Bojan Tunguz (@tunguz) | Robocrunch.Net

Bojan Tunguz @tunguz

#MachineLearning at @Nvidia. @Kaggle Quadruple Grandmaster. Data Scientist. Physicist. Catholic. Husband. Father. @Stanford Alum. Opinions my own.

Tweets by Bojan Tunguz

In the age of increasingly ubiquitous and powerful AI, your datasets are your moat.

Shared by Bojan Tunguz at 5/14/2022

This week DeepMind has released Gato - a single transformers-based architecture that is capable of working on over 600 different tasks. 1/3 https://t.co/C3KlA90MIU #ArtificialIntelligence #AI #MachineLearning #ML #DeepLearning #DL

Shared by Bojan Tunguz at 5/14/2022

LOL, now I’m being quoted in various articles. 🤣 ‘“This is something that many of us in the Python community have been hoping for for a long time,” noted Bojan Tunguz, a machine learning modeler at @Nvidia.’ https://t.co/FQSDqoXpGp via @DEVCLASS

Shared by Bojan Tunguz at 5/5/2022

From the very beginnings of the research in AI, the ability to understand natural language has been seen as the pinnacle and the end goal of what it means for an agent to be intelligent (think of the Turing Test). 1/6

Shared by Bojan Tunguz at 5/5/2022

That’s why only college dropouts should be in charge of startups. https://t.co/POhaOvnT0C

Shared by Bojan Tunguz at 5/5/2022

"A fox knows many things, but a hedgehog knows one big thing." - Archilochus This was a quote that Isaiah Berlin expanded into one of his essays, first published in 1953. The quote was used to classify writers and thinkers into two big categories - those who view the world 1/18

Shared by Bojan Tunguz at 5/3/2022

•Python ecosystem: Run many popular packages of Python and the scientific stack (such as numpy, pandas, scikit-learn, and more) •Python with JavaScript: Bi-directional communication between Python and Javascript objects and namespaces 3/6

Shared by Bojan Tunguz at 5/2/2022

Key features include: •Python in the browser: Enable drop-in content, external file hosting, and application hosting without the reliance on server-side configuration 2/6

Shared by Bojan Tunguz at 5/2/2022

How Banks Are Winning with AI and Automated Machine Learning https://t.co/zCpFPf04Iu via @rightrelevance thanks @datarobot

Shared by Bojan Tunguz at 4/30/2022

DeepSolar Dataset on #kaggle via @KaggleDatasets https://t.co/c9cxs056VN

Shared by Bojan Tunguz at 4/30/2022

If you are working in Python, and find yourself in a need of encoding categorical variables for various ML tasks, I'd strongly encourage you to take a look at the Category Encoders library. It's a very good and fairly comprehensive sklearn-style library. https://t.co/N8wubPlKDF

Shared by Bojan Tunguz at 4/28/2022

I’ve reached 40,000 followers on Twitter. Grateful for every one of them. 🙏 I tweet about: 🤖 Machine Learning/AI 🐍 Python 🟩 NVIDIA 🤪 Memes 👨‍💻 Future of work 🔭 Science 🔐 Blockchain Follow me to learn more about those topics.

Shared by Bojan Tunguz at 4/27/2022

Three ways you can provide insights: 1. Analysis 2. Visualization 3. Models

Shared by Bojan Tunguz at 4/27/2022

I honestly don't believe that ML/AI will ever have a substantial impact on TradFi. For ML to have a substantial impact on *any* domain, we need to go through a major digital transformation first. And that seems to be virtually impossible to pull off in TradFi.

Shared by Bojan Tunguz at 4/23/2022

Here are my top used ML algos for tabular data problems: 1. XGBoost 2. HistGradientBoosting 3. Logistic regression/ Ridge regression 4. LightGBM 5. MLP 6. A blend of 1. And 5.

Shared by Bojan Tunguz at 4/22/2022

One of the things that people who have never given Kaggle a serious try don’t understand is what it *really* takes to be successful there, and in particular what it takes to win a competition. No, it’s NOT about squeezing the maximum performance out of an algorithm. 1/5

Shared by Bojan Tunguz at 4/22/2022

I believe that going forward the biggest bottleneck for the continuing growth of the tech sector will be the availability of well trained tech talent. The demand is far outstripping the supply that's coming down the educational pipeline. 1/2

Shared by Bojan Tunguz at 4/21/2022

“We face danger whenever information growth outpaces our understanding of how to process it.” - @NateSilver538

Shared by Bojan Tunguz at 4/18/2022

All the best machine learning modelers I know are making their best models on @numerai now. Probably nothing.

Shared by Bojan Tunguz at 4/18/2022

Them: large language models are getting too powerful! Large language models:

Shared by Bojan Tunguz at 4/18/2022

Trying to build a tech startup at the steepest part of the sigmoid adoption curve is a very precarious thing to do.

Shared by Bojan Tunguz at 4/14/2022

I am excited to share that a paper that I contributed to “Combinatorial optimization of mRNA structure, stability, and translation for RNA-based therapeutics” has been published in Nature Communications. https://t.co/R7cDHDlWyw #mRNA #biology #vaccine #machinelearning 1/5

Shared by Bojan Tunguz at 4/14/2022

One of my favorite things about XGBoost is how easy it is to install. Here it was a breeze to do so on a Raspberry Pi Zero! From supercomputing clusters, to the edge micro devices #XGBoostIsAllYouNeed.

Shared by Bojan Tunguz at 4/13/2022

I decided that in order to understand the large language models I needed to inspect all the weights by hand, so I printed them all out.

Shared by Bojan Tunguz at 4/8/2022

Just FYI, there is *much* more to Kaggle the platform than the Kaggle-sponsored ML competitions. If your concern is that those competitions are not representative of the "full" DS workflow, you *could* setup your own that mimic the rest of the DS pipeline as you see it. 1/2

Shared by Bojan Tunguz at 4/4/2022

As I was saying earlier … DataRobot’s vision to democratize machine learning with no-code AI https://t.co/dg1v6eTFSu

Shared by Bojan Tunguz at 4/3/2022

NB: there are currently at least two tech unicorns, many other high-value startups, and countless number of extremely successful careers that were built by Kagglers using the skills and insights they gained on that platform.

Shared by Bojan Tunguz at 4/3/2022

XGBoost Is All You Need Deep Neural Networks and Tabular Data: A Survey https://t.co/Z2KsHP3fvp

Shared by Bojan Tunguz at 3/30/2022

Here is a great post about the genealogy of the most popular transformers models by Xavier Amatriain. @xamat https://t.co/w2Z7Kc7opG #NLP #ML #DeepLearning #ArtificialIntelligence #AI

Shared by Bojan Tunguz at 3/28/2022

Deep Learning with PyTorch https://t.co/yjuSB9QxvK Introducing Python https://t.co/iUmywBeHRw Machine Learning Using TensorFlow Cookbook https://t.co/bKlQkiiJMV Natural Language Processing with Transformers https://t.co/D1X8dPgHNW 5/6

Shared by Bojan Tunguz at 3/23/2022

Hands-On Gradient Boosting with XGBoost and scikit-learn https://t.co/sRLIFOI8DT Deep Learning with Python https://t.co/TjL36JKjqA Effective Pandas https://t.co/2UPkRsvDa7 Machine Learning with PyTorch and Scikit-Learn https://t.co/J0sRKyC2Wb 6/6

Shared by Bojan Tunguz at 3/23/2022

Books: Approaching (Almost) Any Machine Learning Problem https://t.co/fYVtINpZ4z Feature Engineering for Machine Learning: Principles and Techniques for Data Scientists https://t.co/d4myngd69o Hands-On Unsupervised Learning Using Python https://t.co/VbuK177Clr 4/6

Shared by Bojan Tunguz at 3/23/2022

Today @NVIDIA has announced a new version of the high-end server class GPU - H100 Tensor Core GPU. This is an order-of-magnitude leap over the previous generation in computational performance. #GPU #gpucomputing #machinelearning #artificialintelligence https://t.co/rQBIm0ui5b

Shared by Bojan Tunguz at 3/22/2022

What is something that you can say during sex, but also while your ML model is training?

Shared by Bojan Tunguz at 3/19/2022

Wordceling is just shape rotation in the latent embedding space.

Shared by Bojan Tunguz at 3/15/2022

Don't forget to shift your learning rate one hour today. Otherwise none of your models will converge.

Shared by Bojan Tunguz at 3/13/2022

Very interesting work on Unsupervised Feature Selection using Supervised Algorithms. It turns out that even for unsupervised learning, XGBoost is all you need. :) Blog post: https://t.co/o2nC7wUl7d Code repo: https://t.co/aOR3mJ9VEy #machinelearning #unsupervisedlearning

Shared by Bojan Tunguz at 3/12/2022

4. Many thousands to about a billion datapoint is where Gradient Boosted trees rule. If you need just one algorithm, go with this. You'll never go wrong. 3/

Shared by Bojan Tunguz at 3/10/2022

OK, here is my honest take on when to use which approach/technique with a given dataset. These are my rules of thumb, and caveats could fill out the entire internet. 1. Up to a few hundred datapoint, use stats 2. For few hundred to few thousand use linear/logistic regression 1/

Shared by Bojan Tunguz at 3/10/2022

Woohoo, I'm winning! Thanks @abhi1thakur for organizing this one! "Overfitting"? Nay, it's called #winning. #kaggle - https://t.co/Cl9JscwRy2

Shared by Bojan Tunguz at 3/9/2022

Hot take: Statistics is just poor man's Machine Learning. Send tweet.

Shared by Bojan Tunguz at 3/9/2022

LinkedIn and DeepMind co-founders form AI startup to help humans talk to computers https://t.co/w08cXvpwSg via @engadget

Shared by Bojan Tunguz at 3/8/2022

Everyone is like 4D chess this, 10D chess that, and I am like dude ever heard of Hilbert space??? Unless you are playing chess in infinite dimensional vector spaces, you are ngmi.

Shared by Bojan Tunguz at 3/7/2022

but we are only scratching the surface of what massive parallelization can achieve. Once your datasets contain all of human knowledge, it's hard to imagine where else you can find more data to feed your algorithms with. 5/7

Shared by Bojan Tunguz at 3/6/2022

For at least a decade now we have been enjoying an unprecedented AI Springtime. A perfect storm of major advances in algorithms (deep learning), computational architecture (GPUs) and availability of large high quality datasets has enabled the field to grow - exponentially! 2/7

Shared by Bojan Tunguz at 3/6/2022

You think your image segmentation deep learning model is SOTA? Try it on this image.

Shared by Bojan Tunguz at 3/5/2022

I just came across this cool deep learning image library. Anomalib is a deep learning library that aims to collect state-of-the-art anomaly detection algorithms for benchmarking on both public and private datasets. https://t.co/hThDksM5m5 #computervision #dl #ml #ai

Shared by Bojan Tunguz at 3/3/2022

Being able to think in terms of probabilities is a superpower, period. And that includes your own beliefs and opinions. Probabilistic and statistical thinking unfortunately doesn’t come naturally to us. Takes years of training and experience to develop.

Shared by Bojan Tunguz at 3/3/2022

Just received my review copy of Machine Learning with PyTorch and Scikit-Learn. Thank you @rasbt for making this happen! The book is huge, much bigger than Python Machine Learning. Seems really comprehensive, covers almost all relevant areas of modern ML. https://t.co/aEFpjo9gek

Shared by Bojan Tunguz at 3/2/2022

I’ve gotten a lot of support for this tweet, far more than I had expected. I also got some pushback, but NONE of the pushback challenged the basic facts of my allegations. Almost all of the pushback was just empty ideological posturing.

Shared by Bojan Tunguz at 3/1/2022

way forward. I feel that we are getting close to the limit of what this line of attack can achieve. My *intuition* is that we'll probably need to take a step or two back, and try to somehow incorporate more insights form how humans understand and process meaning in language. 15/

Shared by Bojan Tunguz at 2/28/2022

I just got my review copy of “Natural Language Processing with Transformers” from @OReillyMedia. Thank you @_lewtun for making this happen! Really looking forward to digging deeper into transformers and @huggingface. 🤗 https://t.co/g1NGNPLXlz

Shared by Bojan Tunguz at 2/26/2022

I’ll never stop being fascinated by the fact that 99.999% Artificial Intelligence researchers know absolutely nothing about a century+ of solid fundamental research in *human* Intelligence. 1/

Shared by Bojan Tunguz at 2/26/2022

200,000+ Jeopardy! Questions on #kaggle via @KaggleDatasets https://t.co/nyohzyj1mX

Shared by Bojan Tunguz at 2/23/2022

Took a look at my spam folder for the first time in a long while. One thing that struck me was how *visually* obvious the spam message titles are these days. I wonder if the spam detection software might be using vision ML models on titles - I know I would try it.

Shared by Bojan Tunguz at 2/23/2022

Adversarial Attacks on Deep Neural Networks: an Overview - https://t.co/ZUQ1k5gTpS https://t.co/kU1Imy1Y7I via @rightrelevance thanks @kirkdborne

Shared by Bojan Tunguz at 2/20/2022