Thread by @pcmasuzzo

Opening the day with a message by @crevits - but I must admit I didn't get much of it. Quality was not top, and would have been nice to have closed captions or translations of some sort. AI-driven, maybe?! 😬

https://twitter.com/ai4growthBE/status/1503638391488921601

And the technical hiccups are being fixed as I tweet.

We hear now about AI FLANDERS & the research program hereby by Sabine Demey #EUAIWEEK.

The AI Flanders research program, directing research efforts into the following domains - health, industry, energy, and government & citizens.
#EUAIWEEK

What seems like a constructive and useful way to collect data and help practitioners take important decisions: msdataalliance.com/covid-19/covid…
#EUAIWEEK

And a platform for #AI training, VAIA, the Flanders AI Academy
vaia.be/en/
#EUAIWEEK

@ai4growthBE I was trying to ask why this Keynote is being held in Dutch but didn't get a chance to. Is it possible to switch to English?

ELSA labs for human-centered AI
nwo.nl/en/news/more-1…

Starting now track B on research.
First topic is around "Hybrid AI: Embedding expert knowledge in your AI". I always argue that an AI system without expert knowledge does very little. But here we add the expert knowledge explicitly.

#EUAIWEEK
Anomalies detection, fault detection and remaining useful life applications. What do they have in common? They all need a lot of data. #MachineLearning

A lot of data? #DeepLearning might help. #EUAIWEEK

I love this slide. Plenty of challenges in both cases, but the often black box on the left side is what worries me the most when I develop and/or apply #MachineLearning algorithms.
#EUAIWEEK

Nice models on papers, useless in practice. This is the curse of every #DataScience #AI project LOL
#EUAIWEEK

Or, as I like to put it, #MachineLearning will predict your input, not your output.
All this is by Sofie Van Hoecke, by the way. Can't find her on Twitter though.
#EUAIWEEK

Hybrid AI comes to the rescue? Domain ontologies and expert rules become crucial in the modelling part.
Nice use case by on smart home in collaboration with Renson.
#EUAIWEEK

Semantic reasoning and AI. This seems like an interesting read. #EUAIWEEK
medium.com/oxford-semanti….

Hybrid AI to learn partially unknown dynamics - we can't always model all the phenomena at hand, those that we don't know, we replace with Neural Networks. The NN will then hopefully learn the physics parts that are missing in the model. #EUAIWEEK

Knowledge graphs to understand how to transfer learning from some known contexts to a new one. #EUAIWEEK

Now we hear from Peter Hellinckx speaking about hybrid models for control. #EUAIWEEK

A somehow abused term - Digital twins - simulations of real environments, so that we can understand them better and learn how to control them. #EUAIWEEK

Next session - Conversing with your AI system. First talk is towards emphatic conversational agents. Sounds cool! #EUAIWEEK

Leaving this here as I feel soooo many different things about this. #EUAIWEEK

Model dynamic emotional trajectories. This sounds absolutely nuts to me. So curious to see where this is going.
#EUAIWEEK

From sentiment analysis to emotion analysis. Most of the current tools on the market don't model implicit sentiment. It's not so evident to me though what implicit sentiment is.

Another thing we humans should have - common sense knowledge. Can chat bots have and express this type of knowledge?

More on conversational agents and a sort of human-like AI (I'm very skeptical with this term, but seems a lot of research is being done on this in Flanders).

The approach used in the research presented at this session / computational construction grammar is the first building block.
FCG is a software that maps linguistic utterance onto a representation of its meaning.
#EUAIWEEK

Second component: hybrid procedural semantics. The idea is to build an executable network that, when executed,gives an answer to a specific question. The nodes of this network are called primitive operations.

When you combine images and conversation memories as your primitive operators, the network becomes an hybrid procedure.

The third block is the conversation memory. Very simply put, keep track of what the agent and the human being have been saying so far. The conversation memory is expanded dynamically.

And we close the session with a quick demo - always fun to have live demos during conference talks! #EUAIWEEK
(The CLEVR-Dialog dataset and tasks were used for the demo).

"Automating your AI: MLOPS, Data Wrangling & Feature Extraction" is the last session before lunch (I'm already hungry lol) #EUAIWEEK

We start off with #MLOps and features store. I leave this here cause it speaks volumes. #EUAIWEEK

And the other side of it.

Increase confidence and trust in the insights and products delivered by #DataScience models seems to me reason number 1 to move towards MLOps practices.

Bottom line message is - a data scientist needs to do a bit of everything, and needs to be able to do a bit of everything (and I'm realising there's no way out of this).

And here it is. I'll stretch it a bit more saying that data scientists cannot work without the domain experts. Translating data & the models therein into value and insights is an effort that requires hard work from many stakeholders. #EUAIWEEK

Now we move on to a specific application - time series and predictions of anomalies, failures, etc.
Usually we do feature extraction and we then try to learn something from the data. This can be either via a domain expert system or automated via #MachineLearning or #DeepLearning.

Combining more data sources is useful, but only when it's done in an intelligent way. Convert them all to a common representation could be one way of doing this.

Adding knowledge seems to be the way to go. This keeps coming back and I think it's where we need to invest future research efforts indeed.

More on time series & open source tools to extract features from them and actually visualise them better:
tsflex and plotly-resampler!

Plotly-resampler is available here
github.com/predict-idlab/…. The library is designed to support both interactive behaviour and be able to scale well to large datasets.
I'll definitely try it out!

While tsflex has more to do with the data wrangling step - filtering, resampling, signal decomposition,...
The library is here
github.com/predict-idlab/…
And let me just say it's so cool these folks made their research outputs #OpenSource and accessible :)
#EUAIWEEK

I also love that these packages don't reimplement the wheel but rather leverage and support what other packages in the community already have. This is truly building community resources that serve everybody in the best way! #EUAIWEEK

And here too we have a demo!

And it's a wrap! Time for some food ;)

And we're back with a keynote by Jonathan Burez from @Belfius - the OneApp strategy, a single place to engage with the customers.

MLOps comes back - challenge is as always to translate local analyses into products that become operational and help the business. And we close with responsible AI.

Python predictors github.com/PythonPredicti… for model explainability and fairlearn.org for assessment of fairness of AI systems.
#EUAIWEEK

And now back to track B again on AI-Driven Mobility Solutions at sentiance.com

Motion sensing, distraction insights and digital coaching - the three major blocks of this system. A system that tries to educate, coach and challenge the drivers into adopting best practices when driving.

Drive behaviour change through an app:
sentiance.com/behavior-chang…

Now the time for OTIV - teaching vehicles to drive autonomously:
otiv.ai

Followed by ML2gGrow
ml2grow.com - and we're reminded of this quote from Andrew Ng "AI is the new electricity" - it's no longer hype. #EUAIWEEK

And after my 3rd ☕ (I know I know), we're back with "AI-Enabled Retail Analytics" by Crunch Analytics. Curious to see what this session will be about (though I can guess from the title lol)
crunchanalytics.be

"The sustainable retailer of tomorrow needs to embrace data and AI" (today, I would add). These data come from e-commerce and omnichannel platforms. And here's a book worth having a look at - Prediction Machines -
predictionmachines.ai

Some use cases on pricing and marketing. Price markdowns - special discounts on items that are at their end of life cycle (when seasons change, for example). Overstock is a waste.

This problem comes with high uncertainty (very hard to say if the discount will influence the sales momentum); there are often also very diverse objectives (the optimization process is unclear); and it's usually managed by humans via cumbersome Excel spreadsheets + business rules

Another aspect is the price elasticity - how much does a price influence the sale momentum and volume? A 10% discount can for example result in a 1% sale uplift or a 30%. These are two very different things.

Other factors to consider are Product lifecycle / inventory levels / business rules. The goal becomes to optimize the value of revenue, for example, and find the price markdown accordingly. The objective could also be to minimize the overstock, however.

Last contribution on Marketing in retail - a perspective. Buying, distributing, marketing and selling. At every step you need to ask yourself who? what? where? and when?

We hear a story on Customer Lifetime Value, which can help you predict next purchase date, monetary value, etc.

Based on these predictions you can build a new set of tools to communicate better with your customers (for example sending emails at specific times).

I wonder though how can these types of model help profiling or understand customers better when the retail items span a much longer time period (you don't buy as many TVs as you buy shoes, of course). But hey, that's the end of this session! Back to plenary now.

No, I was wrong, another parallel session first on AI in food and farm. Major challenges here are seasonality of data, variation and variability present in nature. Tomayto, tomahto (I couldn't resist).

AI applications that don't involve all key players (and from the very initial first step, I might add) are faulty by design.

Amen.
#EUAIWEEK

And now it's time to move to the plenary for the final talk of the day! Dries De Dauw from Colruyt is on stage.

I'm impressed by how Colruyt is using and leveraging AI to create more efficient and more sustainable processes and businesses!

And of course we close with some challenges. Data quality is and will always be number 1 on the list.

Is it easy to embrace AI and make sure we produce valuable output with it? No. Is it worth it? Absolutely yes.

And that's a wrap! #EUAIWEEK

@threadreaderapp please unroll :)