Professionals creating edu content, successful solutions & innovative research in responsible ML & explainable AI • Affiliated with @WUT_edu & @UniWarszawski 🤖
Hey everyone! 👋
We’re a group of scientists creating:
➡️ edu content
➡️ successful solutions
➡️ innovative research
in responsible ML & explainable AI 🤖
We will post daily so you can learn with us - in general & about work we feel proud about.
Let us know who to follow!
If you want to test your resilience against manipulative AI, please play this game: https://t.co/1rzh4SS92H created by @mi2_ai The game is a simplified version of Who Wants to Be a Millionaire? with AI as your best/worst advisor ...
1⃣You should definitely check out the new level glocal explanations we proposed in our latest preprint 📰'Glocal Explanations of Expected Goal Models in Soccer' w/Adrian Stańdo and @PrzeBiec from @mi2_ai, if you only use local and global #XAI tools to explain black-box ML models!
Our proposed solution with details can be found in our recent pre-print, w/ Adrian Stańdo and @PrzeBiec from @mi2_ai, titled 'The Effect of Balancing Methods on Model Behavior in Imbalanced Classification Problems'
You can access the paper on arXiv: https://t.co/d49rO5QZ0U (5/6)
Model explanation is a process. Read more about it in 'The grammar of interactive explanatory model analysis' (https://t.co/pCBpwWO8J0) just published in the SI on Explainable and Interpretable Machine Learning and Data Mining (by @hbaniecki and Dariusz Parzych @mi2_ai#RedTeam)
Many thanks for awesome presentation and organisation to @LindauerMarius @FrankRHutter@AutoML_org!
You can definitely expect @mi2_ai during next editions. 💪
This is what they shared 👇
"Throughout the week we participated in many lectures, hands-on sessions and a hackathon. Lots of knowledge and new insights into the future of #AutoML!
It was also a unique opportunity to meet many people working in this area."
#DataScienceSummit#RadaProgramowa#DataScience
Przedstawiamy Przemysława Biecek (@PrzeBiec), który w tym roku zasiada w Radzie Programowej Data Science Summit.
--
Rejestracja już trwa! Nie zwlekaj i zarejestruj się już dziś!
Today on #ECMLPKDD2022, I have a pleasure of presenting a poster based on our newest paper “Interpretable meta-score for model performance” published in Nature Machine Intelligence.
➡️ It answers a question - how to improve evaluation metrics in ML? 👇🧵
Today, I am presenting our work on adversarial #xai at #ECMLPKDD2022 in Grenoble, France.
Can we trust model explanations?
We highlight that Partial Dependence can be manipulated with adversarial data perturbations!
👇 1/3
One of the topics we focus on in MI2 is #datavisualization.
➡️Ania is our expert in visualization & will definitely help how to do it right! 🔥
For the start, you should visit Ania's profile and check out awesome posters that she & her students prepared about climate topics 👇
Last academic year, together with @hbaniecki, I had the pleasure to teach Data Visualization Techniques course at @WUT_edu 👇
➡️Good #datavisualization is extremely important, so during the course, we introduced students to the world of ggplot2 and other visualization tools.
Another achievement in our lab! 🔥
➡️ @krzyzinskim just published the details of his & his colleagues work on interpretability in survival models with a time-dependent SHAP! Time dependence is a major breakthrough in this case.
Follow him for more on XAI & Survival Analysis 👇
📉 SurvSHAP(t): Time-dependent explanations of machine learning survival models
The lack of #interpretability in #ML survival models is a significant challenge, as in many areas, including medicine, it is crucial to know what affects the model output.
[1/4]
We're very proud of you Kasia and waiting for more cutting-edge research! 🔥
And if you're interested in:
➡️ Hyperparameter Optimization
➡️ AutoML
➡️ Meta Learning
make sure to follow @woznicakat for more content on it!
I had the great pleasure of presenting the results of our last year’s work at #COSEAL22 workshop!
Main topic: Consolidated Learning. It’s a novel look at meta-learning in a specific domain of predictive problems, closer to reflecting real-world applications of ML models.
Ceteris Paribus is a useful method for local explainability which:
1️⃣ for a single example in data (eg. Ben's spendings)
2️⃣ shows how the output changes (eg. his credit score)
3️⃣ when we modify 1 feature (eg. spendings on fuel)
4️⃣ and keep the rest of the features intact.
📦{fairmodels} provides useful tools to check the fairness of the ML models.
I highly recommend the {fairmodels} package for responsible ML!
#rstats#responsibleML
Plotting a graph in readable way is an actual art! 🖌
How to do it right?
➡️ Tip 4: Intuitiveness!
If your data is related to real life concepts, think about intuitive colours for them. Is blue the first colour you think of when describing a hot dessert?
More tips to follow!
Starting with dalex package for Explainable AI is super easy! 🔥
➡️ The idea is to create an Explainer wrapper to your model and data.
In the upcoming flashcards, we will explore what we can do with it. 👀
Link to the full tutorial in the comment! 👇