Varun Velpula

Interesting article about Generative AI! Subscribe to Rishabh Yadav's blog for many more!

10

🎇 On National Day, I went for a leisurely drive in San Francisco and ended up "stress-testing" a Waymo self-driving car on the road. 🚗 As an autonomous driving practitioner, who wouldn't be curious about the real-time robustness of the cutting-edge Waymo One driving system? While cruising downtown, I accidentally noticed a Waymo car tailgating me. While this isn't unusual for SF citizens, a wild, "evil" idea suddenly hit me: Why not directly "adversarially attack" the world's autonomous driving status quo? I executed an unexpected maneuver—suddenly reversing—to see how it would react. 🌟 The response was stellar! At the moment I reversed, Waymo One honked instantly—quicker than any human could—activated its hazard lights, and backed away to maintain a safe distance. This reckless move on my part served as an edge case to test the algorithm's robustness under extreme conditions and, potentially, could be a challenging training sample to enhance Waymo's future autonomous systems. 💫 Thrilled to personally experience how current cutting-edge autonomous algorithms handle rare driving behaviors—and how stable and safe Level 4 autonomy is in dealing with diverse scenarios. However, it also prompted deep reflection as an AI researcher in this field: 🤔 In an industry with little room for error, how can we ultimately avoid or minimize issues that AI fails to handle? 💡 I believe two research directions are particularly promising for achieving Level 5 level autonomy in future mobility systems: - 1️⃣ Development and deployment of vehicle-to-everything (V2X) cooperative systems (including V2V, V2I, V2P, etc). Our initial studies (e.g., V2X-ViT, ECCV'2022 arxiv.org/abs/2203.10638) show that in scenarios with severe occlusions or noise, such cooperative systems can significantly enhance the robustness of perception systems, thereby eventually improving traffic safety. - 2️⃣ Adversarial scenarios generation (including Sim-to-Real, generative modeling). Research done by my colleagues at UCLA (V2XP-ASG, ICRA'2023 https://lnkd.in/gu5nKVHD) shows that adversarial learning techniques can effectively simulate adversarial scenarios, greatly improving model robustness in complex "corner case" situations. Of course, it's often infeasible to collect such collision scenarios. 👨‍🏫 As a new Assistant Professor of CS at Texas A&M University, I will lead a group focusing on these exciting research directions, which can be a proactive approach to reducing accidents and improving safety for all. 🔥 I look forward to future collaborations with governments, academics, and companies to research and develop data and algorithms that can help enhance the safety of vulnerable road users, especially seniors and children. We envision a people-centered intelligent transportation system in the future. Interested in these topics? Let's connect and discuss further! #AutonomousVehicles #AI #MachineLearning #SmartCities #Transporation #Humanity #Mobility

144

3 Comments

Here is another fun app using local LLMs! #Ollama Debater lets you set up two #AI agents and have them debate on an arbitrary topic! Check out the code here: https://lnkd.in/eR3X5vNt The app is built using #Gradio

38

3 Comments

Join me this Saturday for a seminar with Aditi Paul, PhD! 🗓️ We’ll explore the “EB1 algorithm”, discuss strategies for providing unique, in-depth evidence in EB1 petitions and how to reduce variability to increase approval chances. Don’t miss this opportunity to gain valuable insights. #EB1A #EB1B #Immigration #TechProfessionals #AI #ML

22

2 Comments

As Waymo enters an era of scaling they are making a big show of publishing safety data. But not all is as they want it to seem. Scaling, especially to serve new cities in addition to adding vehicles to the fleet, is where the risk of edge cases to safety is going to be the most challenging. Higher operational speeds (e.g., for airport service) additionally introduce a higher fatality risk to any crash they do have. The safety statements being made by their public and government relations apparatus are misleading. Maybe this will work out, and maybe it won't. But this data is in many ways just a distracting sideshow from the core safety question of how fatalities will turn out. Waymo hopes that will turn out well (and so do I), but it is irresponsible to practice what amounts to transparency theater and make misleading safety statements in the pursuit of fast scale-out. Details here in this week's substack posting:

67

10 Comments

Check out my recent work with Richard Lehoucq and Brian Franke: A Stochastic Calculus Approach to Boltzmann Transport. https://lnkd.in/ggXxnQXc I am thrilled that this paper is finally published. It shows how the power of #probability and #StochasticCalculus can uncover unique relations in #TransportProblems. We described a common Monte Carlo procedure used to simulate a particular class of Boltzmann transport problems with a stochastic differential equation. Through the use of a representation theorem, we demonstrated that the Monte Carlo process also naturally gives rise to the adjoint solution.

28

Built an AI Agent for complex decision making in a strategic zero-sum game using Reinforcement Learning & Monte Carlo Tree Search. Thrilled to share the completion of a challenging yet rewarding project in Reinforcement Learning: an AI agent that can play zero-sum games, inspired by AlphaZero! This project was a unique blend of strategy, complex decision making, game theory, and deep reinforcement learning Play against the AI Agent on a 6x6 Tic-Tac-Toe board with pie rule in Streamlit: https://lnkd.in/gMcyrKRE Experience firsthand how the AI makes its moves! The project involved training an AI agent to play a zero-sum game. The AI agent uses two key techniques: - A Deep Neural Network analyzes the game state and predicts the best move, while also estimating the potential outcome. - Monte Carlo Tree Search (MCTS) simulates many possible game scenarios based on the neural network's predictions and finally chooses a move. The Neural Network is trained using a method called Actor-Critic Reinforcement Learning. This involves cycles of evaluation, prediction, and simulation - the core aspect of reinforcement learning. The agent learns and improves by receiving feedback (positive or negative) based on the outcomes of its simulated games played against itself. Trained in a CPU. Github link: https://lnkd.in/gm3s6XPa #reinforcementlearning #ai #artificialintelligence #deeplearning

13

"Foundations for Machine Learning: New Course Launch" In 2022, I graduated with a PhD in Mechanical Engineering from MIT. Although a big component of my research was purely hands-on experiments, my exposure to foundational graduate-level ML courses at MIT, research courses, and Scientific Machine Learning via Julia gave me the confidence of a Machine Learning researcher. I incorporated ML into my research, and it solved a problem that are otherwise difficult to solve theoretically or experimentally. When I co-founded Vizuara AI Labs with Raj and Rajat, our core principles were depth and foundational knowledge. Now I have co-authored multiple AI-ML research papers and two of them are accepted to the upcoming NeurIPS workshop. Behind all of this effort, there is the confidence that stems from knowing what happens underneath the ML algorithms. Most of the online courses have little emphasis on fundamentals. People are so used to spending time on toy Kaggle projects. Very few people I know can build a neural network from scratch or explain what happens behind them. For the last 4 months, I have been working to launch a new course on Vizuara's YouTube channel on "Foundations for Machine Learning." This will be a 45-hour course with ~65 lectures. I will be hosting all lectures on this playlist: https://lnkd.in/gazvPG_b My singular goal with this course is to teach you the entire foundations required to learn ML from scratch. There are no prerequisites. If you have basic logical thinking capability and a willingness to dedicate time, consistently, you can follow this course. I have split the course into 5 modules. 1) First I will be covering the 4 mathematical pillars of ML: Linear Algebra, Probability, Statistics, and Calculus. 2) In the second module I cover the basic programming fundamentals for a complete beginner. I will teach you Python from scratch and it's some of the most important packages for ML including NumPy and PyTorch. 3) In the 3rd module we will learn about optimization and gradient descent. I wanted to dedicate an entire module to optimization because when you actually build ML models, you will be spending a lot of time on optimization. 4) In the 4th module, I will give you an overview of the AI landscape. What happened from 2010-2020 and what does it look like from 2020-2030? This overview will help you understand overall where ML, DL, NLP, CV, and GenAI are heading. 5) In the final module, I will cover the most important 2 steps you will have to master as a Data Scientist or ML engineer: processing data and communication via storytelling. I will teach you some of the most powerful preprocessing and visualization techniques. I have already published the first lecture. Check out here. I am sure you will enjoy and learn a lot: https://lnkd.in/gj3CX2Eh

153

6 Comments

New dataset from Stanford Energy Control Lab 🚨 I am excited to share that a new dataset from our lab, “Second-life lithium-ion battery aging dataset based on grid storage cycling”, co-authored with Dr. Kevin Moy, Simone Fasolato, Dr. Gabriele Pozzato, Ph.D., Dr. Anirudh Allam, and Dr. Simona Onori, is now available online in Data-in-Brief! Retired batteries with unknown usage histories require advanced battery management systems (BMS) to ensure safe operation and to perform accurate health diagnostics and prognostics to extend their lifespan. However, the lack of publicly available second-life datasets has been a significant hurdle in advancing research on the feasibility of second-life batteries for stationary applications, such as grid storage. Our work addresses this challenge by sharing a publicly available dataset of second-life batteries. 💡 Highlights: ·      A batch of used NMC/graphite+silicon cells, aged using an Urban Dynamometer Driving Schedule (UDDS) profile as described in Pozzato et al., 2022 (https://lnkd.in/gA44hfN6), was cycled using synthetic grid-storage duty cycles, simulating residential and commercial applications, over a 24-month period. ·      Seasonal temperature variations were replicated by alternating between 20°C and 35°C during cycling. ·      Cells experienced both cycling and calendar aging during the experimental campaign, and characterization was conducted using a combined capacity and pulse test, and electrochemical impedance spectroscopy (EIS). This dataset underpins the importance of developing battery aging models and algorithms based on application-specific data to maximize performance over the battery's lifetime. We hope it will inspire exciting research and further advancements in the field of second-life batteries! ⚡ Bonus: We are continuing to collect more aging data on these cells, and we will periodically update the data repository! Paper link: https://lnkd.in/g-_9XiQm   Data repository: https://osf.io/8jnr5/

71

1 Comment

From Part 2 of my Waymo Emails: Not many engineers pay attention to fluids and how they impact sensors, but Waymo does. In their blog post « Guided by the Flow » they show how using Computational Fluid Dynamics (CFD) simulations helps reducing drag and improving sensor longevity. It’s not a big deal on one trip with one car, but when preparing for mass production (as I believe Gen 6 of sensors is doing), it could mean a big difference over the years. #selfdrivingcars #robotics #lidar #computervision #deeplearning #ai

78

4 Comments

🚗 Excited to Share My Work on Autonomous Vehicles! 🚗 I recently had the opportunity to collaborate with a Forbes writer to publish the article, “What Do Consumers Really Feel About Autonomous Vehicles?” 📰 A huge thank you to our CEO, Sharon Joseph, for providing me with this incredible opportunity at CREWASIS.AI! Through this project, I offered insights into user sentiments around autonomous vehicles—uncovering public perceptions, concerns, and expectations. It was a fantastic chance to delve into consumer behavior analysis and contribute meaningful data to an evolving area in tech! Check out the article here: https://lnkd.in/dgnsqccN #AutonomousVehicles #ConsumerInsights #ForbesCollaboration #DataScientist #Crewasis #Innovation #FutureOfMobility

26

10 Comments

A New Theory for Managing Chaotic Systems: The Proposal of Weighted Stochastic Riccati Equations Many systems around us involve inherent "chaos." For instance, in remote robotic control systems that experience random communication delays, designing a control law to achieve a goal, such as moving a robot to a target location, requires considering the chaotic nature of these delays. However, chaotic systems are mathematically challenging to handle, making traditional methods limited in scope, computationally intensive, and lacking generality. To address this issue, Y. Ito from TCRDL, in collaboration with Kyoto University, has developed the Weighted Stochastic Riccati Equation (WSR), a theoretical framework for designing control laws that accommodate diverse goals while accounting for stochastic characteristics. By incorporating "weights" into stochastic elements, the WSR equations enable the design of various control strategies. These strategies range from simple ones that average out the chaos to more complex ones, such as "moving slowly to avoid falling" or "prioritizing speed even at the risk of instability." This groundbreaking research is expected to have applications in fields involving chaotic systems, such as robotics and energy systems, contributing to the future development of safe and efficient system designs. https://lnkd.in/gmbYEzn9 https://lnkd.in/gK8GSJEH https://lnkd.in/g93HRugw

8

Robot taxis are encountering ongoing challenges, as highlighted in a recent article by Bloomberg. The piece provides sobering examples of recent setbacks, emphasizing the hurdles that companies like Waymo, Cruise, and Zoox are facing as they strive to advance in the #robotaxi race. Bloomberg's newsletter, aptly titled "Getting out of neutral," offers insightful perspectives on the current state of autonomous vehicle development. In my graduate course, we delve deep into the technical obstacles and core principles of autonomy governing autonomous vehicles from an AI perspective. Rather than merely focusing on control mechanisms, we explore the nuanced realm of AI-driven decision-making and navigation in these intricate systems. I'd like to extend a special shoutout to my students Jan D. and Sidd Argawal, who have recently joined the ranks of industry leaders. Jan, who pursued his PhD, went on to join Waymo, while Sidd, who completed his MS, embarked on a journey with Ford. They've taken the AI robotics principles learned in our course and applied them to their careers, driving innovation in the field of autonomous vehicles. https://lnkd.in/gmVwPFfT #SelfDrivingCars #AutonomousVehicles #Robots #AI #GraduateCourse #RoboticsExpert

24

3 Comments

Predicting how many years it will take for LLMs to replace 90% of software engineers is speculative, as it depends on numerous factors including technological advancements, industry adoption, regulatory frameworks, and societal acceptance. However, let's break down the likely timeline based on the current progression of LLM capabilities and industry trends: Current Capabilities - Code Generation: LLMs already generate high-quality code snippets. - Testing: LLMs are increasingly used for writing and running tests. Short-term (1-3 years) - Bug Fixing: LLMs will understand task management tools like Jira and fix minor bugs. - Simple Front-end Development: LLMs will write code for simple web pages within an existing codebase. - Website Building: LLMs will create basic websites from scratch. Medium-term (3-7 years) - Contextual Understanding: LLMs will better understand development environments and implement CRUD operations. - Complex Backend Operations: LLMs will handle more complex backend tasks, integrating with various systems and APIs. - Teamwork Integration: LLMs will start training other LLMs, improving collaborative coding environments. Long-term (7-15 years) - Widespread Adoption: More companies will adopt LLMs for a significant portion of software development tasks. - Advanced Project Management: LLMs will take on more comprehensive project management roles, understanding larger systems and contributing to design and architecture decisions. - Replacement of Roles: By this stage, LLMs could potentially replace a significant portion of software engineering roles, especially those involving repetitive or straightforward tasks. Beyond 15 years - Majority Replacement: If technological, ethical, and societal conditions are favorable, LLMs might replace up to 90% of software engineering roles, focusing humans on high-level strategic and creative tasks. Factors Affecting Adoption 1. Technological Advancement: Speed and direction of AI research and development. 2. Industry Adoption: Willingness of industries to integrate AI into their workflows. 3. Regulation and Ethics: Legal and ethical considerations regarding AI in the workforce. 4. Human-AI Collaboration: Evolution of hybrid working models where humans and AI complement each other. Given these points, a realistic estimate for LLMs replacing 90% of software engineers could range from 15 to 20 years, contingent on the rapidity of advancements and societal readiness.

4

1 Comment

This is #icra2024 Week, and I want to tell you about a very cool paper I spotted from Waymo's research team: Stateful Tracking with Transformers for Autonomous Driving! It's a 3D Multi-Object Tracker... but with Transformers rather than Kalman Filters. It's very "edgy", and I'm doing a complete breakdown in this video. Dragomir Anguelov (Drago) Qinru Li How did I do? Did I miss anything? #deeplearning #computervision #lidar #selfdrivingcars #robotics

119

6 Comments

Happy to share that our latest paper is out! In this study, we delve into the world of Slow Feature Analysis, a powerful technique to transform measured data into uncorrelated signals ranging from slow to fast. We've introduced a novel approach that goes beyond traditional methods, addressing the challenge of nonstationary and oscillating features. Our semi-supervised encoder-decoder architecture incorporates a statistical preference for these characteristics, paving the way for more accurate modeling. Curious about the results? We put our approach to the test on both simulated and real industrial processes, with promising outcomes. Thanks to my supervisor, Prof. Biao Huang. Read more about our findings in the full paper! https://lnkd.in/gK6naG5U

61

2 Comments

Thrilled to share that I recently participated in Nvidia's "Generative AI Agents" contest. The world of generative AI is evolving rapidly, and I jumped at the chance to explore its potential. While I didn't win the contest, the experience was a fantastic opportunity to dive deep into the world of generative AI agents and explore its potential to revolutionize the way we interact with technology. I built a "Personal Financial Advisor" application that leverages the power of AI to simplify financial planning and empower users to make informed decisions about their money. My application offers a range of features including: - Automated Expense Categorization - Targeted Expenditure Search - Personalized Savings Projections - Predictive Expenditure Forecasting - Budget Optimization Recommendations This project pushed me to delve deeper into the world of AI, learning valuable skills in prompt engineering, large language model fine-tuning, and agent design. I'm eager to see how generative AI transforms various industries and continue exploring its potential across the board. #Nvidia #GenerativeAI #AI #Contest #Innovation #Tech #Future #PersonalFinance #FinancialWellness

49

5 Comments

I'm excited to share my latest publication in the AIAA Journal, titled "Multifidelity Methodology for Reduced-Order Models with High-Dimensional Inputs." In the conceptual design phase of aerospace systems, leveraging physics-rich, high-fidelity information is crucial for exploring innovative designs and configurations. However, traditional high-fidelity techniques like CFD and FEA can be computationally expensive, especially for applications like design optimization, design space exploration, and uncertainty quantification. To address this, my research introduces a multi-fidelity reduced-order modeling framework that integrates machine learning concepts of manifold alignment and dimensionality reduction. This approach provides a computationally efficient and accurate solution for creating reduced-order models in high-dimensional design spaces. We validated our method using various aerodynamic test cases and found it offers a significant cost vs. accuracy benefit compared to current state-of-the-art methods. This methodology paves the way for integrating multiple data sources and techniques, like digital twins, in aerospace design. I extend my heartfelt thanks to my co-authors Dr.Christian Perron, and Dr.Dr. Dimitri Mavris for their invaluable support. You can read the full article on the AIAA Journal website. https://lnkd.in/eJAMwTvP

114

2 Comments