doXray B.V.

07/05/2024

Thrilled to share that our colleague Tin Ferković will be presenting at DSC Adria conference!
Tin will delve into model calibration, covering techniques for aligning confidence predictions with output probabilities across discriminative and generative scenarios, from text classification to advanced tasks like TrOCR.

https://www.dscadria.com/

DSC Adria is just around the corner! 🔥

Meet our next set of speakers:

🎙 Tin Ferković, an ML Research Engineer at doXray specializing in document understanding, holds a recent degree in Data Science from FER, Zagreb, and emphasizes the dual importance of technical excellence and societal impact.

He will be joining us from Croatia to deliver his talk titled: Model Calibration

Tin holds a strong educational background in Data Science, including a Master's degree from FER and a master exchange semester at RWTH Aachen University.

In his talk, Tin Ferkovic will explore the process of model calibration, discussing various techniques to align confidence predictions with output probabilities in both discriminative and generative scenarios, including simpler methods for classification problems like text classification and named entity recognition, as well as more advanced techniques like TrOCR for generative tasks.

🎙 Valentina Zadrija is an AI Tech lead at Yaak Technologies. She will be joining us from Croatia to present her talk titled: Large Multimodal Models: scaling from browsers to Embodied AI.

With over 18 years of experience and a PhD in the field of AI, Valentina has been developing AI models for autonomous mobile robots and self-driving cars, while also leading teams for their development. Her research interests include large multimodal models, their security, and their deployment beyond browsers - into smart homes, warehouses, fields and roads.

In her talk, Valentina will discuss the challenges of implementing large multimodal models for embodied AI, focusing on safety alignment and validation with human core values.

🎙 Sandro Skansi, an Associate Professor of Logic and AI at the University of Zagreb / Sveučilište u Zagrebu' is a leading figure in the field of artificial intelligence.

He will be joining us from Croatia to deliver his talk titled: A Lost Croatian AI Program from the 1950's

With two published books on AI and numerous peer-reviewed papers, Sandro has also applied his expertise to develop AI models for Croatian companies and government agencies. Holding a Ph.D. in Logic from the University of Zagreb, his contributions bridge the gap between technical AI advancements and philosophical inquiry, making him a valuable asset to both academia and industry.

In his talk, Sandro will cover the early AI landscape from 1943 to 1960, focusing on the first Zagreb circle of cybernetics in the 1950s. He'll explore their work in developing a machine translation system, discussing technical aspects and drawing parallels with modern transformers like BERT and GPT.

May 21-24| 📍 Zagreb| 🏟 Mozaik Event Centar

Tickets ➡ https://lnkd.in/dc78rnPJ

Check out our previous speakers ➡ https://lnkd.in/e3EidK9S

06/11/2023

📢AI and language enthusiasts, we have exciting news!
Introducing open-sourced project LinguaAI - Your Personalized Language Learning Companion 🗣️🌍

In our latest blog, Mihaela Bakšić, delves into prompt engineering techniques for the development of a language learning chatbot powered by GPT models with the objective of creating a highly adaptable chatbot tailored for language learning and practice via prompt engineering.

Beyond Traditional Learning: The GPT-4 Edge in Adaptive Language Tutoring for any Language and any Level.

21/08/2023

The problem with transfer learning paradigm is that it requires fine-tuning a new large language model (LLM) for each new task, which is unsustainable in terms of time, storage, and energy. ⏳💾⚡️

🚀In our latest blog post "Multi-Task Learning with Intermediate Continual Learning for Industry NLP Use Cases", Tin Ferkovic wrote about finding an efficient, yet effective, multi-task learning (MTL) method which would be able to handle learning multiple tasks using less time and storage space, eventually leading to less energy consumption, whilst preserving the performance as in single-task learning (STL). 🧠🔋

This method should be easily accessible to all small to medium-sized companies, so it should be able to run on a single or few mid-range graphics processing units (GPUs). 💼💻

Check out how Tin Ferkovic utilized adapters and hypernetworks to efficiently, effectively, and continuously train multiple tasks. 🛠️🌐🤖

Full blog post:

Utilizing adapters and hypernetworks to efficiently, effectively, and continuously train multiple tasks

08/08/2023

We're thrilled to share insights from our latest study on Document-level Entity Relation Extraction! 🧠🔍

In this cutting-edge research, we delve into the complexities of identifying and categorizing relationships between entities within whole documents, going beyond individual sentences. 📜🔗

🔥 Key Highlights:

🔹 Challenges Beyond Sentences: Document-level entity relation extraction unveils novel challenges, as entities may relate across different sections of a document. We tackle these complexities head-on to capture meaningful connections at a broader scale.
🔹 Investigating Generative Models: We explore state-of-the-art generative models to deepen our understanding of entity relations within documents. This endeavor unlocks potential for more comprehensive information extraction.
🔹 Innovative Experiments: We conduct three experiments employing various approaches, from Named Entity Recognition (NER) to explicit tagging. These experiments yield intriguing results that shed light on the effectiveness of each approach.
🔹 Power of Prompting: Prompting emerges as a powerful technique, guiding our models in generating coherent responses. We uncover how different prompts impact the model's performance, enhancing our grasp on adaptability.
🔹 Selecting the Right Model: Our study evaluates multiple models, including T5, mT5, FLAN-T5, GPT-3.5 Turbo, and GPT-4. We weigh their strengths, performance, and adaptability for accurate entity relation extraction.
🔹 Results & Implications: Through meticulous evaluation and analysis, we present compelling results, demonstrating GPT-4's prowess in capturing entity relationships. Yet, model selection nuances arise, highlighting the importance of context and fine-tuning.
🔹 Future Directions: We acknowledge limitations and pave the way for future research, aiming to refine accuracy, generalize to diverse domains, and tackle longer document challenges.
This study offers insights that ripple across various applications. Join us in embracing the dynamic realm of document-level information extraction! 🌐💼
👉 For an in-depth dive, read the full post here https://blog.doxray.com/p/exploring-state-of-the-art-models

Document-level entity relation extraction involves identifying and categorizing relationships between entities mentioned within whole documents, rather than individual sentences. Document-level entity relation extraction presents unique challenges due to complex dependencies and context beyond indiv...

28/07/2023

🏆 Introducing "Named Entity Recognition Using Question Answering in Zero- and Few-Shot Settings" 🏆

In today's fast-paced world, document processing plays a pivotal role in effective decision-making. Our research is centered around the transformative concepts of Zero- and Few-Shot Learning, where we explore the true potential of training models with minimal data. 🌐💼

🔹 Unleashing the Power of Limited Data: Dive into our blog post to uncover how Large Language Models (LLMs), like the cutting-edge Flan-T5, can excel in document analysis even with as little data as possible. 📈📝

🔹 Advancing with Pre-Training and Fine-Tuning: Discover the cost-effective paradigm that empowers your document processing models to achieve remarkable results, from masked language modeling to sequence reshuffling and beyond. 💰💡

🔹 A Leap Forward in Document Analysis: Our research sheds light on the potential of Zero- and Few-Shot Learning, drastically reducing costs, training time, and resource requirements without compromising on performance. 📊⏱️

🔹 Propel Your Business with Innovation: Stay ahead in the game by harnessing the prowess of Large Language Models and shaping the future of document processing in your organization. 💼🚀

👉 Ready to revolutionize your document processing models? Explore our blog post and embark on the journey of Zero- and Few-Shot Learning for Named Entity Recognition! 🔗📚

The human race produces copious amounts of documents daily to aid its effective functioning. Even though digitalization allows the automatic processing of documents when they are stored in an appropriate format, this is not the case for documents that are printed or hand-written. The introduction of...

03/07/2023

🔒🚀 Exciting news! Our talented expert, , has just published a must-read blog post on bolstering the security of your Kubernetes web applications. Don't miss out on this insightful guide! 🔐🌐

Discover how to seamlessly integrate into your web apps using Keycloak and Gatekeeper, without making core code modifications. From setting up Keycloak to configuring user and group settings, Bojan covers it all! 🛠️💡
Whether you're a novice or an experienced developer, this step-by-step approach empowers you to safeguard your applications with ease. Strengthen your app security and stay agile in today's connected world. 🚀🔐

Ready to dive in? Click below to access the full blog post.

In certain scenarios, you may need to integrate an authentication layer into a basic application or tool that doesn't demand complex authentication. This technical article addresses such situations, aiming to help users safeguard their applications without modifying the underlying code. To facilitat...

19/06/2023

📢 Attention all BPO companies and organizations dealing with vast amounts of unstructured documents!

Our company is proud of our game-changing product that has revolutionized data extraction from unstructured documents! 💼💪

For the past 5 years, we've been using this cutting-edge technology in production, at scale, to extract valuable insights from millions of documents. 📚🔍

💡 Unlock the power of our proven solution to effortlessly extract and organize data, saving time and resources for your business. 💻✨

Join the ranks of satisfied customers and experience the power of our industry-leading data extraction solution! 💼💪

Doxray explainer video. Inbox Use Case

02/06/2023

🎉 Thank You for Visiting DoXray at Job Fair 2023! 🙌

We had an incredible time connecting with talented students at the Job Fair 2023 conference in Zagreb. It was a pleasure sharing our vision, mission, and exciting career opportunities with all of you!

🤝 We extend our heartfelt gratitude to everyone who stopped by our booth, engaged in insightful conversations, and showed enthusiasm in learning about our innovative solutions in the field of AI!

👉 We believe in fostering a culture of growth and continuous learning, and we were thrilled to see so many talented individuals eager to make a positive impact through their work. Remember, the future belongs to those who are bold enough to embrace challenges and explore new frontiers.

🌟 If you missed us at the Job Fair, don't worry! Reach out to us through our website or social media channels to learn more about exciting career opportunities and how you can be a part of our dynamic team. We look forward to hearing from you!

🙏 Once again, thank you to all the students and organizers who made the event a resounding success. Stay tuned for more updates on our latest projects and initiatives!

18/05/2023

20/02/2019

doXray applies, tailors and customizes it’s proprietary deep learning based NLP software to customer challenges in the legal, financial, real estate and BPO verticals. The doXray software uses entity extraction, categorization, comparison and other NLP services on unstructured electronic or physical data and embeds these services into the customer’s respective workflow.