This year, SABO Mobile IT became a Gold partner of the largest European conference on machine learning and artificial intelligence held in Prague – Machine Learning Prague 2022. Now it is time for a look back.
The conference took place in „La Fabrika” in the center of Prague for three days from 27-29 May, including a workshop that was held at the “CEVRO institute” in Prague. More than 500 international participants made it a great success!
An interesting organizational feature was used – the conference was “hybrid”, so it ran both in person and online in combination with the eventee app (https://eventee.co). For attendees onsite it was very useful as a platform, as well to ask questions to speakers during their talks. The program and abstracts of the talks were also available in the app.
SABO Mobile IT presented its applications for industrial use, as well as its chatbot SABOT – a voice assistant for MIWE baking ovens. The practical example was presented in the partner section. SABO Mobile IT’s contribution to machine learning research was also presented in the talk by Timo Leitritz (Fraunhofer Institute for Manufacturing Engineering and Automation) - Zero to Hero: AI based assistance in industrial machine operation.
The conference was very interesting, inspiring, and fun. I was glad to see what machine learning scientists do, how they think, and what concerns them. I was able to learn many interesting things, though obviously with the amount of technical information provided, I only received a big-picture overview.
For us, the most interesting presentations were:
Really amazing application of NLP models for smarter search and recommendation of proteins in scientific papers based on necessary properties. Machine learned model by reading tens of thousands of scientific papers related to specific molecules. Then, this trained model was used to highlight 20 out of 20.000 molecules which may have significant impact in cancer treatment.
Very nice overview of object detection models, their architectures, and pros and cons. Also, Yauhen Babakhin mentioned that Transformers are used more and more in the field of image recognition, and that they are already achieving state of the art results for image recognition and object detection tasks. It will be very interesting to see what Transformers can achieve in the future in this field of AI.
SLEM stands for self-learning and self-explaining machine, and is the name for this collaborative project between Fraunhofer IPA, Knowtion, and SABO Mobile IT. It is funded by the Baden-Württemberg Ministry of Economics, Labor, and Housing as part of the Baden-Württemberg AI Innovation Competition.
The first step is to train SLEM‘s recognition models on a new scene or application scenario (e.g. a specific machine) using annotated data. Then, an expert executes a relevant process that the SLEM observes and learns. A less experienced operator performs the process with the assistance of SLEM. Errors are detected and the correct steps are communicated to the operator.
With a deep dive into the activity recognition using pose data, Timo Leitritz generated a lot of interest - the many questions after the presentation could not even all be answered.
A nice example of a machine learning application in business: analysis of the efficiency of marketing channels – it replaces the method that uses 3rd party cookies.
Bayesian models have certain advantages for B2B customer: explainability (why the model shows this result is much easier to tell than in a neural network), setting up constraints, and embedding them into the model of existing business knowledge.
With 10 million inhabitants, the Czech Republic has its own search engine, which is quite unique and successful. Starting out in 1996, Seznam today runs about 30 different web services and associated brands. Interesting to see how a local but big tech company works, what it has achieved, and how it enhanced the organic search on seznam.cz by introducing semantic vectors.
Which features can be extracted from songs and how. Yama Anin Aminof described how Meta analyzes uploaded music tracks, splits them into lyrics, melody, and rhythm and finds similar audio in a database of known songs. Analysis of popular music and some other results. Yama Anin Aminof also showed that sometimes boybands and heavy metal bands have very similar lyrics in their songs which are close in the feature space even though the way they perform those lyrics is very different.
Felt back at CTU study halls again :)
The process of creating a chatbot for social topics: music, movies, sports, politics. “Alquist” has made it several times to the finals and won the Alexa price last year. Really interesting. One of the requirements was that the average conversation should be at least 20 minutes in length!
SABO Mobile IT would like to thank the organizers of Machine Learning Prague 2022 for their amazing work- everything just worked as expected. The program schedule was accurate without any delays, there was a lot of good coffee, drinks, meals, and desserts, and after a long time an inspiring event that was held live.
Nikita Evstigneev, Lubos Brat, Timo Tränkel
