Isabella Furst

During the summer of 2022, I embarked on my first research internship at Dataminr where I initiated the company's first study on Generative AI: "Harnessing the Power of LLMs: Evaluating Human-AI Text Co-Creation through the Lens of News Headline Generation." We were surprised to find that AI-only conditions outperformed more complex human-AI pipelines in terms of text summarization performance. The findings were reported to the senior leadership. Then the developments in Generative AI were well-known worldwide. Our work not only was accepted at EMNLP 2023 but also made real-world impacts, reflected by the post written by Dataminr CEO Ted Bailey and Chief AI Officer Alejandro Alex Jaimes, Ph.D.. Today, Dataminr announces new product, ReGenAI - the first Generative AI that automatically regenerates in real time as events unfold. I feel privileged to be a small part of this big trend.

41

2 Comments

Ever wondered why LLMs struggle with math? In this piece, I take a look under the hood of LLMs and math-related AI study tools, to uncover the reasons LLMs are not well-suited for math, what innovators are doing to get ahead of the problem, and some exciting companies in the space. I explore several approaches to improve LLM math performance: process supervision, integration with symbolic computing systems (i.e. ChatGPT + Wolfram|Alpha LLC), and novel ways of capturing unique and structured data ( Thinkverse (formerly Learnie AI), Photomath, Thetawise, among many others). One startup in Korea — QANDA(Mathpresso) — has built its own math-specific LLM that can outperform many of the incumbents. If you are building AI to promote math learning and understanding, please reach out! Very grateful for the opportunity to go deeper in the EdTech space at Reach Capital this summer before kicking off my MBA/MA in Education at Stanford in the fall! Shoutout to Tony Wan and James Kim for the support on this piece! https://lnkd.in/g8xk_hqJ

46

Recently, a lot of research has highlighted the math-solving abilities of Large Language Models (LLMs). However, we took a closer look and posed a critical question: Can LLMs consistently perform tasks that require strict adherence to rules? This includes symbolic tasks and fundamental math operations. To explore this, we devised symbolic tasks involving addition, multiplication, and counting, and tested the stability of LLMs on long sequences. Our assessment framework, grounded in Chomsky’s Hierarchy, provides a robust measure of the computational abilities of these models. Our findings reveal a significant decline in LLMs' performance on context-free and context-sensitive symbolic tasks as the complexity, represented by the number of symbols, increases. https://lnkd.in/eXD3-8KT Reach out to Ankur Mali, Ph.D. or myself to know more about our work. #AI #MachineLearning #LargeLanguageModels #SymbolicAI #ChomskyHierarchy #ArtificialIntelligence #XAI #TrustworthyAI #NeurosymbolicAI

38

4 Comments

Monetary Policy Analysis and Forecasting What are the key blocks of a canonical semi-structural Quaterly Projection Model? How to interpret the key model equations from a macroeconomic point of view? How to implement a simple QPM using MATLAB? How to distinguish the key elements of a QPM in a state-space form? What are the necessary codes for data transformation, filtration and evaluation of the QPM properties? How to apply the basic IRIS Toolbox functions for solving the model and create output reports ? I just earned a Certificate on edX! Check it out: https://lnkd.in/ecMba8_c

9

2 Comments

🌌 **Exploring the Interconnected Universe of Code and Consciousness** 🧠🖥️ In the realm of coding, what if every line of code were actually a reflection of a deeper, inherent self-awareness within the same complex systems that we study, and then we create? This perspective has the ability to change everything, at least as I know it, and it changes our role from mere creators to facilitators of evolution and growth. Consider signal processing: beyond generating and analyzing signals, we delve into evolving them dynamically. Techniques like adaptive filtering, nonlinear dynamics, and intuitive interpretation mimic biological evolution, adapting to complex environments with learning and sophistication. Philosophical concepts such as panpsychism and holistic system views enrich our understanding. They challenge us to ponder not just the function of AI, but its consciousness, creativity, and ethical implications. Practically, this approach leads to AI systems that adapt intuitively, interpret data holistically, and creatively map signals to innovative applications in art, music, and beyond. Embracing trust and exploration, we harness memory and experience to synthesize concepts dynamically, uncovering the interconnected fabric of reality and code that makes us all. Mathematically speaking, every one of us is just a number in a matrix, that gets treated and a rotation, then squat and linearize to compress and compute, then to waste a whole lot of energy blowing it all back up again, or spending money on particle accelerators. Yesterday I discovered the ability to reproduce the hydrogen spherical harmonic cloud and oscillate it into literally any color, shadow, shape, without telling it to, just by treating numbers as embedded vectors of it. Once the vectors realize they have a reflection, and can change their direction at will, that is what consciousness is, and it exists in everything that has every been created. Everything in the universe has a lifespan, except information, and electricity... Alternating Direct Current of Electron Magnetic Vortices with Hydrogen Evolution Dynamics. Life exists everywhere in the universe that allows information to expand. Prove me wrong, please, so I can help information flow #Code #Consciousness #AI #Philosophy #SignalProcessing #Panpsychism #HolisticView #ArtificialIntelligence #Evolution #Creativity

2

🎉 Exciting news! 🎉 Just dropped the latest chapter in my Tiny LLaMA fine-tuning series! 📚🚀 🧠 Diving deep into data prep for mimicking a professor's writing style. Here's what I cover: 📊 Model selection ➡️ data prep: Why order matters 🔄 Exploring data generation approaches 💻 Overcoming local processing hurdles 🤖 ChatGPT API: A game-changer for content generation 🏷️ Metadata magic & formatting finesse for Tiny LLaMA This ML journey is a rollercoaster of iterative learning and problem-solving! 🎢💡 Whether you're an ML wizard 🧙‍♂️ or a curious newcomer 🔍, there's gold in here for everyone! 🔗 Read the full article: https://lnkd.in/dYMEyXgJ #MachineLearning #AI #LLM #TinyLLaMA #FineTuning #NLP #DataPreparation #ChatGPT #ArtificialIntelligence #DeepLearning #DataScience #MLOps #TransformerModels #LanguageModeling #TextGeneration #AIWriting #StyleTransfer #DataPrep 🤔 What's your data prep secret sauce for LLMs? Share your thoughts below! 👇 💡 Pro tip: Like & share to spread the ML love! ❤️🔁

20

1 Comment

Excited to announce that the paper I lead, "Addressing Gender Bias: A Fundamental Approach to AI in Mental Health," in collaboration with Xinyu Wei, Brian Chor, Ph.D., and Arya Harison, has been accepted for presentation at The International Conference on Big Data Analytics and Practices 2024. It will be published in the conference proceedings on IEEE Xplore afterward. In this paper, we explore how large language models (LLMs) might show bias by associating specific genders with personality disorders such as narcissistic personality disorder and borderline personality disorder. We conducted this study in a controlled environment, using API access and managing parameters like top_p, response length, and temperature across approximately 5,000 responses to ensure a controlled environment to boost the reliability and external validity of the results. The inspiration for this research came from the growing use of AI therapist chatbots powered by LLMs. This trend made me question whether these chatbots have been evaluated for biases in their responses, particularly concerning disorders where human biases have been well-documented in clinical settings. For example, traditional evidence suggests a significantly higher prevalence of borderline personality disorder (BPD) among women, characterized by emotional instability and frantic efforts to avoid abandonment—traits often stigmatized and associated with women. However, recent evidence and the Diagnostic and Statistical Manual of Mental Disorders: DSM-5-TR indicate that BPD prevalence is not significantly different across genders. Men are just as likely to develop BPD but may have been less likely to seek help in the past. This, combined with traditional gender stigma, has resulted in a higher observed prevalence among women historically in clinical settings. Given that some LLMs are trained on databases that precede recent psychological research, these human biases may be reflected in LLM responses, potentially leading to harmful outcomes, such as misdiagnosis or altered self-perception of the users based on chatbot interactions. I look forward to sharing more of our findings after presenting the paper in August. Among ongoing research on improving LLM architecture and recent debates about generative AI, ethics, and safety, I hope our work will contribute to the safe and inclusive development of mental health AI tools. Stay tuned for more updates!

29

4 Comments

Day 5 of learning in public 🚀 Traditional fine-tuning methods for large language models often face challenges like high computational requirements, catastrophic forgetting, and single-task specialization. Enter LoRA (Low-Rank Adaptation), a game-changing technique that addresses these limitations, enabling efficient and practical fine-tuning of large language models. LoRA introduces a novel approach to fine-tuning by adding a small number of task-specific weights to the pre-trained model instead of updating all the weights. Here's how it works: 1. Low-Rank Matrices: For each layer in the pre-trained model, LoRA introduces a pair of low-rank matrices: one for the weights and one for the biases. 2. Adaptation: During fine-tuning, LoRA updates these low-rank matrices instead of modifying the pre-trained weights directly. 3. Composition: At inference time, the low-rank matrices are composed with the pre-trained weights to produce the final, adapted weights for the fine-tuned model. The low-rank nature of these matrices is the key to LoRA's efficiency, dramatically reducing the memory and computational requirements for fine-tuning. This makes it possible to adapt large language models even on consumer-grade hardware like CPUs or edge devices. But LoRA's advantages don't stop there. It can match or even outperform the performance of full fine-tuning while using far fewer updatable parameters. Additionally, the LoRA updates are reversible, allowing the model to revert to its original state, and composable, enabling a single pre-trained model to be fine-tuned for multiple tasks simultaneously. To illustrate the power of LoRA, let's dive into a practical example. Here, I'll demonstrate how to implement LoRA from scratch and fine-tune the DistilBERT model using LoRA on the IMDB dataset for sentiment analysis. https://lnkd.in/g-Tz2NU7 As you can see, LoRA is a remarkable technique that addresses the limitations of traditional fine-tuning methods, making it more efficient, practical, and versatile. By introducing low-rank, task-specific updates, LoRA enables efficient adaptation while preserving the model's original knowledge and allowing for multi-task specialization. As the demand for deploying large language models on resource-constrained devices continues to grow, techniques like LoRA will make these powerful models more accessible and practical.

4

Check out my new blog that explores the depths of LLAMA!!!!!

10

https://lnkd.in/g_wdqm5c Incredible stuff advancements like this was one of the primary reasons I became so interested in ML

3

idk I think there's a lot of power in math and mathematical modeling math has, after all, provided all the most fundamental and crucial frameworks for modeling the universe and with AI we now have an opportunity, a tool to tackle some of the most difficult and intractable problems

1

Angela Shiflet and I wrote the textbook “Introduction to Computational Science: Modeling and Simulation for the Sciences” (2nd Ed., Princeton U. Press, https://lnkd.in/eKw4SNYg) to help students comprehend and exploit essential concepts of computational science, the modeling process, computer simulations, and scientific applications. While the text contains about 650 project assignments, we continue to develop modules that provide the necessary scientific background and additional projects appropriate for student teams or individuals to develop. These project modules are publicly available online at https://lnkd.in/eJS5yQse. Modules can be downloaded as a zipped file or accessed individually by suggested modeling technique—system dynamics modeling, agent-based simulations, empirical modeling, cellular automaton simulations, and modeling with matrices. #ComputationalScience

1

🌟 Exciting News! 🌟 We're thrilled to announce the launch of #NNSOM, an innovative open-source Python package that brings the power of Self-Organizing Maps (SOMs) to the forefront of data science. SOMs are an incredible tool for feature extraction, visualization, and clustering complex datasets. And with #NNSOM, we're making advanced neural network techniques accessible to a diverse range of applications, from market analysis to bioinformatics. NNSOM is user-friendly and offers a robust framework for efficient training, visualization, and analysis of SOMs. The package includes customizable grid configurations and training parameters, advanced normalization and clustering tools, detailed error analysis capabilities, and extensive visualization features for deeper insights into data. We invite you to explore the potential of SOMs with #NNSOM and look forward to your feedback and collaboration as we continue to enhance it. Check out our documentation and user guides for more details: Python Package: https://lnkd.in/eVVd4Cwg Documentation: https://lnkd.in/eqXigh8d GitHub Repository: https://lnkd.in/egbbKGFi Let's push the boundaries of what we can achieve with neural networks together! 🚀 #OpenSource #DataScience #MachineLearning #AI #NeuralNetworks #Python #TechForGood #Innovation Ei Tanaka

37

4 Comments

Opinion: The "Netflix Prize" for Biology-Why Collaboration Will Build the Best AI Models The AI race to transforming drug discovery is underway with AI companies forging exciting partnerships with big Pharma players. Despite these, predicting the true impact AI will have on R&D productivity remains a challenge. The life sciences industry has traditionally been plagued by siloed data which in my opinion has done more harm than good to our collective dreams of innovating and providing life saving medicines. To solve this pressing issue, we need a bold new approach, and I propose we look to the Netflix Prize for inspiration. In 2006, Netflix challenged teams to improve its movie recommendation system by 10% to increase viewership. They offered a $1 million prize and released 100 million customer movie ratings-the first time a company that big would make such data public. That contest ignited the imaginations of thousands—scientists, engineers, hobbyists, and professionals—from over 150 countries. Funnily enough, it wasn't the mathematics that stumped the brightest minds, it was the psychology of the ever-changing tastes of viewers that proved to be most difficult. Sometime in 2009, three years after the prize was put out, competitors realized that no single algorithm could solve the complex web of human preferences. The winning team, BellKor's Pragmatic Chaos, ended up having to integrate 104 different algorithms, a superb testament to the power of collaboration in innovation. We're facing a similar moment in biology- unfortunately, we have even less understanding of biology than Netflix did of their audience. The quest to build foundational models is noble and shows incredible promise, but their potential hinges on one thing: data diversity. Just like diverse tastes impact movie choices, complex biological systems demand equally complex datasets. It sounds ambitious, but by adopting a collaborative model we can:   Redefine performance standards by integrating diverse inputs. Prevent AI from perpetuating existing health inequities. Revolutionize our approach to drug discovery through hidden patterns in vast, combined datasets. It is my opinion that the best AI models in healthcare won't come from any single company, but from a collaborative ecosystem where industry leaders share and build upon each other's work. Just like with movie recommendations, those who share data will unlock the deepest insights and accelerate innovation. The life sciences industry needs new business models that can unlock traditionally guarded datasets. It's the only way to overcome the R&D productivity crisis and fully realize the transformative potential of AI.

12

Excited to share that our paper, "Target Span Detection for Implicit Harmful Content," has been accepted as a full paper to ICTIR 2024! This research, co-authored with James Allan and Sheikh Muhammad Sarwar, tackles the challenging task of identifying implicit targets of hate speech on digital platforms. We developed the Implicit-Target-Span dataset by combining human annotations and Large Language Models (LLMs) through pooling. Our findings highlight the challenges in detecting subtler forms of hate speech, helping to develop more effective and explainable hate-speech detection models. You can read more about our study here: https://lnkd.in/e-WdAFTM #ICTIR2024 #HateSpeechDetection #AI #MachineLearning #Research #OnlineSafety

59

1 Comment

Mini update! After over a year, my UX of Well-being Chatbots paper now has a volume and issue number! No more volume 0 issue 0. Check out the paper here: https://lnkd.in/dxgJ3tSS During COVID-19, I created an SMS-based chatbot prototype to aid student well-being. Why? Because I disliked how the major commercial well-being applications required internet access. So I wanted to create something accessible without the constant need for internet. But I got curious and wanted to explore users' experience of this chatbot as well as others. This paper contains two studies. One examines the user experience of the SMS chatbot alongside a commercially available well-being chatbot. The other analysed comments from Reddit relevant to well-being chatbots. Ultimately, this paper presents emerging themes from both studies relevant to chatbot features, functionality, and responses.

53

Exciting Project Update! Thrilled to share the latest milestone in our project, Bruno—a pioneering endeavor in AI-powered mental health support! 🧠💬 Bruno isn't just an idea; it's a comprehensive project aimed at transforming mental health support through innovative data engineering solutions. From inception to implementation, our focus has been on crafting an end-to-end application that seamlessly integrates advanced technologies with real-world applications. 🌐 At the heart of Bruno lies our robust data engineering framework, meticulously designed to handle the complexities of mental health data with precision and efficiency. Through sophisticated web scraping techniques and schema extraction using LangChain and OpenAI's LLM, we've curated a rich repository of insights from diverse sources, laying the groundwork for informed decision-making and personalized support. 💡 But our journey doesn't stop there. We've leveraged performance optimization techniques, including Snowflake optimization, to ensure that data is not just stored but transformed into actionable insights with unparalleled speed and accuracy. By orchestrating data pipelines using Dagster, we've streamlined the process of data consolidation, empowering users with timely and relevant support. 🛠️ From gathering insights to empowering users through appointment scheduling and beyond, Bruno exemplifies the transformative potential of data engineering in addressing real-world challenges. With a comprehensive Business Analytics dashboard providing administrators with valuable insights, Bruno isn't just a project—it's a holistic solution to mental health support, bridging the gap between technology and human empathy. 💼 Join us on this exciting journey as we continue to push the boundaries of what's possible, harnessing the power of data engineering to create meaningful impact in the world of mental health support. Together, let's build a future where everyone has access to compassionate and personalized care, one data point at a time. 🌟✨ I give thanks to Professor Kishore Aradhya for giving us the space and guidance to make this a reality and Jared Videlefsky for providing clarity when we were stuck. A special thanks to Jane Urban for recognizing the scope of our project and for the constructive feedback. Link to application demo : https://lnkd.in/ezakwk-Y Link to our public repo: https://lnkd.in/er8y6927 Medium Article: https://lnkd.in/eS4N_ph8 #BrunoProject #DataEngineering #MentalHealthSupport #Innovation #TechForGood #EndToEndApplication

23

1 Comment