Schedules

This research addresses challenges in taxpayer risk assessment faced by financial departments worldwide. Outdated algorithms and a lack of standardized metrics hinder accurate identification of risky taxpayers. To combat this, the study introduces an innovative approach utilizing Large Language Models (LLMs). It involves integrating taxpayer data into templates to create comprehensive natural-language profiles, which are then used to fine-tune LLMs for precise risk predictions. Comparative evaluations show superior performance over traditional methods, revealing nuanced insights often missed by older algorithms. This approach not only enhances accuracy but also deepens understanding of taxpayer behavior, aiding informed decision-making. Embracing LLMs promises improved fiscal governance amid evolving financial landscapes, highlighting the necessity of modernizing taxpayer risk assessment methods in governmental financial departments.

In today's rapidly evolving technological landscape, Generative AI stands at a pivotal crossroads, much like the early days of the internet. While the technology is undeniably powerful, its concrete business applications remain somewhat elusive. This talk delves into the groundbreaking capabilities of Google's cutting-edge Gemini model. We will explore innovation frameworks and strategies for applying these capabilities to revolutionize your business. By leveraging Generative AI, you can forge unparalleled market differentiation, craft novel customer experiences, and unlock a wellspring of untapped opportunities.

Traditionally retailers employ a strategic integration of digital screens and printed media in hypermarkets to captivate customers, convey brand messaging, and increase sales of their products. During these media campaigns, vast amounts of product transaction data are recorded that require extensive analysis, comparison, and the ability to quickly export specific data for non-technical media planners to be able to visualize, understand, and plan media campaigns more effectively. This paper introduces an innovative approach to building a chatbot interface for the transformation of natural language into SQL queries by utilizing the large language model NSQL 350M, which can be used to perform select operations on databases to retrieve and analyze specific data. This enables media planners to ask the chatbot any query about their historical campaign data in English, and the chatbot can translate that into an SQL Query which is executed on the database, thereby retrieving the necessary information. The paper emphasizes the process of prompt engineering and finetuning the language model to ensure its accuracy is up to the mark and language model hallucination is minimal, and it highlights the potential of the chatbot in several applications for retail media campaigns.

In this session you will learn how you can leverage Generative AI on Snowflake for your business using your enterprise data and unlock value. Explore how we are empowering organisations globally to move from a POC mindset to an enterprise solutioning mindset with GenAI.

CLIP (Contrastive Language-Image Pre-training) excels in zero-shot image classification across diverse domains, making it an ideal candidate for pre-labelling unlabelled datasets. This paper introduces three pivotal enhancements designed to elevate CLIP-based pre-labeling efficacy without the need for labeled data. First, we introduce prompt refinement using a large language model (GPT-3.5- Turbo) to generate more descriptive prompts, significantly boosting accuracy on various datasets. Second, we address overconfident predictions through confidence calibration, achieving improved results without the need for a separate labeled validation set. Lastly, we leverage the inductive biases of CLIP and DINOv2 through ensembling, demonstrating a substantial boost in zero-shot labeling accuracy. Experimental results across various datasets consistently demonstrate enhanced performance, particularly in handling ambiguous classes. This work not only addresses limitations in CLIP but also provides valuable insights for advancing multimodal models in real-world applications.

A peek into how traditional sales is reinventing its approach using custom LLMs, RAG, and Autonomous agents. We'll talk about technical challenges in various domains and the approaches that we're taking to solve them.

The session emphasizes the strategic alignment of competency skill sets with the ever-changing technology landscape. The synopsis underscores the importance of staying updated by learning the latest advancements, including LLM (Large Language Models) and VLM (Vision Language Models) with Generative AI Evolution. This approach aims to cultivate competencies that effectively tackle contemporary challenges, ensuring professionals stay pertinent in the swiftly evolving digital era. As individuals, the key focus is on intelligently navigating dynamic career paths, remaining at the forefront of innovation, and adeptly contributing solutions to the intricate challenges present in today's professional landscape.

AgTech merges Agriculture and Technology, harnessing AI and ML to revolutionize farming. Scarce comprehensive data impedes its full potential. Our focus starts with a limited image set, crafting a diverse dataset mirroring real-world complexity: weather shifts, soil variations, and lighting dynamics. Employing LoRA fine-tunes a stable diffusion model swiftly. DinoV2 enhances segmentation models by integrating LoRA-trained data. Synthetic datasets bolster model performance, evident in AgTech scenarios. This research introduces LoRA, Stable Diffusion, and DinoV2, addressing limited data challenges. These methods elevate model performance, emphasizing the cost-effectiveness of synthetic datasets. Beyond AgTech, this blueprint for synthetic datasets optimizes ML in diverse domains, paving the way for large vision models.

The talk is a forward-looking dialogue that illuminates how SingleStore, a high-performance scalable SQL database, is at the forefront of this transformative wave, enabling organizations to harness the full potential of Generative AI. We will delve into the complex tapestry of challenges and opportunities that organizations face in today's need-for-speed demands. With the advent of GenAI applications, the need for a robust, versatile, and high-speed database is more evident than ever. SingleStore, with its unparalleled real-time data analytics capabilities and support for diverse data models, stands as a crux of innovation, powering the GenAI future backed by the Global Fortune 500 & Unicorns. Discover how SingleStore isn't just keeping up with the pace of change – it's setting the pace, one real-time insight and AI-driven decision at a time. Key Highlights of the Talk: Fueling Generative AI: Learn about the critical role of SingleStore in powering generative AI applications. A powerful combination of vector storage with full-text search, hybrid search, and multiple vector indexes that paves the way for groundbreaking AI-driven solutions. Real-Time Data Architecture: Discover how SingleStore's real-time data analytics revolutionize decision-making processes, enabling businesses to instantly act on insights gleaned from vast datasets. Multi-Modal Database Prowess: Imagine SQL+JSON. SingleStore's multi-model capabilities embrace relational, key-value, document, geospatial, time-series data, and full-text search, simplifying complex data architectures and significantly reducing operational complexity and costs. Case Studies in Transformation: From real-time fraud detection to powering data-driven innovation in fintech, witness real-world cases and scenarios powered by SingleStore across multiple industries and applications.

The Indian Energy Exchange (IEX) serves as India's energy nucleus, enabling electricity, renewable energy, and certificate trading. Expanding into cross-border electricity trade, it forges a cohesive South Asian Power Market. IEX's success hinges on user-focused technology, simplifying price discovery and procurement. This paper unveils a tool harnessing IEX's pricing data to forecast electricity prices over seven days, refining predictions through historical analysis. Integrated with Generative AI (GenAI), the tool not only forecasts but also generates insightful reports via natural language, empowering informed energy procurement. This initiative aims to enhance decision-making and market transparency, propelling a data-driven, efficient energy market in India and potentially across South Asia.

Personally Identifiable Information (PII) detection is critical due to the increasing exploitation of individual data, particularly in the text analytics domain. With the rise in the application of large language models (LLMs) for Natural Language Processing (NLP) solutions, data security concerns call for effective on-premises solutions and privacy-centric methods. This paper explores the use of LLMs fine-tuned on limited domain-specific datasets for detecting and masking PII and benchmarking this solution against existing NLP methods such as BERT and GPT3.5. Our approach includes fine-tuning the Vicuna-7B LLM using the Quantized and Low Rank Adaptation (QLoRA) technique, enabling cost-effective fine-tuning and deployment on consumer GPUs; The proposed approach offers several advantages, including improved performance and reliability compared to GPT3.5, enhanced data security by keeping data within the company's cloud, domain adaptability through model fine-tuning, and on-premise usage benefits such as reduced dependence on proprietary models, quota limitations, and flexible scaling of model hosting infrastructure. Overall, this paper presents an efficient and secure solution for domain specific PII detection tasks using LLMs.

This paper navigates challenges faced by foundational language models, particularly combating Hallucinations in models like GPT. Detailing a methodical strategy, it regulates GPT's responses by tapping into knowledge solely from PDF, HTML, and Word Doc sources. The approach, integral to a regulatory Query Bot application, prioritizes hallucination-free responses. Leveraging Retriever Augmented Generation, unique prompt engineering, and a scoring mechanism for context passages, it ensures accurate responses despite GPT's innate limitations. This solution significantly reduces turnaround time by 150%, introduces Query Status Tracking, and enhances User Experience, marking a pivotal advancement in controlling misinformation while harnessing the potential of large language models.

We are witnessing unprecedented innovation in AI. How to adapt your role as a Data Science professional. What core skills to take forward and how to stay well-versed with new developments?

This research paper innovates traditional market survey methods by integrating advanced Generative Pre-trained Transformers (GPTs) and ensemble Language Models (LLMs). Emphasizing efficient synthetic data generation while upholding stringent privacy and ethical standards, our methodology utilizes these AI techniques for transparent dataset synthesis. Through real-world scenarios and detailed case studies, it demonstrates substantial efficiency gains, acknowledging limitations transparently. Prioritizing user privacy and data security, our approach complies with rigorous regulations, accelerating survey processes and fostering personalized engagement. This pioneering paradigm shift reshapes market survey practices, establishing an industry benchmark for ethical and efficient data synthesis.

In recent months, the landscape of Generative AI has undergone a transformative shift, challenging the long-held dominance of Big Tech giants. This talk delves into the rapidly innovating world of open-source LLMs, highlighting how they are reshaping the field. The latest advancements in both model development and infrastructure are not only democratizing access to generative AI but also presenting cost-effective alternatives to proprietary systems. Join us in this session, tailored for businesses looking to leverage the power of GenAI, as we unravel the intricacies of building and scaling an in-house Generative AI platform. We will compare the financial implications, performance metrics, and long-term viability of proprietary versus open-source solutions.

his paper conducts an in-depth investigation of the visual capabilities of OpenAI's advanced language model, GPT-4, focusing on image comprehension and visual content creation. It demonstrates how GPT-4 processes and translates visual information into detailed data and its proficiency in generating diverse visual content. The study further emphasizes the integration of GPT-4's visual abilities with its advanced language processing capabilities, leading to more complex AI applications. It also discusses the impact of GPT-4's capabilities in various sectors and potential ethical concerns. The paper ends with a perspective on the future implications of these AI advancements.

In the year 2023, we saw the rise of a new industrial revolution - Large Language Models (LLMs). Bringing in a step function growth in how we process information, provide support to customers, and even build applications. We saw a glimpse of what Artificial General Intelligence (AGI) could be with autonomous agents working together. Following the hype of LLMs was not AGI but RAG, Retrieval Augmented Generation. So, what is RAG? If LLM is the most powerful car engine, RAG is the wheels. You can’t reach your destination with the most powerful engine without converting the mechanical energy of the engine to the kinetic energy of the wheels as motion. With the current LLM architecture, the knowledge of the world and the understanding to process the information are embedded in the network in an intertwined manner, which means the model’s view of the world is static and not easy to refresh without training. This is a big limiting factor. Even if we choose to refresh LLM’s knowledge of the world, LLMs are not good at information recall (or retrieval) because they learn a compressed probabilistic view of the world. The fragmented information storage can’t provide retrieval. LLMs, being generative models, will always generate based on existing understanding, which leads rise to the problem of hallucination. Retrieval Augmented Generation (RAG) is a technique of providing LLMs with context relevant to user queries to help LLM ground answers and provide citations. But, building a RAG application hasn’t been easy, it is a continuous cycle of build -> learn -> improve. A lot of this is because of the paradigm shift in how we build applications. Our programming language has changed from structured programming languages to unstructured language of English. This directly impacts how we process data and build execution flows, requiring us to rethink and build new mental models. In this talk, I will take the first principles approach to break down the shift we are seeing, how we are building RAG applications, and leave the audience with thoughts to think about while building RAG application themself.

This research highlights the integration of Deep Reinforcement Learning (DRL) with Deep Deterministic Policy Gradient (DDPG) in Adaptive Cruise Control (ACC), a key part of Advanced Driver Assistance Systems (ADAS). Using a simulation environment featuring a lead car and a Battery Electric Vehicle (BEV), the DRL controller manages the BEV's velocity, addressing multi-objective ACC goals. Driving objectives like tracking accuracy, ride comfort, and safety are translated into reward functions for the DRL. Comparing it with Model Predictive Control (MPC) in highway scenarios, the DRL excels in ride comfort while matching MPC in accuracy and safety. Moreover, the DRL exhibits robustness and efficiency across various driving scenarios, showcasing superior performance and strong adaptability.

The evolution of Computer Vision has witnessed a transformative shift towards Video Analytics, enabling real-time interpretation of visual data. This evolution empowers systems to analyze and derive meaningful insights from dynamic video streams, enhancing applications in surveillance, object recognition, and automated decision-making. The integration of advanced algorithms and deep learning techniques has propelled Video Analytics to the forefront of cutting-edge visual intelligence technologies.

This talk will cover the Generative AI landscape from the AI scientist’s perspective. The playground of technologies and the range of dev-toys at the disposal of the modern-day AI scientist has exploded in a period of a few months. Generative AI technologies are being viewed mostly from the end user’s perspective in common parlance, due to the disruptive nature and the promise of productivity. However, the ‘behind the scenes’ technologies are equally fascinating, as these are being introduced to the world at a rapid speed. This is especially true after the two announcements of ChatGPT almost putting the Turing test to shame, and the AI engineering world being put on notice when MS mentioned usage of tens of thousands of A100s. There has been an incessant day-on-day announcements about newer topics like Langchain, RAG, anchoring, GANs, LLMs, VLMs, Diffusion, RLHF. Even an age-old AI topic (‘Reasoning’) coming to the forefront -- is fascinating to see. This journey will shed light on the transformative power of Generative AI, uncovering some of the challenges, the breakthroughs, and the promising horizon that lies ahead.

In the dynamic world of investment firms, the investment committee's role in evaluating potential investments is pivotal. The foundation lies in meticulously prepared investment memos by advisors. Traditionally, this demanding process involves extensive due diligence, analyzing company documents, extracting insights, and market research. This research explores Large Language Models (LLMs) as tools to enhance efficiency and accuracy. LLMs streamline data synthesis from varied sources, aiding in generating investment hypotheses, memo drafting, and precise Q&A support. Preliminary results show a twofold efficiency boost in tasks from due diligence to committee reviews and an 85% accuracy rate in responses. This signifies LLMs' potential to transform manual investment decision-making in firms.

Increasingly Al solutions have become an integral part of businesses. This has led to significant investment in AI - private investment in AI was around USD 70 billion in 2021, increased to USD 90+ in 2022 and is expected to be USD 160 billion in 2025; up 70-75%. These investments come with a glaring expectation of appropriate returns. As data scientists we are intrigued by the technological advancements and the pursuit of excellence - very rightly so. However, most often than not we forget to or ignore that eventually these solutions should be able to provide net returns. This requires a hard look at the solutions that we have built with so much love. It requires us to evaluate the economics of our solutions before we are asked for it. I have found that either teams shy away from doing this - or do not have the right methodology to evaluate. From a survey conducted in my data science communities I found that the majority of the people do not carry out such an exercise and were also not able to list all the aspects of such an evaluation. In the new world of GenAI - this has become much more relevant. A great data science team is the one that has a solid process and culture for this self scrutiny. This makes our work relevant, long lasting and trustworthy.

This session would be a journey exploring how machine learning, a powerful tool, is enhancing the field of neurotech and contributing to advancements in healthcare. It will delve into real-world applications where machine learning assists in analyzing brain scans, potentially leading to earlier diagnoses and improved treatment and surgery options. It will cover recent research and experiments utilizing machine learning models, offering us a glimpse into the potential of this technology. Additionally, it will compare and contrast how machine learning's impact on healthcare differs from its contributions in other areas, highlighting its unique potential to improve human lives. This session promises to be informative and thought-provoking, offering valuable insights into the exciting intersection of AI and neurotechnology.

In this talk, Sanyam Bhutani will share the cutting edge details of how to fine-tune your LLM effectively using open source tools. We will start from Zero, get a whole-picture overview and then learn the details hands on.

This research delves into lookalike modeling's methodology, seeking congruent traits among novel customers mirroring high-value ones. Targeting expanded market reach, it pinpoints potential customers keen on products or services. Employing a designated subset (seed-set) of high-value customers, machine learning uncovers commonalities. Leveraging job-related attributes, the model identifies similar users from vast datasets via a fast multi-graph approach, boasting efficiency. Employing a distributed architecture ensures scalability, managing 500 recommendations/second with 80 million interactions. Evaluation shows up to 90% precision enhancement for 200 Lookalike candidates in digital marketing. Further, preprocessing optimization mitigates noise issues, enabling broader model applicability across diverse datasets. This study champions collaboration and innovation, promoting an accessible open-source framework.