DLDC 2023

Generative AI is revolutionizing the finance industry by transforming investment strategies and risk assessment. In this session, we will explore the applications and impact of Generative AI in finance. We will discuss how Generative AI models are being used to optimize portfolio management, automate investment strategies, and enhance risk assessment. Furthermore, we will delve into specific use cases, such as fraud detection and anomaly detection, where Generative AI is revolutionizing traditional financial practices. Join us to discover how Generative AI is reshaping the financial landscape and enabling more accurate, efficient, and data-driven decision-making processes.

This talk explores the integration of computer vision, IoT, and large language models (LLMs) to enhance roller coaster safety while preserving the customer experience. Leveraging deep learning techniques, sensor data, camera feeds, and LLM-derived insights are analyzed for real-time detection of safety incidents and prediction of risks. The fusion of computer vision and IoT enables comprehensive monitoring, while LLMs extract meaningful information from textual data. Striking a balance between safety and excitement, this research aims to revolutionize roller coaster safety standards through data-driven approaches. The goal is to ensure an enjoyable and secure experience for riders worldwide.

In this session, we will explore the application of advanced language models in the field of cheque processing. Discover how these models combine natural This session aims at exploring LLM’s language processing and computer vision techniques to extract valuable information from cheques efficiently and accurately. This session will cover the significance of multi-modal data analysis and its impact on automating cheque extraction processes. Join us as we explore the capabilities, benefits, and real-world applications of large language models in revolutionizing cheque processing.

This research paper explores the ad tech industry’s evolution in digital advertising campaigns. It proposes an advertiser-agnostic deep learning solution to recommend relevant YouTube channels, enhancing targeting and campaign effectiveness in the ad tech industry.

This research paper explores the use of deep learning models to accurately predict the shelf life of fruits, focusing specifically on bananas. The study compares two object detection algorithms, Faster R-CNN and You Only Look Once (YOLO), to determine their effectiveness in predicting banana maturity based on shelf life and appearance. The authors created a dataset of banana images representing different stages of ripeness and applied preprocessing and augmentation techniques to improve accuracy. The results demonstrate the potential of deep learning algorithms for predicting fruit shelf life, with implications for other types of fruits as well.

This research paper introduces a cross-attention-based transformer model that utilizes multi-modal data (such as heat signature and acoustic signals) to detect faults in mechanical systems. The proposed model shows improved accuracy in fault diagnosis compared to traditional methods by leveraging the correlations between different modalities and employing a two-stage classification process. The results demonstrate the potential of multi-modal vibration analysis for accurate mechanical fault detection.

The world of artificial intelligence is constantly evolving, and recent developments in generative AI and Large Language Models (LLMs) have been nothing short of groundbreaking. However, even with these advancements, there are still challenges that need to be addressed in order to fully realize their potential. Enter LangChain, a powerful framework for building applications that take advantage of various LLMs to create innovative and game-changing solutions. In this session, we will have an overview of LangChain and explore its potential industry-relevant use cases. Join us to discover how LangChain can transform the way you approach AI and help you unlock new possibilities in your industry.

Engineering product improvements are often driven by user and engineer feedback. B2C products track event-level actions and failures to enhance performance, but dealing with the volume of unstructured failure logs can be challenging. In this discussion, we will have an introduction to a new approach that uses a language model for embedding generation, which is then clustered to group failures into themes. We will also explore methods to manage embeddings for improved performance and computation time. We will understand how this approach performs similarly to the latest language models while using less than one-tenth of the computation time.

Modern vehicles have lots of signals to collect data like engine rpm, vehicle speed, odometer etc. They use ECU and CAN protocol to communicate and store the data. This data is absolutely vital to analyze and diagnose vehicle performance. Generally, OEMs do encode this data making it unreadable for technicians. Decoding this data is very difficult and time-consuming task. We have developed deep learning based technique for decoding the CAN bus data. It utilizes interpretive convolutions of data along with sequence model for classification. This model is trained to differentiate patterns of each signal and segregate correct signal type like mapping of different CAN ids to signals (vehicle speed, odometer, rpm etc.). We have also attempted to address the massive data imbalance problem for the CAN decoding task.

The advent of LLMs such as GPT3, ChatGPT, and GPT4 has brought about a new paradigm shift in the world of AI. However, these models are currently not open-source, and users can only access them through user interfaces or APIs. This lack of access to the source code of these LLMs has impeded innovation and progress in the state-of-the-art.

Recent development of several open-access LLMs, such as LLaMA, StableLM, and BLOOM, has accelerated research, but the enormous size of these models discourages their adaptation to specific downstream tasks through training on task-specific datasets, making their usage limited in production.

To address this limitation, adapter-based parameter-efficient fine-tuning (PEFT), such as LoRA, BitFit, S-Adapter, and P-Adapter, has emerged as one of the most promising topics. PEFT only requires the fine-tuning of a few external parameters instead of the entire LLMs, achieving comparable or even better performance. Additionally, performing knowledge distillation from an ensemble of teacher LLMs to guide smaller student LLMs further boosts performance, making LLMs applicable, especially in limited-resource and high-throughput scenarios.