Hello and welcome back to the AAIP

This is the second part of my interview with Eldar Kurtic and his research on how to optimiz inference of deep neural networks.

In the first part of the interview, we focused on sparsity and how high unstructured sparsity can be achieved without loosing model accuracy on CPU's and in part on GPU's.

In this second part of the interview, we are going to focus on quantization. Quantization tries to reduce model size by finding ways to represent the model in numeric representations with less precision while retaining model performance. This means that a model that for example has been trained in a standard 32bit floating point representation is during post training quantization converted to a representation that is only using 8 bits. Reducing the model size to one forth.

We will discuss how current quantization method can be applied to quantize model weights down to 4 bits while retaining most of the models performance and why doing so with the models activation is much more tricky.

Eldar will explain how current GPU architectures, create two different type of bottlenecks. Memory bound and compute bound scenarios. Where in the case of memory bound situations, the model size causes most of the inference time to be spend in transferring model weights. Exactly in these situations, quantization has its biggest impact and reducing the models size can accelerate inference.

Enjoy.

## AAIP Community

Join our discord server and ask guest directly or discuss related topics with the community.

https://discord.gg/5Pj446VKNU

### References

Eldar Kurtic: https://www.linkedin.com/in/eldar-kurti%C4%87-77963b160/

Neural Magic: https://neuralmagic.com/

IST Austria Alistarh Group: https://ist.ac.at/en/research/alistarh-group/

Hello and welcome back to the AAIP

If you are an active Machine Learning engineer or are simply interested in Large Language models, I am sure you have seen the discussions around quantized models and all kind of new frameworks that have appeared recently and achieve astonishing inference performance of LLM's on consumer devices.

If you are curious how modern Large Language Models with their billions of parameters can run on a simple laptop or even an embedded device, than this episode is for you.

Today I am talking to Eldar Kurtic, researcher in the Alistarh group at the IST in lower Austrian and senior research engineer at the American startup Neural Magic.

Eldar's research focuses on optimizing Inference of Deep Neural Networks. On the show he is going to explain in depth show sparsity and quantization works, and how they can be applied to accelerate inference of big models, like LLM's on devices with limited resources.

Because of the length of the interview, I decided to split it into two parts.

This one, the first part, is going to focus on sparsity to reduce model size and enable faster inference by reducing the amount of memory and compute that is needed to store and run models.

The second part is going to focus on quantization as a mean to find representations of models with lower numeric precision that require less memory to store and process, while retaining accuracy.

In this first part about sparsity, Eldar will explain fundamental concepts like structured and unstructured sparsity. How and why they work and how currently we can achieve performant inference of unstructured sparsity only on CPU's and far less on GPU's.

We will discuss how to achieve crazy numbers of up to 95% unstructured sparsity while retaining the accuracy of models, but why it is difficult to leverage this under quoutes, reduction in model size, to actually accelerate model inference.

Enjoy.

## AAIP Community

Join our discord server and ask guest directly or discuss related topics with the community.

https://discord.gg/5Pj446VKNU

### References

Eldar Kurtic: https://www.linkedin.com/in/eldar-kurti%C4%87-77963b160/

Neural Magic: https://neuralmagic.com/

IST Austria Alistarh Group: https://ist.ac.at/en/research/alistarh-group/

## Summary

Today on the show I am talking to Veronika Vishnevskaia. Solution Architect at ONTEC where she specialises in building RAG based Question-Answering systems.

Veronika will provide a deep dive into all relevant steps to build a Question-Answering system. Starting from data extraction and transformation, followed by text embedding, chunking and hybrid retrieval to strategies and last but not least methods to mitigate hallucinations of LLMs during the answer creation.

## AAIP Community

Join our discord server and ask guest directly or discuss related topics with the community.

https://discord.gg/5Pj446VKNU

## TOC

00:00:00 Beginning

00:03:33 Guest Introduction

00:08:51 Building Q/A Systems for businesses

00:16:27 RAG: Data extraction & pre-processing

00:26:08 RAG: Chunking & Embedding

00:36:13 RAG: Information Retrieval

00:48:59 Hallucinations

01:02:21 Future RAG systems

## Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: Software that Saves the Planet: The Future of Energy Begins Here - https://belichberg.com/

### References

Veronika Vishnevskaia - https://www.linkedin.com/in/veronika-vishnevskaia/

Ontec - www.ontec.at

Review Hallucination Mitigation Techniques: https://arxiv.org/pdf/2401.01313.pdf

Aleph-Alpha: https://aleph-alpha.com/de/technologie/

# Summary

Today on the show I am talking to Manuel Reinsperger, Cybersecurity Expert and Penetration Tester. Manuel will provide us an introduction into the topic of Machine Learning Security with an emphasis on Chatbot and Large Language Model security.

We are going to discuss topics like AI Red Teaming that focuses on identifying and testing AI systems within an holistic approach for system security. Another major theme of the episode are different Attack Scenarios against Chatbots and Agent systems.

Manuel will explain to use, what Jailsbreak are and methods to exfiltrate information and cause harm through direct and indirect prompt injection.

Machine Learning security is a topic I am specially interested in and I hope you are going to enjoy this episode and find it useful.

## AAIP Community

Join our discord server and ask guest directly or discuss related topics with the community.

https://discord.gg/5Pj446VKNU

## TOC

00:00:00 Beginning

00:02:05 Guest Introduction

00:05:16 What is ML Security and how does it differ from Cybersecurity?

00:25:56 Attacking chatbot systems

00:41:12 Attacking RAGs with Indirect prompt injection

00:54:43 Outlook on LLM security

## Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: Software that Saves the Planet: The Future of Energy Begins Here - https://belichberg.com/

## References

Manuel Reinsperger - https://manuel.reinsperger.org/

Test your prompt hacking skills: https://gandalf.lakera.ai/

Hacking Bing Chat: https://betterprogramming.pub/the-dark-side-of-llms-we-need-to-rethinInjectGPT: k-large-language-models-now-6212aca0581a

AI-Attack Surface: https://danielmiessler.com/blog/the-ai-attack-surface-map-v1-0/

https://blog.luitjes.it/posts/injectgpt-most-polite-exploit-ever/

https://github.com/jiep/offensive-ai-compilation

AI Security Reference List: https://github.com/DeepSpaceHarbor/Awesome-AI-Security

Prompt Injection into GPT: https://kai-greshake.de/posts/puzzle-22745/

## Summary

At the end of last year, the EU-AI Act was finalized and it spawned many discussions and a lot of doubts about the future of European AI companies.

Today on the show Peter Jeitschko, founder of JetHire an AI based recruiting platform that uses Large Language models to help recruiters find and work with candidates, talks about this perspective on the AI-Act.

We talk about the impact of the EU AI-Act on their platform, and how it falls into a high-risk use-case under the new regulation. Peter describes how the AI-Act forced them to create their company in the US and what he believes are the downsides of the EU regulation.

He describes his experience, that the EU regulations hinder innovation in Austria and Europe and how it increases legal costs and uncertainty, resulting in decision makers shying away in building and applying modern AI systems.

I think this episode is valuable for decision makers and founders of AI companies, that are affected by the upcoming AI Act and struggle to make sense of it.

## AAIP Community

Join our discord server and ask guest directly or discuss related topics with the community.

https://discord.gg/5Pj446VKNU

## TOC

00:00:00 Beginning

00:03:09 Guest Introduction

00:04:45 A founders perspective on the AI Act

00:13:45 JetHire - A recruiting platform affected the the AI Act

00:19:58 Achieving regulatory goals with good engineering

00:35:22 The mismatch between regulations and real world applications

00:48:12 European regulations vs. global AI services

## Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: Software that Saves the Planet: The Future of Energy Begins Here - https://belichberg.com/

## References

Peter Jeitschko - https://www.linkedin.com/in/pjeitschko/

Peter Jeitschko - https://peterjeitschko.com/

JetHire - https://jethire.ai/

https://www.holisticai.com/blog/requirements-for-high-risk-ai-applications-overview-of-regulations

# Summary

For the last two years AI has been flooded with news about LLMs and their successes, but how many companies are actually making use of them in their products and services?

Today on the show I am talking to Markus Keiblinger, Managing partner of Texterous. A startup that focus on building custom LLM Solutions to help companies automate their business.

Markus will tell us about his experience when talking and working with companies building such LLM focused solutions.

Telling us about the expectations companies have on the capabilities of LLMs, as well on what companies need to have in order to be successfully implementing LLM projects.

We will discuss how Textorous has successfully focused on Retriever Augmented Generation (RAG) use cases.

RAGs is a mechanism that makes it possible to provide information to an LLM in a controlled menner, so the LLM can answer questions or follow instructions making use of that information. This enables companies to make use of their data to solve problems with LLMs, without having to train or even fine-tune models. On the show, Markus will tell us of one of these RAG projects and we will contrast building a RAG system based on Service Provider offerings like OpenAI or self hosted open source alternatives.

Last but not least, we talk about new use cases emerging with multi-modal Models, and the long term perspective that exists for custom LLM Solutions Providers like them in focusing on building integrated solutions.

## AAIP Community

Join our discord server and ask guest directly or discuss related topics with the community.

https://discord.gg/5Pj446VKNU

## TOC

00:00:00 Beginning

00:03:31 Guest Introduction

00:06:40 Challenges of applying AI in medical applications

00:17:56 Homogeneous Ensemble Methods

00:25:50 Combining base model predictions

00:40:14 Composing Ensembles

00:52:24 Explainability of Ensemble Methods

## Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: Software that Saves the Planet: The Future of Energy Begins Here - https://belichberg.com/

### References

- Markus Keiblinger: https://www.linkedin.com/in/markus-keiblinger

- Texterous: https://texterous.com

- Book: Conversations Plato Never Captured - but an AI did: https://www.amazon.de/Conversations-Plato-Never-Captured-but/dp/B0BPVS9H9R/

## Summary

Hello and welcome back to the Austrian Artificial Intelligence Podcast in 2024.

With this episode we start into the third year of the podcast. I am very happy to see that the number of listeners has been growing steadily since the beginning and I want to thank you dear listeners for coming back to the podcast and sharing it with your friends.

Gabriel is a Bioinformatician at the Austrian Institute of Technology and is going to explain his work on ensemble methods and their application in the medical domain.

For those not familiar with the term, an Ensemble is a combination of individual base models that are combined with the goal to outperform each individual model.

So the basic idea is, that one combines multiple models that each have their strength and weaknesses into a single ensemble that in the best case has all the strengths without the weaknesses.

We have seen one type of ensemble methods in the past. These where homogeneous ensemble methods like federated learning, where one trains the same algorithm multiple times by multiple parties or different subsets of the data, for performance reasons or in order to combine model weights without sharing the training data.

Today, Gabriel will talk about heterogeneous ensembles that are a combination of different models types and their usage in medical applications. He will explain how one can use them to increase the robustness and the accuracy of predictions. We will discuss how to select and create compositions of models, as well how to combine the different predictions of the individual base models in smart ways that improve their accuracy over simply methods like averaging over majority voting.

## AAIP Community

Join our discord server and ask guest directly or discuss related topics with the community.

https://discord.gg/5Pj446VKNU

## TOC

00:00:00 Beginning

00:03:31 Guest Introduction

00:06:40 Challenges of applying AI in medical applications

00:17:56 Homogeneous Ensemble Methods

00:25:50 Combining base model predictions

00:40:14 Composing Ensembles

00:45:57 Explainability of Ensemble Methods

## Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: Software that Saves the Planet: The Future of Energy Begins Here - https://belichberg.com/

## References

Gabriel Alexander Vignolle - https://www.linkedin.com/in/gabriel-alexander-vignolle-385b141b6/

Publications - https://publications.ait.ac.at/en/persons/gabriel.vignolle

Molecular Diagnostics - https://molecular-diagnostics.ait.ac.at/

# Summary

I am sure that most of you are familiar with the training paradigm of supervised and unsupervised learning. Where in the case of supervised learning one has a label for each training datapoint and in the unsupervised situation there are no labels.

Although there can be exceptions, everyone is well advise to perform supervised training when ever possible. But where to get those labels for your training data if traditional labeling strategies, like manual annotations are not possible?

Well often you might not have perfect labels for your data, but you have some idea what those labels might be.

And this, my dear listener is exactly the are of weak supervision.

Today on the show I am talking to Andreas Stephan who is doing is PhD in Natural Language Processing at the University of Vienna in the Digital Text Sciences group led by Professor Benjamin Roth.

Andreas will explain about his recent research in the area of weak supervision as well how Large Language Models can be used as weak supervision sources for image classification tasks.

# TOC

00:00:00 Beginning

00:01:38 Weak supervision a short introduction (by me)

00:04:17 Guest Introduction

00:08:48 What is weak supervision?

00:16:02 Paper: SepLL: Separating Latent Class Labels from Weak Supervision Noise

00:26:28 Benefits of priors to guide model training

00:29:38 Data quality & Data Quantity in training foundation models

00:36:10 Using LLM's for weak supervision

00:46:51 Future of weak supervision research

# Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: We do digital transformations as your innovation partner - https://belichberg.com/

# References

- Andreas Stephan - https://andst.github.io/

- Stephan et al. "SepLL: Separating Latent Class Labels from Weak Supervision Noise" (2022) - https://arxiv.org/pdf/2210.13898.pdf

- Gunasekar et al. "Textbooks are all you need" (2023) - https://arxiv.org/abs/2306.11644

- Introduction into weak supervision: https://dawn.cs.stanford.edu/2017/07/16/weak-supervision/

## Summary

I am sure my dear listener, you have heard about genAI that is driven by gigantic foundations models like GPT4 or Stable Diffusion. Generating Texts, Images and even Videos. But what about the 3D Space? What about 3D models that are used as digital twins for the metaverse, for digital cities that are required for self-driving cars to navigate safely in an metropolitan area?

What does 3D genAI look like today?

Well to give us an insight into what is possible today and what they are building for tomorrow, I have Cristian Duguet on the show. Co-Founder of Nuvo. A company that is focused on building the 3D Foundation models of the future.

On the show Cristian will explain to use some of the limitations of current 3D reconstructions from 2D images or sensors. Like issues with shiny surfaces or transparent materials, but as well, how we are currently lacking a good 3D representation for future foundation models.

We will discuss how one at the moment has to choose between two conflicting representations. 3D point clouds and 3D meshes. Where point clouds are easy to create from real world object with sensors like the newest IPhone, but are required in high density and therefore compute in order to generate photo realistic renderings.

Where 3D meshes on the other hand, made up of polygons in 3D space, need far less points and provide a very easy way to modify and change 3D objects, but are difficult to generate out of sensor data.

Cristian explains what is needed to build future foundation models and how they will be able to combine the benefits of point clouds and 3D meshes, as well how Nuvo are working towards building those next generation 3D foundation models.

---

## TOC

00:00:00 Beginning

00:02:27 Guest Introduction

00:06:40 Nuvo 3D

00:17:56 What motivates the demand in 3D models

00:25:50 GenAI for 3D present and future

00:40:14 How Nuvo wants to build 3D foundation models

00:52:24 How will future foundation models look like

## Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: We do digital transformations as your innovation partner - https://belichberg.com/

## References

Cristian Duguet - https://www.linkedin.com/in/cristianduguet/

## Summary

I have a awful memory, but its good enough most of the time, so I can remember where I left my coffee mug or when I am searching for it, where I have looked before. Imagine a person that has no recollection of what happened in their past. They might be running between room A and room B trying to find their coffee mug for ever, not realising they put it in the dishwasher.

What this person is lacking, is an episodic memory. A recollection of their, personal, previous experiences. Without them, they can only rely on what they observe and think about the world at the present moment.

Today on the Austrian Artificial Intelligence Podcast, I am talking to Fabian Paischer, PhD Student at the JKU in Linz and the ELISA PhD Program. Fabian is going to explain his research, developing an episodic memory system for reinforcement learning agents.

We will discuss his Semantic HELM paper in which they have been using pre-trained CLIP and LLM models to build an agents biography that serves the agent as an episodic memory.

How pre-trained foundation models help to build representations that generalize Reinforcement learning systems and help to understand and navigate in new environments.

This agent biography serves as a great help for the agent to solve specific memory related tasks, but in addition provides ways to interpret an agents behavior and thinking process.

I hope you enjoy this very interesting episode about current Reinforcement learning research.

## TOC

00:00:00 Beginning

00:02:08 Guest Introduction

00:07:15 Natural Language and Abstraction

00:10:37 European Ellis PhD Program

00:13:14 Episodic Memory in Reinforcement Learning

00:18:35 Symbolic State representation & Episodic Memory

00:27:04 Pre-trained Models for scene presentation

00:36:25 Semantic Helm Paper & Agent Interpretability

00:45:47 Improvements and Future research

## Sponsors

- Quantics: Supply Chain Planning for the new normal - the never normal - https://quantics.io/

- Belichberg GmbH: We do digital transformations as your innovation partner - https://belichberg.com/

## References

Fabian Paischer: https://www.jku.at/en/institute-for-machine-learning/about-us/team/fabian-paischer-msc/

Ellis PhD Program: https://ellis.eu/

SHELM Paper: https://arxiv.org/abs/2306.09312

HELM Paper: https://arxiv.org/abs/2205.12258

CLIP Explained: https://akgeni.medium.com/understanding-openai-clip-its-applications-452bd214e226