dida develops tailor-made machine learning software for clients, integrating it into existing IT infrastructure.

dida

dida is your partner for process automation and develops tailormade software applications for your business.

We automate your internal processes through AI-powered software. Our team operates at the intersection of mathematics and computer science and is specialized in Machine Vision and Natural Language Processing (NLP) projects.

dida Machine Learning

𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐈𝐧𝐬𝐢𝐠𝐡𝐭: 𝐅𝐫𝐮𝐢𝐭𝐍𝐞𝐑𝐅

Recently, we examined the paper "FruitNeRF", which investigates the use of Neural Radiance Fields (NeRF) for counting fruit on trees. NeRF is generally applied for 3D reconstructions in controlled environments, but this study explores its potential for a practical agricultural problem.

FruitNeRF uses NeRF to create detailed 3D models of tree canopies, allowing for accurate counting of individual fruits. This represents a shift from NeRF’s typical applications to a real-world problem in crop monitoring and management.

The paper demonstrates how this method can improve efficiency and accuracy in assessing crop yields and managing resources.

𝐅𝐨𝐫 𝐟𝐮𝐫𝐭𝐡𝐞𝐫 𝐝𝐞𝐭𝐚𝐢𝐥𝐬, 𝐲𝐨𝐮 𝐜𝐚𝐧 𝐫𝐞𝐚𝐝 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥 𝐩𝐚𝐩𝐞𝐫 𝐡𝐞𝐫𝐞: https://lnkd.in/eskGPyP5 

#research #AI #technology #machinelearning #ML #innovation

NVIDIA 𝐫𝐞𝐜𝐞𝐧𝐭𝐥𝐲 𝐩𝐮𝐛𝐥𝐢𝐬𝐡𝐞𝐝 𝐧𝐞𝐰 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐨𝐧 𝐂𝐨𝐦𝐩𝐚𝐜𝐭 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 𝐯𝐢𝐚 𝐏𝐫𝐮𝐧𝐢𝐧𝐠 𝐚𝐧𝐝 𝐊𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐃𝐢𝐬𝐭𝐢𝐥𝐥𝐚𝐭𝐢𝐨𝐧.

In this work, they focused on reducing the size of LLama3.1 8B by pruning 50% of its layers—specifically, they removed 16 layers, shrinking the model from 8 billion parameters to 4 billion. The layers were selected using an activation-based importance estimation strategy to ensure the most critical parts of the model remained intact.

After pruning, they retrained the 4B model using knowledge distillation, where the original 8B model served as the teacher, guiding the smaller model during retraining.

This approach led to impressive results: the model required far fewer training tokens—approximately 100 billion, a 40x reduction—and achieved a 16% improvement in MMLU scores compared to training from scratch.

𝐂𝐡𝐞𝐜𝐤 𝐨𝐮𝐭 𝐭𝐡𝐞 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐡𝐞𝐫𝐞: https://lnkd.in/e_JbyqVX

#machinelearning #ML #technology #research #news #LLM

𝐖𝐞’𝐯𝐞 𝐣𝐮𝐬𝐭 𝐬𝐞𝐧𝐭 𝐨𝐮𝐭 𝐨𝐮𝐫 𝐬𝐞𝐜𝐨𝐧𝐝 𝐧𝐞𝐰𝐬𝐥𝐞𝐭𝐭𝐞𝐫 𝐨𝐟 2024!

In this edition, we share some exciting updates, including a feature on our solar panel estimation project by Google Cloud, insights from our recent dida conference, and highlights from Lorenz Richter's presentation at ICML 2024. We also introduce a selection of newly started projects.

If you haven’t seen it yet, you can read the full newsletter on our website here: https://lnkd.in/eMbcJjwh 

Thank you for staying connected with us at dida!

#ML #machinelearning #innovation #technology #education #google #project

𝐎𝐮𝐫 𝐧𝐞𝐰 𝐛𝐥𝐨𝐠 𝐚𝐫𝐭𝐢𝐜𝐥𝐞 𝐨𝐧 𝐞𝐱𝐩𝐥𝐚𝐢𝐧𝐚𝐛𝐥𝐞 𝐀𝐈 𝐢𝐬 𝐡𝐞𝐫𝐞!

Learn about eXplainable Artificial Intelligence (XAI) and its human-oriented approach to machine learning. Instead of focusing solely on evaluation methods, we present why no single method can be considered perfect and emphasize the importance of understanding the diverse needs of different user categories.

The article highlights the importance of catering to the unique demands of various user categories and underscores the ethical implications and transparency challenges that come with AI systems. Particularly in healthcare, where the stakes are high, XAI plays a pivotal role in ensuring that both medical professionals and patients can trust and understand AI-driven decisions.

Read the full article here: https://lnkd.in/etT5GmZb 

#XAI #artificialIntelligence #machinelearning #healthcare #transparency #AI #innovation #technology #future

𝐖𝐞’𝐫𝐞 𝐞𝐱𝐜𝐢𝐭𝐞𝐝 𝐭𝐨 𝐚𝐧𝐧𝐨𝐮𝐧𝐜𝐞 𝐭𝐡𝐞 𝐤𝐢𝐜𝐤𝐨𝐟𝐟 𝐨𝐟 𝐨𝐮𝐫 𝐧𝐞𝐰 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐩𝐫𝐨𝐣𝐞𝐜𝐭, 𝐆𝐄𝐍𝐈𝐀𝐋4𝐊𝐌𝐔, 𝐢𝐧 𝐜𝐨𝐥𝐥𝐚𝐛𝐨𝐫𝐚𝐭𝐢𝐨𝐧 𝐰𝐢𝐭𝐡 University of Hamburg!

GENIAL4KMU aims to harness the power of Retrieval-Augmented Generation (RAG) systems to make unstructured company data more accessible and accurate. By focusing initially on structured data and later extending to unstructured data, we plan to implement a system that integrates relevant contexts into queries, ensuring reliable and efficient access to information.

This project not only leverages the latest in AI and language model advancements but also aims to create tailored solutions for German-language content. We’re looking forward to exploring new frontiers in AI and contributing to both academic and practical advancements in the field.

Stay tuned for updates as we progress with GENIAL4KMU!

#AI #machinelearning #research #innovation #collaboration

𝐌𝐞𝐞𝐭 𝐮𝐬 𝐚𝐭 𝐭𝐡𝐢𝐬 𝐲𝐞𝐚𝐫'𝐬 𝐈𝐂𝐌𝐋!

Our CTO Lorenz Richter is currently on his way to the [ICML] Int'l Conference on Machine Learning in Vienna. 

Lorenz will present our new paper "Bridging discrete and continuous state spaces: Exploring the Ehrenfest process in time-continuous diffusion models".

The paper advances the understanding of generative AI by studying time-continuous discrete-space diffusion models and for the first time relate them to the popular score-based generative modeling, prominently used e.g. for image generation.

Reach out to Lorenz on LinkedIn, if you are also there and would like to meet up!

Link to the paper: https://lnkd.in/dmRFagbq

𝐖𝐢𝐭𝐡 𝐬𝐦𝐚𝐫𝐭𝐞𝐱𝐭𝐫𝐚𝐜𝐭 𝐰𝐞 𝐣𝐮𝐬𝐭 𝐥𝐚𝐮𝐧𝐜𝐡𝐞𝐝 𝐨𝐮𝐫 𝐟𝐢𝐫𝐬𝐭 𝐀𝐈 𝐩𝐫𝐨𝐝𝐮𝐜𝐭. 

It's a SaaS software that allows you to extract information from complex documents or emails, helping you to automate your document processing tasks. 

The first feedback has been great so far, but we would love to get some more, and also show some demos. If you are interested, please reach out!

𝐓𝐚𝐥𝐤 𝐚𝐭 𝐝𝐚𝐭𝐚2𝐝𝐚𝐲 𝐜𝐨𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞

Join us at the Data2day conference where Jakob Scharlau, our ML Scientist, will present "Semantic Search and ML: Finding the Right Document" on September 18th and 19th in Heidelberg, Germany.

This talk introduces the application of machine learning and pre-trained language models in developing search algorithms that comprehend text semantically. Jakob will cover recent advancements, including text embeddings and the utilization of LLMs. He will provide an overview of different approaches to crafting tailored search algorithms, illustrating key methods with real-world customer projects.

𝐖𝐞 𝐥𝐨𝐨𝐤 𝐟𝐨𝐫𝐰𝐚𝐫𝐝 𝐭𝐨 𝐬𝐞𝐞𝐢𝐧𝐠 𝐲𝐨𝐮 𝐭𝐡𝐞𝐫𝐞! 

#research #technology #AI #innovation #LLM #conference

𝐓𝐚𝐧𝐠𝐞𝐫𝐞24 𝐰𝐢𝐭𝐡 𝐒𝐢𝐞𝐦𝐞𝐧𝐬 𝐆𝐌𝐒

Last week we had the pleasure of attending Tangere24, hosted by Siemens Global Marketing Services (GMS) in Berlin. Siemens GMS, known for their expertise in marketing, sales, and communication, are leaders in leveraging AI within their operations.

Our Managing Director, Philipp Jackmuth, and Machine Learning Researcher, Jan Macdonald, represented us at the event. They had the opportunity to gain valuable insights into Siemens' comprehensive AI strategies, exploring how AI is transforming their global processes.

The event featured collaborative workshops focused on developing practical solutions for real-world challenges in marketing, sales, and communication. It provided a dynamic platform for meaningful exchanges with AI experts, discussing the latest advancements and best practices. We had the opportunity to engage directly with the Siemens GMS team, gaining valuable connections and exploring the potential for joint initiatives. 

We thank Siemens Team for this insightful event and look forward to future collaborations:

Lena Serdarusic, Dr. Hannes Fleischer, Stefan Schönmann, Larissa Anna Epp, Christoph Bischoff, Keito N., Samuel Kuhn & Julian Garber

#Tangere24 #SiemensGMS #AI #innovation #collaboration #networking

𝐖𝐨𝐮𝐥𝐝 𝐲𝐨𝐮 𝐥𝐢𝐤𝐞 𝐭𝐨 𝐬𝐩𝐞𝐧𝐝 126 𝐄𝐔𝐑 𝐨𝐫 7.150 𝐄𝐔𝐑 𝐟𝐨𝐫 𝐲𝐨𝐮𝐫 𝐝𝐨𝐜𝐮𝐦𝐞𝐧𝐭 𝐩𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠 𝐰𝐢𝐭𝐡 𝐆𝐏𝐓?

Pricing of LLMs such as OpenAI’s GPTs can be hard to grasp. We therefore played it through based on the processing amounts of a medium-sized company. 

Let’s take a medium-sized company that automatically processes its documents with OCR and one of OpenAI’s GPT models: 

With for example an active sales department and regular customer interactions, this medium-sized company might process around 4,500 documents monthly, totaling about 55,000 annually, each averaging 2.6 pages. 

	◾ It could either spend 126 EUR for the overall task or up to 7.150 EUR 
	◾ GPT-3.5 is 50 times cheaper than GPT-4
	◾ GPT-4o is still 5 times cheaper than GPT-4

This calculation obviously only shows the cost side of the GPT model choice, but it shows that a) for some tasks the price aspect might be more important than the best performance and b) why OpenAI decided to publish GPT-4o as a balanced option of performance and costs.

So when selecting a GPT model, consider your specific needs. GPT-3.5 is ideal for budget-conscious applications, while GPT-4 is best for tasks requiring top-tier performance. GPT-4o strikes a balance for those needing both affordability and advanced features.

Choose wisely based on your business priorities.

#AI #LLM #GPT #businessautomation #OCR #technology #innovation

DIDA

DO:DIDA

dida