DLDC 2024

Exploring whether Generative Adversarial Networks (GANs) are adequate for handling expanding time series data. In fields like money, medical services, and natural examination, time series data augmentation assumes a crucial part in upgrading the perceivability of different artificial intelligence models. To grow the preparation dataset and work on model guess, this study utilizes GANs to create manufactured time series data that intently looks like the attributes of affirmed data. Generally, humble series datasets will probably involve Significant Learning calculations for data augmentation.

Assessments of Generative Adversarial Networks (GANs) for use in expanding time series data certainly stand out in the writing. We depict and dissect the consequences of a pilot study that expands another assessment of two classes of time series data augmentation strategies (i.e., change based systems and model mixing procedures), and we propose headings for future exploration in this squeezing field.

In this paper, we present a methodology for converting unstructured text into a structured question- and-answer format, specifically targeting 11 Indian languages. The scarcity of question-and-answer datasets for these languages poses a significant challenge for fine-tuning Large Language Models (LLMs) for specific tasks. We employed a ternary quantized model based on the LLaMA-2 architecture to achieve this conversion efficiently. Our model, BitNet 1.58, leverages a unique computation paradigm, reducing memory consumption and enhancing computational efficiency. The dataset was created in the Alpaca format and trained on 47 million data tokens over 3 epochs. Evaluation challenges were addressed using Grice’s Maxims and AI-assisted evaluation techniques. The research demonstrates significant potential for improving data quality and expanding usability across more languages.

As the adoption of Generative AI accelerates, the focus has primarily been on the cost of infrastructure and token usage. However, the challenges extend far beyond these concerns. This session delves into the critical role of XOps—DevOps, DataOps, DataScienceOps, MLOps, and LLMOps—in addressing the broader operational complexities of deploying and managing Generative AI systems.

In today’s landscape, where models are becoming increasingly sophisticated and data-hungry, the need for robust XOps frameworks is paramount. These frameworks enable seamless integration across development, data management, and machine learning operations, ensuring that AI models are not only scalable and reliable but also secure and compliant. We’ll explore how XOps facilitates the automation of workflows, enhances model governance, and supports continuous improvement, all while mitigating risks associated with model drift, data quality, and ethical AI usage.

By extending the conversation beyond infrastructure costs, this session aims to provide developers with a comprehensive understanding of the operational ecosystem required to successfully deploy Generative AI at scale. Attendees will gain insights into best practices and strategies for implementing XOps, ultimately driving more efficient and sustainable AI operations in their organizations.

In recent years, Large Language Models (LLMs) like GPT-4 and its successors have demonstrated remarkable capabilities in generating human-like text across a wide range of applications. However, a significant challenge in deploying these models is their inherent unpredictability and lack of control over generated content. This session will explore the latest research in controllable text generation, where the goal is to guide LLMs to produce outputs that adhere to specific constraints, stylistic requirements, or content guidelines.

The talk will delve into various techniques that have been proposed to achieve controllability, such as reinforcement learning, prompt engineering, and the use of control tokens. We will also discuss the trade-offs between flexibility and control, and how these techniques can be applied in real-world scenarios like content creation, bias mitigation, and user interaction. Additionally, the session will highlight emerging trends in the field, including the integration of external knowledge sources and multi-modal inputs to further enhance the controllability of LLMs.

Attendees will gain insights into the state-of-the-art methods for harnessing the power of generative AI while maintaining the ability to direct the output in meaningful and practical ways. This discussion will be particularly relevant for researchers, developers, and practitioners looking to apply LLMs in contexts where precision and adherence to guidelines are crucial.

In the evolving landscape of financial crime and regulatory compliance, the financial institutions face increasing pressure to detect and report suspicious activities effectively. The Suspicious Activity Reporting (SAR) Assistant solution leverages advanced Generative AI capabilities to enhance and expedite the investigation of suspicious activities in financial systems. This solution integrates natural language processing (NLP) and machine learning (ML) techniques to automate and streamline the SAR process, ensuring timely and accurate reporting. The Generative AI-driven SAR Assistant utilizes extensive training on diverse datasets, enabling it to recognize complex patterns and anomalies indicative of fraudulent activities. By stitching together domain expertise and Generative AI techniques, the solution can generate detailed, contextually relevant narratives and summaries for SARs. This automation streamlines the SAR process, allowing compliance officers to focus on higher-level investigations and decision-making.

This paper aims at a) unveiling the transformative role of Generative AI in the investigation of financial crimes, and b) detailing out the key components of the case investigation system including automated data processing, conversion into data embeddings, vector database set up, information retrieval, case summarization and the eventual AI-based recommendation.

This research addresses a critical business challenge faced by commodity trading companies in buying and selling commodities from international markets. Typically, trades can happen in two ways: fixed bid basis or index-linked basis. We explore the use of Large Language Models (LLMs) for time series forecasting to develop a swap-based hedging strategy that aids in deciding between fixed bid and index-linked purchases. Our approach integrates LLM-driven forecasts with swap instruments to create a hedging strategy that mitigates market volatility risks. Through comprehensive empirical analysis and back-testing, we demonstrate the efficacy of LLMs in generating long-term forecasts, thereby enhancing traditional hedging methods. This study also covers various experiments evaluating five LLMs and the finalized ensemble approach that combines traditional ML models with LLMs, providing a robust framework for effective risk management in commodity trading.

This research aims to enhance large-scale language models (LLMs) by leveraging human preferences through conditional reinforcement learning. It addresses shortcomings in current LLMs affecting authenticity, security, and user engagement. By integrating human preferences and structured decision explanations, the study aims to mitigate discrepancies between LLM outputs and human responses, advancing AI towards more accurate and contextually aware communication models.

Learn why and how data scientists and business professionals can collaborate effectively to maximize the value of data and AI projects. Discover why cross-functional collaboration is essential for successful data projects, how you can overcome common challenges in aligning technical and business perspectives and bridge the communication gap between data teams and stakeholders. This session will equip you with strategies to enhance cross-functional partnerships and drive data-informed success.

This talk explores the integration of human feedback loops into generative AI models, emphasizing how such feedback can refine outputs and steer the learning process more effectively. It examines research findings on the benefits of incorporating human insights, aiming to enhance the accuracy and relevance of generative AI systems through targeted guidance.

This talk delves into the intricacies of understanding generative AI models, highlighting the challenges in deciphering how these models produce their outputs. It underscores the necessity for developing interpretable and transparent AI systems to enhance trust and accountability. Emphasizing the critical need for explainability, the discussion aims to bridge the gap between AI complexity and user comprehension.