Ganesh Murugappan

🌉 Dead Week at UC Berkeley 🌉 May has swept through Berkeley, bringing with it the annual tradition known as Dead Week. As the buzz of Telegraph Avenue dims slightly, we undergrads know it’s our cue to hunker down and hit the books one last time before finals. Here are some key points I've been focusing on as we wrap up this dead week! Lesson 1: Structure in Chaos Berkeley’s Dead Week is ironically alive with a flurry of review sessions and last-minute project deadlines. I've found that a structured approach—breaking down study sessions into focused blocks—helps me navigate this intense week. The serene backdrop of the Bay Area, from the calming waters by the marina to the redwoods around campus, provides the perfect study break scenery to clear my mind and reset. Lesson 2: The Strength of Shared Struggles There's something about the communal silence of Mainstacks Library during Dead Week that reminds you you're not alone. Joining forces with other students, we grind out practice exams and Pomodoro sessions until 2 am. It's these moments of collaboration that transform challenging content into shared victories. Lesson 3: Well-being as a Priority Dead Week at Berkeley is not just about academic survival but also about maintaining well-being amidst the madness. I make it a point to keep up with my runs through the Fire Trails, ensuring that my mental and physical health aren’t sidelined during these critical times. Lesson 4: Reflection and Readiness As I review notes and practice problems, I’m also reflecting on how each semester builds more than just technical skills. It hones resilience, deepens patience, and expands my understanding of what I can achieve under pressure. With finals on the horizon, I feel prepared and eager to tackle the challenges ahead. To my peers also in the midst of Dead Week, let’s push through to the finish line with determination and spirit! 💪 What are your Dead Week strategies? Let’s share some wisdom and encouragement as we gear up for finals! 🚀 #UCBerkeley #ComputerScience #DeadWeek #FinalsPreparation #GoBears

39

6 Comments

🌟 𝐃𝐫𝐞𝐚𝐦𝐢𝐧𝐠 𝐁𝐢𝐠, 𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐁𝐢𝐠𝐠𝐞𝐫 🌟 Last May, I embarked on an incredible journey with WisdomAI right after graduating from UIUC, and it has been nothing short of a dream run! Starting my career with a passionate, dynamic team like this one has been an amazing experience. 🚀 Here’s what I’ve absolutely loved about my time at WisdomAI so far: 1️⃣ Ownership Sparks Passion When you’re empowered with the right motivation and a sense of ownership, the task transforms into a pursuit. I’ve cherished the freedom to take charge of my projects, driven by curiosity for achieving results rather than just deadlines. 2️⃣ Culture of Growth & Impact I’m deeply grateful to the founders and my team for cultivating an environment of insane growth while delivering great value to our customers every single day. Huge thanks to the founders - Soham, Sharvanath, Kapil, and Guilherme for fostering this environment. 3️⃣ The Power of Team Energy The motivation and drive of the people around you can make all the difference. Being surrounded by such an energetic, supportive team has truly set the trajectory for my career. As we step into this new year, we’re doubling down with more energy, sharper focus, and even bigger goals for our startup. Here’s to a year of innovation, impact, and scaling new heights together! 💡🌍 #StartupJourney #GrowthMindset #WisdomAI #Teamwork #Innovation #NewYearGoals #CareerJourney

136

7 Comments

I recently completed my summer research at Columbia Summer Undergraduate Research Experiences in Mathematical Modeling (CSUREMM)! Throughout the program, I explored how subway station centrality influences the performance of public high schools in Manhattan. This opportunity allowed me to delve deeply into Graph Theory, Multiple Linear Regression, Principal Component Analysis, and K-Means Clustering, which led to intriguing findings. 🔍𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀: 🔹𝗠𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝗟𝗶𝗻𝗲𝗮𝗿 𝗥𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝗼𝗻: The models exhibited limited predictive power of subway station centrality on school performance. 🔹𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗮𝗹 𝗖𝗼𝗺𝗽𝗼𝗻𝗲𝗻𝘁 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀: Identified three key components (Neighbor-driven Centrality, Distance-driven Centrality, Performance Metrics) which captured most of the variance in the data. 🔹𝗞-𝗠𝗲𝗮𝗻𝘀 𝗖𝗹𝘂𝘀𝘁𝗲𝗿𝗶𝗻𝗴: Revealed that neighbor-driven and distance-driven centrality PCs generated similar clusters, indicating that the clustering approach effectively captured centrality-performance patterns. Showcasing my research at: 🔹Columbia Undergraduate Pathway Programs Symposium 🔹Summer Undergraduate Math Symposium 🔹Summer Math Day at Jane Street allowed me to gain new insights, deepen my enthusiasm for the research topic, and connect with like-minded peers. I would like to thank Caroline Smyth and Kathy Xu for their incredible collaboration on this project, as well as George Dragomir and Vihan Pandey for their invaluable mentorship and guidance throughout the challenges we faced.

193

27 Comments

🚀 BostonHacks 2024! Last weekend, Andrew Sun and I won 2nd place for the CometCare track at BostonHacks 2024, which was our first Hackathon. We had a lot of fun working from the ground up to create a final product called GoodDeeds, a platform designed for NGOs, non-profits, and other local organizations looking to post about events requiring community involvent across the spectrum of urgency, ranging from severe post-disaster recovery efforts to more relaxed community service initiatives. 🏢 For Organizations Organizations can sign up to create events, specifying a location, urgency rating, volunteer needs, and a description of the event. 🙋 For Volunteers Individuals like you and me can sign up as volunteers with a specified “home” location and a distance radius we’re willing to travel. Volunteers can view all events within their set radius, filter by date and start time, and sign up. Their email is added to a mailing list for updates, and they can reach out to event organizers or ask questions to an OpenAI chatbot directly within the app. 🎮 XP Levels To encourage community involvement, we introduced a game-like feature: XP levels for individual users that increase with volunteering attendance. Higher XP points are awarded for events with higher severity ratings (like post-hurricane cleanup). Users and organizers can see a leaderboard displaying top volunteers with their XP points and approximate location. 📣 Real-Time Notifications We integrated a notification system, allowing organizers to alert individuals within a set radius about upcoming events or updates for events they’ve signed up for, powered by the Twilio SendGrid API. The emails look nice, too! 💻 Tech Stack We had a lot of fun learning Streamlit, a frontend library built in Python, along with building a Flask backend and using a PostgreSQL database hosted on Neon. The frontend was deployed on Streamlit Cloud, while the backend was deployed on Render. Given the short time constraint, we’re thrilled with what we accomplished, and we look forward to future development! 🌐 Check It Out! Demo Video: https://lnkd.in/ecJyGufP Live Site: https://lnkd.in/ey5ht6yG (starts cold, may take a while on first signup) Source Code: https://lnkd.in/ew7PJySZ

37

14 Comments

Week 4 of CUNY Tech Prep: I'm excited to be working with Daniel Chen and Andrew Lam, RT(R) throughout this semester in building a simple reinforcement learning data science project. Daniel is amazing at coding and has won several Hackathons in the past, including the CTP Hackathon that I participated over this past summer. Andrew and I met through a mutual friend, and we have many interests in common, including a background working in healthcare. This week, I watched a video on Clustering in Data Science, which I have attached below for your reference. This video introduced me to the concept of Clustering, which is the grouping of data based off a measure of similarity. The video continues to show a demo of creating dendrograms from a dataset in R. I thought that it was interesting because I recently worked with R and RStudio in a research project for Collegiate Science and Technology Entry Program (CSTEP). I am excited to recreate the same demonstration using Pandas and NumPy in Python.

7

CUNY Tech Prep: This week, we explored open-source models with Hugging Face. For my wordle group project, I’m planning to incorporate Meta's NLLB (No Language Left Behind) translation model, which supports over 200 languages. The model uses Torch tensors, essentially multi-dimensional arrays that efficiently process large-scale data, to handle translations smoothly. By integrating NLLB, I can offer translations and personalized hints in real-time, making the game accessible to a diverse audience. Checkout Meta's model here: https://lnkd.in/eCbZJz6x #MachineLearning #DataScience #CTP #HuggingFace #Meta

27

Tonight I had the pleasure of meeting Astrophysics Professor and Chair of the Physics Department at Carnegie Mellon University, Dr. Scott Dodelson in a lecture entitled “Challenges in Modern Cosmology”, arranged by UChicago Ryerson Astronomical Society. Dodelson is a leading expert in cosmology, focusing on the large-scale structure of the universe and the physics of dark matter and dark energy. Author of ‘Modern Cosmology’ and—one of my favourites—‘Gravitational Lensing’, his work bridges theory and observation, using tools like lensing and the cosmic microwave background to unlock the mysteries of our universe’s origins and evolution. In this talk, Dodelson discussed how the standard cosmological model, while successfully fitting all current data, relies on phenomena like dark matter, a cosmological constant, and inflation—none of which are observed in the lab—and posed the critical question of how to advance as we prepare for the most ambitious astronomical data sets in history… not to mention (as he bluntly put it) to disprove the ΛCDM model!

31

My team and I were proud to present our research at UC Berkeley College of Computing, Data Science, and Society’s Discovery Symposium. Through the Data Discovery Program, I had the opportunity to conduct machine learning research with BART (Bay Area Rapid Transit). Our project focused on developing neural network-based models to forecast overheat events in train control rooms. By analyzing complex temperature trends, along with their velocities and accelerations, we aimed to enhance operational reliability for one of the Bay Area's most vital transit systems. I am proud to share that our Gated Recurrent Unit (GRU) neural network outperformed the previously deployed predictive model. This has been such a fulfilling opportunity to dive deeper into machine learning and explore different areas within data science. It's an exciting time in my life as I continue to discover my passions and expand my technical skills! A huge thank you to our mentor, Dr. Yu (Aaron) Shen, for guiding us, and to my teammates Ayumu Ueda, Luca Butera, Yuwei Chen, and Tiange Chen for their hard work and intellectual contributions. I'm incredibly grateful for this experience and for the chance to showcase our work alongside so many talented Berkeley students! 🚉📊

87

14 Comments

The Babel Effect: Multilingual Performance Discrepancies in LLMs Excited to share my research, "The Babel Effect," which explores performance gaps in Large Language Models (LLMs) between high-resource languages like English and low-resource ones like Nepali. The paper dives into tokenization challenges, data scarcity, and training inefficiencies. We propose solutions like data augmentation and improved tokenization to address these issues. Read more here: https://lnkd.in/deWwJ4uN #AI #NLP #Tokenization #ML

13

4 Comments

Stevens just wrapped up its annual Algorithmic Trading Competition, and I had an incredible experience working on the implementation of Reinforcement Learning in High Frequency Trading. Thanks to this opportunity, I gained invaluable hands-on experience beyond pure theory in RL. I'm excited to continue exploring this field and see where it takes me. #Stevens #Algorithmic_Trading_Competition

12

3 Comments

What society values may not be the best thing for you. When I applied for grad school, I got admits from MIT, Stanford, Purdue, CMU and UIUC. I knew I wanted to do a PhD. I chose MIT immediately, because it was ranked number 1. I never thought whether it was a good decision or not. I just assumed that it’s a good decision because it’s rated higher than the other universities. When I joined MIT, all that rank 1 noise disappeared. All I was surrounded by was my lab, my lab group members and my advisor. I had never carefully considered whether this research group was a good fit for me. I just joined the first MIT research group which came my way. This led to a miserable first 2 years of PhD for me. The group was not a good fit and the research topic was also not a good fit. These lost 2 years were a result of choosing what society values more (university ranking) rather than what is a better thing for me. When I look back, I should have done the following: (1) I should have made a list of all research groups I was interested in, from all universities I was selected to. (2) I should have spoken with the lab advisor from all these research groups and understood their vision. (3) I should have learnt more about the group culture by talking to past members of the group. Graduating members are usually very honest. They would tell you if the culture is toxic, if the advisor is micromanaging etc. (4) I should have looked at the group website to look at the publication rate, what do the alumni end up doing etc? If the average graduation time is 8 years, it’s a red flag. After doing the above 4 things, I should have selected the best research group for me. I should have selected the research group whose culture aligns well with my personality. Society values ranking. However, when you enter the university for a PhD, the ranking does not matter. Your research group matters. What you do every day after waking up matters. And if you blindly select universities based on the ranking, the research group won’t be a good fit and it would lead to a terrible PhD experience. It turns out that society misguides people in many other professions including: - Choosing to prepare for IIT-JEE just because society prefers that you are an IITian - Choosing a job just because society prefers higher salary - Choosing to not do a startup because the society (your parents) don’t approve Always question yourself before blindly making decisions based on what the society values. It might not be the best thing for you. Going against traditional societal values is hard in the short term, but it pays off in the long run. P.S: In the photo, you can see me during my first day at MIT!

308

6 Comments

My time as a research engineering intern at Google DeepMind has come to a close! As someone whose world revolves around networked, computer, and distributed systems, I never imagined I'd land on my dream team – the model partitioning team! This summer, I worked on PartIR's automatic model partitioning capabilities(https://lnkd.in/egp5Ygvy), diving deeper into distributed computing and exploring applications in reinforcement learning. It was definitely a challenging project, but I picked up so many invaluable skills, from software engineering to research in an industry setting. I’d like to give a huge thank you to Sami Alabed, Masha Samsikova, and the entire model partitioning team for making me feel so welcome and helping me successfully complete my project! I’m also thrilled that my final presentation was well received (and even got some laughs!), helping many at GDM understand a complex topic. Outside of my project, I had the chance to meet so many fascinating people across different teams and, of course, the other interns! A big thanks to the GDM Academy organizers, Charlotte Colson and Claudia Pope, for their incredible support in running such a fantastic program while I was balancing my thesis. And to Lucia Lopez Rivilla and Ioana Mihăilescu, former GDM Academy interns, for helping me find my way around the company! This summer wasn’t without its challenges, though. I finally addressed the anxiety issues I’d been struggling with since my teenage years, as they were impacting my studies and relationships. Unfortunately, I wasn’t able to attend my final year exams as I had hoped. However, I'm so grateful to UCL's Computer Science team and Faculty of Engineering for giving me the flexibility to extend my studies by a year, without penalty, and extending my scholarship. I couldn’t be more thankful for their support! All in all, this experience has solidified my determination to pursue research. In my third year, I discovered my passion for networked systems, and during my time at AMD, I knew I wanted to work on distributed systems. Now, after my time at GDM, I’m confident that distributed and networked systems are my future – and with the added experience of AI, I feel more prepared to navigate the machine learning-filled computer science landscape. Here’s to a (hopefully) restful but equally exciting year ahead! See you soon, Morgane 🦆

219

16 Comments

Knock knock! 🚪 Who's there? 🤔 It's GLaMM! 🌟 Excited to share that our paper, GLaMM: Pixel Grounding Large Multimodal Model, will be presented at CVPR 2024! 🎉🎉 tomorrow (Jun 20, 2024, Thursday) at 10:30 AM, poster #326! 🔥 Why GLaMM is special 🙌: 1. First end-to-end LMM tackling multiple vision-language tasks like Grounded Conversation Generation (GCG), image captioning, region-captioning, referring expressions segmentation, visual question answering, and more in a single model. 💥 2. We developed a fully automatic image annotation pipeline to build GranD dataset. 📊✨ 3. Fully open-source: Code, models, dataset, and dataset generation pipeline. 🚀💻 Check it out: GranD Dataset: https://lnkd.in/eYuKia9C GranD-f Dataset: https://lnkd.in/dYcjC_e8 Code: https://lnkd.in/d3yFnPRt (Do not forget to ⭐️ our repo). Come meet the team (Hanoona, Sahal, Salman, Ming-Hsuan, Fahad) at Arch4A-E at poster #326 to know more! 😊👩‍💻👨‍💻 #CVPR2024 #MultimodalModels #VisionLanguage

147

4 Comments

Hello LinkedIn Friends, I’m excited to introduce you to Nolan Maske from NCA&T University. Nolan and his team won first place in the Fall 2024 Hackathon with Goldman Sachs! Even more impressive is the application they developed to win the competition. They created a web app called PredictFi, designed to predict future earnings for financial companies. Congratulations, Nolan and team! If you follow my posts on LinkedIn, you know that Market Leader Game (MLG) also predicts future earnings. At first glance, you might think PredictFi and MLG are similar. Both use statistical modeling to forecast earnings. For example, MLG’s algorithm identifies NASDAQ or NYSE stock symbols that are most likely to experience significant price increases over the next 12 months. @Nolan, it would be fascinating to learn more about the context and objectives behind future earnings predictions in PredictFi. MLG selects stock symbols with the highest probable growth rates by combining and then comparing: - Monte Carlo-generated stock prices, - Actual growth rates from the past 12 months, and - Detailed analysis of income statements and balance sheets. I purchased 3 stocks in 05/24 and 2 more in 10/24 to test the model. The performance (as of today) has been remarkable - 148% growth in 6 months: The difference between MLG and PredictFi became clearer when I asked Nolan if his team used another statistical model to derive future earnings. Nolan replied (paraphrasing): "PredictFi uses ARIMA (AutoRegressive Integrated Moving Average). The application assumes past patterns in data will continue. The team did not have sufficient time to explore the uncertainties and random outcomes that Monte Carlo could address." His response piqued my interest, leading me to dive deeper into the differences between Monte Carlo Simulation and ARIMA. Here’s what I found: ARIMA Modeling Strengths: Time-series focus: Well-suited for capturing trends, seasonality, and autocorrelations in stock prices. Deterministic forecasts: Based on past patterns and relationships within the data. Simpler assumptions: Assumes that historical patterns will persist. Monte Carlo Simulation Strengths: Flexibility: Models multiple potential price paths using randomness and volatility. Captures stochastic behavior: Reflects the uncertainty and randomness of financial markets. Use Case: Best for long-term risk assessment, scenario analysis, and when uncertainty and volatility are key factors. Conclusion Use ARIMA for short-term forecasting and when historical trends dominate. Use Monte Carlo for long-term forecasting, risk assessment, and probabilistic scenarios. Combining these methods or integrating them with other models (e.g., machine learning or GARCH for volatility analysis) often yields the best insights. I hope this comparison sheds some light on these powerful predictive tools. Congratulations again to Nolan and his team for their impressive achievement!

8

🚀 Excited to share some insights from a recent paper titled "TourLLM: Enhancing LLMs with Tourism Knowledge" by Qikai Wei, Mingzhi Yang, Jinqiang Wang, Wenwei Mao, Jiabo Xu, and Huansheng Ning! 🌟 Highlights: Addressing the Gap: Large language models (LLMs) like ChatGPT and Llama excel in various NLP tasks but struggle with tourism-specific knowledge. To bridge this gap, the authors introduced a specialized dataset and model fine-tuning approach. Cultour Dataset: The team developed a supervised fine-tuning dataset called Cultour, tailored for the culture and tourism domain. This dataset comprises: Tourism knowledge base QA data Travelogues data Diverse QA data Introducing TourLLM: TourLLM is a Qwen-based model fine-tuned with the Cultour dataset. It aims to improve the quality of information provided about attractions and travel planning. Evaluation: The effectiveness of TourLLM was validated using both automatic and human evaluation methods. The researchers proposed a novel human evaluation criterion named CRA (Consistency, Readability, Availability). Results: The experimental results demonstrated that TourLLM significantly enhances the quality and accuracy of responses in the tourism domain, offering more reliable travel advice and enriching user experience. 🔍 Key Takeaways: Customized Fine-Tuning: By creating domain-specific datasets and fine-tuning models accordingly, we can significantly improve the performance of LLMs in specialized areas. Innovative Evaluation Metrics: Developing tailored evaluation criteria, like CRA, ensures that the models meet the specific needs and expectations of users in different domains. Research Paper:https://lnkd.in/gruBhP6T #AI #MachineLearning #Tourism #NLP #LLMs #DeepLearning #Innovation

1

Proud to share our work on enhancing LLM efficiency through dynamic early exit strategies! Collaborating with Juan D. Castano and Amir Voloshin, we built on Meta’s LayerSkip framework to introduce interpretable heuristics that identify when token predictions stabilize. Excited about the potential this unlocks for faster, more efficient LLM inference. A big shoutout to my amazing teammates for their hard work and insights!

22

🎉 Thrilled to announce that I have three papers accepted at EMNLP 2024! 🚀 Here’s a glimpse into each: 1️⃣ Enhancing Pre-Trained Generative Language Models with Question Attended Span Extraction on Machine Reading Comprehension (led by myself, co-authored with Zack Hui and Gary Liu) 🌟 In this work, we tackle the limitations of generative models for extractive machine reading comprehension (MRC) by introducing QASE, a lightweight module that significantly improves performance without increasing computational demands. QASE enables PLMs to surpass GPT-4o in few-shot settings, achieving or exceeding state-of-the-art (SOTA) results across multiple datasets.   📄 Read more: https://lnkd.in/esuJC5Ga 2️⃣ Defending Against Social Engineering Attacks in the Age of LLMs (jointly led by myself, Tharindu Kumarage, and Amrita Bhattacharjee) 🌟 This paper explores how LLMs can both facilitate and defend against chat-based social engineering (CSE) attacks. We introduce SEConvo, a novel dataset simulating CSE attacks and defenses in real-world contexts. Our proposed framework, ConvoSentinel, systematically flags malicious messages and improves detection across entire conversations, offering a cost-effective and adaptable solution.   📄 Read more: https://lnkd.in/ebvDQBkm 3️⃣ A Survey on Open Information Extraction from Rule-based Models to Large Language Models (jointly led by Pai Liu, myself, and Ziwei (Sara) Gong, and more) 🌟 This comprehensive survey covers the evolution of OpenIE from 2007 to 2024, mapping its trajectory from rule-based methods to neural and pre-trained LLMs. It highlights key datasets, evaluation metrics, and offers insights into how traditional OpenIE research can inform future developments in the LLM era.   📄 Read more: https://lnkd.in/eJ8D_sBz #EMNLP2024 #NLP #AI #LLMs #Cybersecurity #InformationExtraction

70

3 Comments

🌪 Tracking the Storm: Visualizing Trends in U.S. Hurricanes and Climate Impact🌍📊 Over the past semester, I had the opportunity to work on an exciting project that blended my passion for data visualization📈 with a critical global issue—climate change and its impact on hurricanes. This work was part of the Data Visualization course, guided by the inspiring Prof. YY Ahn. Collaborating with my talented teammates Adithya S., Mohit Sai Tatineni, and Rasaghna Kuturu, we analyzed decades of hurricane data to uncover patterns in their frequency, intensity, and geographic distribution. By using techniques like geospatial overlays, multivariate scatterplots, and temporal trends, we revealed key insights into hurricane behavior, including the relationship between wind speed and pressure. 💡 Key Learnings: 1. Gained hands-on experience in cleaning and analyzing complex datasets. 2. Developed proficiency in geospatial and temporal visualization techniques. 3. Learned how to effectively communicate data-driven insights through storytelling. 4. Strengthened my ability to collaborate within a team, balancing creativity and technical rigor. I’m incredibly grateful to Prof. YY Ahn for his guidance and to my teammates for their creativity and hard work. This project reinforced the value of data visualization in addressing global challenges and bridging the gap between science and actionable policy. 🔗 You can explore the full project here: https://lnkd.in/gNAxprYW. I’d love to hear your thoughts!

22

2 Comments

🌪️ Hurricanes Through the Lens of Data (Because Someone Had to Decode the Chaos) 🌍📊 I’m happy to be a part of this project under the guidance of Prof. YY Ahn, where I took a deep dive into decades of U.S. hurricane data, uncovering trends that make it clear: climate change isn’t kidding around. From rising storm intensity to seasonal peaks and high-risk regions, this project shed light on how hurricanes are getting bolder and angrier—thanks, global warming. With geospatial visualizations and temporal analyses, we untangled the story behind hurricanes, aiming to turn chaotic weather data into actionable insights for disaster response and climate preparedness. After all, someone has to make sense of all this, right? Let me know your thoughts (or if you’ve figured out how to calm hurricanes down). 🌍💡

15