offers the ideal environment to deepen my understanding and push the boundaries of predictive
modeling. In the long run, I aim to start a company that will put these answers into practice by developing a
solution that helps individuals understand where their lives could lead based on the steps they take. By turning
this vision into reality, I hope to empower people to make more informed decisions about their futures.
It's interesting that I hadn't always considered pursuing a master's degree, as I wasn't fully aware of its
importance early on. My interest in data science, in fact, came about almost by accident when my mother
casually mentioned it during my eleventh grade. She brought it up after an exciting discussion at her work-
place about emerging fields, which ended with the conclusion that data science was gaining huge attention
and could be the next major leap in computer science. I didn't think about it at the moment, but as I started
college and continued coming across the term, my curiosity increased. During a seminar in my second year,
a speaker brought up the point that data analysis enables us to forecast the results of our actions. I connected
with this idea. I have had to make tough choices throughout my life, some of which had terrible results. The
possibility of predicting the outcome of my choices felt almost inconceivable, knowing the future based on
the steps I take. This question kept nagging me: is this truly possible? Since then, I have actively pursued
research in data science, co-authoring 9 high-impact research papers and contributing to 3 patents. This very
question is the reason I would like to pursue a master's degree in data science, which would allow me to
come closer to answering it.
During my Information Security Analysis and Audit class, I met Dr. Sahaya, whose thoughts on research
and its potential to drive solutions inspired me. I asked her about predicting outcomes through data science,
and she confirmed its feasibility and highlighted the challenges, especially handling vast datasets securely.
This led us to focus on securing federated learning infrastructures for large-scale data collaboration with-
out compromising privacy. I developed a Secure Federated Learning (SFL) system using digital twin and
blockchain technologies to ensure data security and prevent model poisoning which resulted in an accuracy
of 97%. Later published in IEEE Access, our work titled 'Fortifying Federated Learning in IIoT: Leveraging
Blockchain and Digital Twin Innovations for Enhanced Security and Resilience' marked a crucial milestone
in advancing secure predictive models.
Building on my research and growing interest in data science, I secured an opportunity at highperformr.ai,
where I navigated the complexities of applying theoretical knowledge in real-world scenarios. A key chal-
lenge arose when our site traffic surged from 1,000 to 7,000 daily visitors, leading to unsustainable costs
from frequent ChatGPT API calls. This experience required both technical resource management and strate-
gic thinking. I developed a solution by fine-tuning BERT, GPT, and LLAMA models using PEFT and LoRA
techniques, integrating these models into both our free tools and our social CRM app. The app helped users
manage social media accounts by drafting personalized posts through AI. The ChatGPT API's limitations in
handling multiple interactions at once conflicted with our goal of crafting posts tailored to each user's voice.
By building our own model, we generated a signature for users based on their posting history, aligning it with
content from similar but higher-quality creators to produce more refined, personalized outputs. This journey
instilled in me a deeper understanding of balancing technical and operational needs while fostering creativity
in AI applications. The sense of accomplishment I felt as I saw the content being generated and our efforts
bearing fruit planted the seed for my entrepreneurial aspirations. At CMU, I aim to further nurture this vision,
equipping myself with the skills to build innovative solutions and take meaningful strides toward founding
my own startup
After my internship at Highperformr.ai, I saw an opportunity to make a societal impact when I learned
that local government school children near our campus did not know how to operate a computer, as they
revealed during a conversation. With Dr. Sahaya's support, I took action by organizing a computer science
camp at Kumizhi Government School, teaching the basics of computer use. During the camp, one boy stood
out, not for his participation, but for his isolation. Learning that he was autistic, I was emotionally affected
and began researching ways to help children like him, especially those with Savant Syndrome, a condition
where certain autistic individuals possess extraordinary cognitive abilities. I quickly realized a sprawling
gap, while there were methods to identify if a child is autistic, there was no effective way to assess their
potential, particularly by focusing solely on the child's input. This gap was further complicated by a divide
between technological and medical assessments. Motivated to address this, I developed the idea of a multi-
modal cognitive assessment system that integrates facial expression analysis, response initiation time, and
traditional grading techniques. This system, powered by a deep learning model, not only predicts cognitive
scores but also facilitates the detection of Savant Syndrome, resulting in my first patent titled A System and
Method Facilitating Cognitive Assessment in Users. This project not only fulfilled my desire to make a
societal impact but also marked an important milestone in my journey as a researcher.
Building on this, I began questioning why separate models were necessary for different data types. Is
it impossible for a single ML model to learn from different data types? This question served as the basis
for the following research that I conducted under the guidance of Dr. Vatchala. I worked on developing a
model with shared and modality-specific layers to better correlate diverse data while preserving their unique
attributes. This led to the development of a multimodal biometric authentication system integrating facial,
vocal, and signature data, achieving an accuracy of 94.65%. This work, titled "Multi-modal Biometric Au-
thentication: Leveraging Shared Layer Architectures for Enhanced Security", has been submitted to IEEE
Access. I am eager to further this research in the MultiComp Lab at CMU, where the mission of under-
standing the interdependence between human verbal, visual, and vocal behaviors aligns well with my goal
of enhancing multimodal algorithms to accurately integrate distinct data types for improved security and
interaction insights.
CMU's pioneering research in federated learning and secure AI aligns perfectly with my goals in fair
and privacy-preserving machine learning. I am particularly excited about the opportunity to work with
Dr. Virginia Smith at CyLab, where her research on 'Fair Federated Learning via Bounded Group Loss'
closely aligns with my work in Secure Federated Learning (SFL). Her approach to achieving fairness through
bounded group loss mirrors my own efforts to ensure privacy and robustness in federated settings, especially
when managing diverse, non-IID data across devices. CyLab's commitment to advancing security through its
four core themes is incredibly inspiring, and I would be honored to contribute specifically to the "Secure by
Design Systems" theme, exploring how fairness and security can complement each other in federated learn-
ing. This environment would be ideal for bettering my understanding of secure, fairness-oriented models and
refining my skills in applying privacy-preserving techniques to real-world federated applications.
Additionally, I am drawn to the work of Dr. Ameet Talwalkar, whose research on 'Expanding the Reach
of Federated Learning by Reducing Client Resource Requirements' addresses key limitations in federated
learning by making it accessible to a wider range of users. His work on reducing client-side resource de-
mands aligns well with my interest in creating scalable and inclusive federated systems. Dr. Talwalkar's
collaboration with Dr. Smith, especially in their joint work on 'Weight sharing for hyperparameter opti-
mization in federated learning', offers a unique opportunity for me to learn about hyperparameter tuning in
federated learning. I would be thrilled to contribute my expertise in SFL to their collaborative research, ex-
ploring ways to integrate fairness, efficiency, and scalability across federated learning applications. Working
with both Dr. Talwalkar and Dr. Smith at CMU would be an invaluable experience, enabling me to develop
effective and ethical federated systems.
A fundamental question has driven my journey in data science: can we predict the outcomes of our
actions using data? This curiosity has guided my academic and research endeavors, from developing Secure
Federated Learning systems to filing a patent on cognitive assessment for autistic children. Each step has been
crucial to my journey toward answering this question. The Masters in Computational Data Science program
at CMU offers the ideal environment to deepen my understanding and push the boundaries of predictive
modeling. In the long run, I aim to start a company that will put these answers into practice by developing a
solution that helps individuals understand where their lives could lead based on the steps they take. By turning
this vision into reality, I hope to empower people to make more informed decisions about their futures.
