Austrian Artificial Intelligence Podcast Manuel Pasieka

## 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

# Summary

Every day you can read and hear about the impact of AI on companies of any industry and size. But are really all business at a stage where they can benefit from the wonders of AI? What about small companies that are not in the tech and dont have the budget to hire data scientists and machine learning engineers. For example, like small retailers of fast moving consumer goods; FMCG in short. that might only have a few stores in a city. They are experts in their field, but lack the personal or infrastructure to have their own AI initiatives. How can they benefit from AI to for example optimize their planning and supply chain?



Today on the show I am talking to Eric Weisz, co-founder of Circly, an AI startup that has build a self-service platform for SME's to help them with demand forecasting. Making it possible for none data scientists with little historical data to use their platform and benefit from accurate predictions.



On the show we talk about the challenges of building such a one-fits all platform that has to provide value to all kind of different customers without intensive manual configuration and tuning. We talk about how to verify and maintain data quality, and how approaches from federated machine learning can be used to ensure the effective use of prediction model. So that based on the available data and its characteristics models are selected the are efficient to run as a platform provider, reducing costs, while providing highly accurate predictions for customers.



## TOC

00:00:00 Beginning

00:02:42 Guest Introduction

00:05:98 Circly: Demand Prediction for SME's

00:08:05 Demand prediction as an SaaS offering

00:14:37 Ensuring and maintaining data quality

00:26:09 Prediction model selection based on data and efficiency

00:35:04 Federated Machine Learning & Weight sharing

00:39:58 Feature selection and context enrichment



## 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

Eric Weisz - https://www.linkedin.com/in/ericrweisz-circly-ai-retail/

Circly - https://www.circly.at/

Today on the show I am talking to Michael Trimmel, head of AI at HalloSofia about his journey as an entrepreneur, building AI Startups.

This episode will be most valuable to people that interested in creating an AI startup or at the beginning of this journey.



Michael will tell his personal startup story, describing his troubles and learnings on the way. Its particular important to him to highlighting that one can get into AI without having a traditional computer science background.



We will be talking on how to get started as an Entrepreneur, what makes a good founding team, how to build a support network, how to build first prototypes, how to benefit from accelerator program and what funding options there are in Austria.



I hope this interview will provide you with useful information and tips to get you started on your own journey.







# TOC

00:00:00 Beginning

00:02:31 Guest Introduction

00:07:17 Co-Founder of Cortecs GmbH

00:11:48 Head of AI at HelloSofia

00:20:22 What you need to build a startup

00:23:48 The founding team

00:31:03 The Business Network

00:37:36 Incubators & Accelerators

00:43:56 Funding

00:49:18 Navigating the AI Hype



# 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

Michael Trimml - https://www.linkedin.com/in/michael-trimmel-91aa3a152/

Hallo Sofia - https://www.hallosophia.com/

Startup House - https://www.startuphouse.at/

Austrian Startups - https://austrianstartups.com/

Austrian Startups - ELP - https://austrianstartups.com/elp/

Hummel Accelerator - https://hummelnest.net/

INITS - https://www.inits.at/

FFG - https://www.ffg.at/

# Summary

In our bodies, the Immune system is detecting foreign pathogens or cancer cells, called antigens, with the help of antibody proteins that detect and physically attach to the surface of those cells.

Unfortunately our immune system is not perfect and does not detect all antigens, meaning that the immune system does not have all antigens it would need to detect all cancer cells for example.

Modern cancer therapies like CAR T-cells therapy therefor introduces additional antibody proteins into the system. This is still not enough to beat cancer, because cancer is a very diverse decease with a high variation of mutations between patients, and the antibodies used in CAR T-cell therapy are developed to be for a cancer type or patient group, but not for individual patience.



Today on the austrian AI podcast I am talking to Moritz Schäfer who is working on applying Diffusion Models to predict protein structures that support the development of patient specific, and therefore cancer mutation specific antibodies. This type of precision medicine would enable a higher specificity of cancer Therapie and will hopefully improve Treatment outcome.



Existing DL systems like Alpha Fold and alike fall short in predicting the structure of antibody binding sites, primarily due to lack of training data. So there room for improvement, and Moritz work is focused on applying Diffusion Models (so models like DALL-E or Stable Diffusion), which are most well known for their success in generating images, to problem of protein prediction. Diffusion models are generative models that generate samples from their training distribution based on an iterative process of several hundred steps. Where one starts, in case of image generation from pure noise, and in each step replaces noise with something that is closer to the training data distribution.



In Moritz work, they apply classifier guided Diffusion models to generate 3d antibody protein structures.

This means that in the iterative process of a diffusion model where in each step small adjustments are performed, a classifier nudges the changes towards increasing the affinity of the predicted protein to the specific antigen.



# TOC

00:00:00 Beginning

00:03:23 Guest Introduction

00:06:37 The AI Institute at the UniWien

00:07:57 Protein Structure Prediction

00:10:57 Protein Antibodies in Caner Therapy

00:16:17 How precision medicine is applied in cancer Therapy

00:22:17 Lack of training data for antibody protein design

00:30:44 How Diffusion models can be applied in protein design

00:46:06 Classifier based Diffusion Models

00:51:18 Future in prediction medicine



# 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

Moritz Schaefer - https://www.linkedin.com/in/moritzschaefer/

Unser Institut - [https://www.meduniwien.ac.at/ai/de/contact.php](https://www.meduniwien.ac.at/ai/de/contact.php)

Lab website - [https://www.bocklab.org/](https://www.bocklab.org/)

LLM bio paper: [https://www.biorxiv.org/content/10.1101/2023.06.16.545235v1](https://www.biorxiv.org/content/10.1101/2023.06.16.545235v1)

Diffusion Models - https://arxiv.org/pdf/2105.05233.pdf

Diffusion Models (Computerphile) - https://www.youtube.com/watch?v=1CIpzeNxIhU

# Summary

Today on the show I am talking to Martin Huber Co-Founder and CEO of AMRAX.

We will talk about their product Metaroom; an AI application that is build on-top of consumer smartphones and makes it possible to create a digital twins of buildings for indoor user cases, like interior and light design.



We will focus less on algorithms and Machine Learning Methods, but on the impact that sensors and hardware platform have on the AI applications that can be build on top of them.



Martin will explain how Apple's LiDAR Sensors, available in their pro devices, in combination with the Apple Roomplan API are a unique and powerful platform to build AI applications, but at the same time forces one to focus on vertical integration and solutions. We will discuss how as an AI startup in this space one has to be super focused to be successful.



# TOC

00:00:00 Beginning

00:02:57 Guest Introduction

00:05:08 Building hardware vs. writing software

00:07:12 AMRAX & Metaroom by AMRAX

00:13:51 3D Reconstruction with and without LiDAR

00:24:45 Data processing on device & in the cloud

00:30:55 Strategic positioning as a startup

00:37:11 Digital twin for smart home use cases

00:43:47 Future LiDAR sensors and their impact



# 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

Martin Huber - https://www.linkedin.com/in/martin-huber-b940a084/

Lidar - https://amrax.ai/news/power-of-lidar

Metaroom - https://www.linkedin.com/showcase/metaroom-by-amrax/

Apples Roomplan - https://developer.apple.com/augmented-reality/roomplan/

Sony LiDAR Sensors - https://www.sony-semicon.com/en/news/2023/2023030601.html

# 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/