Christoph Molnar 🦋 christophmolnar.bsky.social
@ChristophMolnar
Followers
32K
Following
7K
Media
774
Statuses
5K
Author of Interpretable Machine Learning https://t.co/gJKlTA2deP | Newsletter: https://t.co/6fQuMr8yI8
Munich
Joined July 2012
Exciting news!. Our book Supervised Machine Learning for Science is now published! 🥳🥳🥳. Available in hardcover, paperback, eBook, and PDF. You can even read it online for free!
5
52
281
BREAKING: IBM makes game-changing move, invests $42 in ChatGPT Pro to revolutionize IBM Watson capabilities.
138
544
9K
I have the solution for detecting AI-generated text.
225
255
4K
2 years, 250 pages, 1,219 commits, and 78,480 words: . I am very proud to say that today I published the 1st edition of "Interpretable Machine Learning". 🎉🎉🎉. Web: Leanpub:
71
1K
4K
When your machine learning model is confronted with data it has not been trained on.
63
703
3K
You can't "train" a model. The model always exists. It existed before you were born and it exists after your death. You can only find the model. "Training" is just your way of looking for the model's location in the infinite hypothesis space and binding its essence to silicon.
197
298
3K
I write my books with vim. Why tough?. While there are many good reasons to use Vim for book writing, one stands out: I don't know how to exit Vim.
76
233
3K
Machine Learning concepts as animal GIFs. A thread 🧵.
53
1K
3K
Watching the k-means cluster slowly converge
33
381
3K
Ready to nerd out on random forests?. Someone wrote an entire Ph.D. thesis on random forests ❤️. 200 pages of gems. Highlights: Why ensembles work, interpretability of forests, and making random forests work on huge datasets.
14
472
3K
A lot of machine learning research has detached itself from solving real problems, and created their own "benchmark-islands". How does this happen? And why are researchers not escaping this pattern?. A thread 🧵
48
685
3K
Evaluating a model on training data is like asking your mom if you look good.
41
276
2K
The mean is the best prediction model. • Needs no features.• Easy to compute.• No overfitting.• Interpretable.• Optimizes L2.• Analytical solution
116
135
2K
Not many people know this, but one milestone in becoming an ML researcher is getting blocked by Lex Fridman on Twitter.
82
91
2K
Neural networks with only linear layers would result in linear models. Because a linear combination of linear functions is again linear. Nothing would be gained by adding more layers.
ML folks, does anyone have a good resource/explanation for *why* the use of non-linear functions (e.g. ReLU) in neural networks and why they work so well? . My simplified version is:. Combine enough straight (linear) and non-straight (non-linear) lines and you can draw a
31
177
2K
"Machine Learning is just statistics." I thought as I entered my first Kaggle competition. Turned out to be a harsh reality check: ML and statistical modeling require different mindsets. This haunted me for many years, so I wrote it all down in my book Modeling Mindsets.
12
223
2K
Where are you on your machine learning journey?
17
186
2K
I just found a great introduction to embedding. The book is comprehensive yet short. Historical encoding tools, neural nets, and production - all covered. Fantastic job by @vboykis. Thanks for making it free to read!. Looking forward to diving in.
5
270
1K
Will this now happen with every new LLM update?
17
183
1K
Some dude at a BBQ once asked me what I do. I said research in ML / AI. Dude went off into an angry rant about surveillance, killer robots, and self-driving cars. No chance to tell him my work is most useful for data stored in Excel for some boring process in an insurance corp.
29
54
1K
"Machine Learning is just statistics." I thought as I entered my first Kaggle machine learning competition. Turned out to be a harsh reality check: ML and statistical modeling require different mindsets. This haunted me for many years, so I wrote the book Modeling Mindsets.
8
149
1K
What do you call a machine learning model that perfectly predicts the training data, but does not work for unseen data?. A database.
38
128
1K
How it started:. f(X) = X. How it's going:. • is f AGI?.• f might be conscious.• f will conquer the world.• You can use premium f for $20 a month.• f is only 14 weeks old and here is what you can do. • Here are 10 products of f(X)=X you haven't heard of.
30
162
1K
I "grew up" as a statistician. When I later learned about machine learning, I found it a mind-blowing new perspective on data modeling. The best data modelers don't ideologically follow one mindset but have many at their disposal. But which modeling mindsets are there? 🧵.
15
193
1K
Progress in generative AI . Past: Cherry-pick good results.Now: Cherry-pick bad results.
15
74
1K
Novice data scientist: Data is data. Expert data scientist: Each data column has a difficult past. A past full of life-altering events (distribution shift), loss (missing data), misunderstandings (wrong encodings), and regrets (no documentation of how data was collected).
11
172
995
My book Interpretable Machine Learning has been career-defining. Even though the book is available on the web for free, many have supported me by buying the ebook or paperback, which ultimately helped me become a full-time writer. I'm grateful for all your support 🙏
12
159
1K
The first version of my online book on Interpretable Machine Learning is out! . I am very excited to release it. It's a guide for making machine learning models explainable. #interpretableML #iml #ExplainableAI #xai #MachineLearning #DataScience
16
449
968
Machine learning sucks at uncertainty quantification. But there is a solution that almost sounds too good to be true: . conformal prediction. • works for any black box model.• requires few lines of code.• is fast.• comes with statistical guarantees. A thread 🧵.
19
140
945
LASSO is the Marie Kondo of linear regression models
5
164
881
This is used as an example of a Gaussian distribution. But it's not a pure Gaussian. We must increase the nerd level:. It's a mixture distribution of a discretized and truncated Gaussian with 0 and 135 (?) cut-offs and a Dirac delta function at 100.
16
96
878
Studying statistics was one of the best decisions I made. After finishing school, I had no idea what industry I wanted to work in. I liked math. So I chose statistics. It's like having a janitor's key ring that gives me access to any industry. I never looked back.
21
91
870
No amount of deep learning can fix bad data. Modeling is fun, good data is hard work. Improving data can mean. - going through hand-written records.- building a labeling interface yourself.- find out why data is missing.- investigating the "trail" of the data.- weeks of curation.
13
129
817
We have a new paper on arxiv 🎉🎉.Interpretable Machine Learning - A Brief History, State-of-the-Art and Challenges. Best to read in a comfortable chair with a cup of coffee/tea. It's an extended abstract to a keynote I gave at the ECML XKDD workshop.
17
208
810
I'm calling for a more minimalistic approach to machine learning:. Regression? Predict the mean.Robust regression? Median .Classification? Always predict majority class.Time series? Same as yesterday: Y(t) = Y(t-1).Text generation? Lorem ipsum.
25
68
818
Bayesians: "Join us! Updating prior believes is exactly how humans learn.". Causal inference: "Join us! Humans think in causal relationships.". Machine learning: "Join us! Models are black boxes and so are humans.". 1/2.
6
118
792
Reinventing the wheel
7
115
787
Most ML interpretation methods have a common enemy:. Correlated features. They ruin interpretation both on a technical and a philosophical level. Why correlation is problematic, how to patch it, and why we have no cure. A thread 🧵.
15
167
780
Machine learning interpretability from first principles:. • A model is just a mathematical function.• The function can be broken down into simpler parts.• Interpretation methods address the behavior of these parts. Let's dive in.
3
120
785
It took me a long time to understand Bayesian statistics. So many angles from which to approach it: the Bayes' Theorem, probability as a degree of belief, Bayesian updating, priors and posteriors, . But my favorite angle is the following first principle :.
8
114
762
Regression models usually just output a point prediction. That's a problem. Because the prediction could be spot on or it could be a wild guess. To distinguish these two scenarios, we need uncertainty quantification. A solution: Conformal Prediction.1/n.
17
117
749
Bayesian modeling from first principle and memes. Let's go.
15
117
716
Statisticians and Data Scientists
7
59
708
SHAP, LIME, PFI, . you can interpret ML models with many different methods. It's all fun and games until two methods disagree. What if LIME says X1 has a positive contribution, SHAP says negative?. A thread about the disagreement problem, and how to approach it:.
19
138
720
Simply apply SHAP and other methods to interpret your machine learning model, paste those charts into a report and you're done?. Unfortunately, if you do, you're likely to make a mistake. Avoid these pitfalls when interpreting your machine learning model.👇.
12
162
723
Been doing ML for some years now. Here's my ranking:. Supervised ⭐️⭐️ ⭐️ ⭐️ ⭐️ .Self-supervised ⭐️ ⭐️ ⭐️ ⭐️ .Reinforcement Learning ⭐️ .Unsupervised 🚮.
51
43
704
Unpopular (?) opinion: Scientific papers are designed to be defensive which comes at the expense of honesty and readability. I blame the peer review system.
30
41
685
Finally, the second edition of Interpretable Machine Learning is here! 🥳🥳🥳. So far only as e-book. Paperback should be available shortly (and in color!). Leanpub: Amazon:
11
117
669
Data science cannot be fully automated. No amount of parameter tuning, benchmarking, automation, model comparison, automated feature engineering, etc., can automatically figure out what data column location_123_old contains and whether it should be a feature or not.
24
74
609
Supervised learning "only" gives you a prediction function. But with the right tools, you'll get a lot more:. • Uncertainty quantification.• Causality.• Interpretability.• Analysis of variance.• . And the best news: tools in this thread work for any black box model. 👇.
8
98
599
It's overwhelming to keep up with research on interpretable machine learning. I say that as the author of the Interpretable Machine Learning book. 😅. I use these 3 questions to quickly understand new interpretation methods:.
11
96
602
Super excited to release my new book Modeling Mindsets tomorrow 🥳. This is my first book since I decided to become a full-time writer. And it's the book I wish I had read a few years ago to save myself a lot of time. 1/n
10
80
593
I'm hoping developers take the AI dev thing seriously. Self-driving cars were introduced in the mid-2010s. Look at all the fools who got their driver's license nonetheless and now have a worthless document. Or think of all the jobless radiologists trying to make ends meet.
30
63
588
Logistic Regression sold as AI
2
46
558
I found it unexpectedly difficult to get into causal inference. (still a beginner, I guess). Here are a few insights that helped me in understanding causal inference. 🧵.
15
64
546
If you were to learn only 1 method for explaining machine learning models, it should be Shapley values (SHAP). - Model-agnostic: Use with any model.- Theoretic foundation: Game theory.- Good software ecosystem.- Local and global explanations. More here:
16
82
563
One of the best arguments for supervised learning was made by a statistician. Statistical Modeling: The Two Cultures. Every modeler should read it. The paper is written by Leo Breiman, the inventor of Random Forests.
11
104
553
Early on, I had made "Interpretable Machine Learning" (web + e-book) available for 0$. I never regretted the decision, and I happily share my knowledge. And yet so many of you opted to pay for the e-book, encouraging me on my path of free knowledge. I am grateful 🙏.
6
29
541
How COVID ML papers are created
9
98
539
The modeling mindset of statisticians in one tweet. 1) Measure random variables, e.g. water temperature.2) Assume distribution, e.g. Normal distribution.3) Goal: Find optimal distribution parameters, e.g. mean.4) Solution: Maximize the likelihood. Parameters = Insight about world
9
78
529
Statistics, Machine Learning, Causal Inference, . People coming from these different modeling cultures often have misunderstandings, and fight over whether logistic regression is "AI" or statistics. Why the big differences?. 1/n 🧵.
8
131
536
This tweet is . 20% food for thought .80% shitposting. Thank you for your attention.
16
6
523
Missing data in a random forest.
10
57
489
What are some limitations of interpretable machine learning methods?. This summer, our students worked on this question. We compiled the results in a free online book, which we release today. 🎉🎉🎉. Find out more:.
12
164
484
To become a statistician, you must study the divine numbers:. 30 represents sufficiency. 0.049 is a sign of good fortune. 1.96 protects you from uncertainty. 0.051 says that your career is in danger.
8
47
485
I don't get why people still use SQL????. • Rejects most prompts. Requires specific prompting. • 100% overfitting to training data.• No generalization to new data.• Non-stochastic parrots.
25
40
493
Supervised machine learning models are deployed everywhere. It's an open secret that all models have a huge problem:. Performative prediction - when predictions change future outcomes. How to spot and handle this problem: A thread 🧵.
19
86
474
I see what you did there. Love the title.
17
32
482
When machine learning researchers reinvent methods from statistics.
8
70
460
Statistical modeling from first principle. But with memes and GIFs. Let's start.
4
117
467
We assume linearity
5
71
455
3 days left. Looking forward to the release of Supervised Machine Learning for Science.
9
50
462
Most machine learning applications fail. Google Flu, Amazon hiring, ML in healthcare, . Why?. A wide range of reasons. But most are due to the same flaw: not thinking through the problem as a statistician would. Adopt a statistician's mindset. Become a better machine learner.
15
67
446
tabular -> import xgboost.text -> from transformers import pipeline.image -> import torch, torchvision.chat -> import langchain. thanks for coming to my TED talk.
10
35
451
Modeling Mindsets is published 🥳. From Bayesian inference to unsupervised learning: Modeling Mindsets provides an intuitive introduction to different modeling approaches and allows you to choose the right one for your problem. Get your copy at:
21
84
443
Is linear regression machine learning or stats?. is like asking. Is an apple a fruit or is it something to eat?.
32
29
433
Some examples how interpretable machine learning (IML) is used for scientific discovery, when the underlying model is a black box model. A few examples. Thread 1/n.
8
109
429
I once met a professor who was OBSESSED with h-index and citations. - Advised me to stop my book, turn it into a paper to get cited.- Output not published at icml/ neurips? Worthless.- Gave a talk - it was only about flashing status symbols. Absolutely corrupted by THE GAME.
21
22
422
The more I dive into machine learning targeting COVID, the more I believe that the research impact is not zero, but negative. Negative, bc. many prediction models are low quality, unusable. Waste of time + resources. Adding noise to an already crazy publishing environment.
16
59
409
Deep learning crowd getting novel models.
5
52
410
Welcome to academia
3
27
421
My new book, Interpreting Machine Learning Models With SHAP is now also available as a paperback (in color)! 🎉 . SHAP is the Swiss army knife of ML interpretability and this book is a comprehensive guide to the theory and application of SHAP.
5
41
411
Looking for a career in AI?. I have you covered.
8
41
385
Follow me for more tips on how to sound smart
8
18
397
7 years ago, I wrote the first chapters of Interpretable Machine Learning. I didn't know back then, but this was the start of my writing career. In 4 days, I'll publish a new book project. Online. Free access. In many ways, it weaves together many threads of my career.
17
27
385
Interpretable Machine Learning has >2000 citations 🥳. This reminds me of the professor who laughed at me when I told him that people were citing the book (8 citations at the time) and told me, "You should turn the book into review paper so you can get citations." 🤷♂️.
6
18
384
I have started to create cheat sheets for model interpretation. So far:. • Logistic regression: • SHAP plots for tabular data: They are free (pay what you want). Do you have any topic wishes for the next cheat sheets?.
16
59
379
Bias-variance tradeoff, generalization error, no free lunch theorem. -> These concepts come from statistical learning theory which views supervised machine learning through a statistics lens. It's heavy on math and theory. Bookmark the best entry point:.
1
56
384
Happy to publish our book project “Supervised Machine Learning for Science” today 🥳. Read it for free here: The book is about the philosophical and pragmatic justification for using supervised machine learning in science.
10
79
373
A lot of interpretation methods for machine learning are connected to other fields. Game Theory -> Shapley Values.Philosophy -> Counterfactuals.Sensitivity Analysis -> model-agnostic methods.Statistical Modeling -> linear model, GAMs.Rule-Based ML -> Decision trees. What else?.
16
56
378
Two types of Bayesians. 1) Beer pub Bayesian: People who want to be seen as intelligent by using words like "Bayesian", "prior", "posterior", "updating" to refer to the fact that people have opinions and sometimes change their minds. 2) Actual Bayesian who analyzes data.
21
28
347
Are researchers putting too much burden on this smol boy?
2
35
354
"You should turn your book into a paper so that it gets cited.". "You have to publish at ICML/ECML to have any impact.". It can be difficult to decide when to follow conventional advice and when not. I am happy I did the unconventional with Interpretable Machine Learning.
12
37
352
@nntaleb Bayes' theorem: P(H | D) = P(D | H) P(H) / P(D). Conspiracy' theorem: P(H | D) = P(H). where. H: Hypothesis.D: Data.P(H): prior belief.P(D): evidence.P(D | H): likelihood.P(H | D): posteriori belief.
9
76
339
Most common arguments against interpretable ML: .- Humans can't explain their actions either.- Performance > Interpretability.- Slows down ML adoption.- Linear model also not interpretable.- Might create illusion of understanding the model. Thread:.
10
149
335