Arnaud Dhaene (@arnaudhaene) profile picture

MSc Thesis @Stanford | Computational Life Science @EPFL

Palo Alto, CA
Joined November 2017
I always felt that some of the functional tools that come with standard Python were insufficient and awfully verbose.
1
This is why I decided to create a little wrapper around the emblematic Python list that allows calling {filter, map, reduce, flatmap} on the list element itself. It's definitely a WIP, but my lists are now unencumbered and my code (more) readable!
1
All you need to do is wrap your original list object in a FuncList object. gist.github.com/arnauddhaene…
Arnaud Dhaene retweeted
Unpopular data science opinion: no body uses dashboards
48
88
80
914
Arnaud Dhaene retweeted
Give fewer choices, get more action.
35
225
24
2,228
What's the probability that you are infected with COVID-19 given that your RT-PCR test came back positive? 🤔 Hint: it's lower than you think! Let's find out using Bayes' Theorem👇
1
For the sake of the calculation, let us assume a 95% true-positive and a 2.3% false-positive rate based on the paper below (UK). Let us use the current infection rate in the UK: 273 per 100k residents (7-day rolling average) means P(C) = 0.00273 👇 coronavirus.data.gov.uk/
2
P(C | T) = 0.95 / (0.95 * 0.00273 + 0.023 * 0.99727) * 0.00273 P(C | T) = 0.1015 or 10.15% Putting all of this together, we find that the probability you are infected given a positive test result is only 10.15%! Wait, is this true? How can this be? 👇
1
People are not getting tested randomly! Most people who get swabbed have a reason, they might display symptoms, recent contact, etc. This increases the probability of getting tested, AND of being infected! Lesson learned: never forget confounders! 🕵️‍♂️
1
Arnaud Dhaene retweeted
In case you wondered what made Pfizer quickly deliver 10M vaccine doses to Israel: "Israel will provide details to Pfizer [...] about the age, gender and medical history of those receiving the jab [...] No identifying information will be given" Hmm... politico.eu/article/israel-c…
1
9
2
21
This week in the @VisiumAI Digest, I discuss privacy considerations in Large Languages Models, DALL·E, and the first-ever deployment of a Deep Learning tumor contouring solution in a Hospital. Direct link: https://nitter.cf/t.co/hOZlmul2Sw More info in the 🧵 below
1
(2) OpenAI has released a new text-to-image synthesizer called DALL-E. The combination of a GPT-3 variant with a VQ-VAE model shows extremely promising results. Waiting for the paper to get juicy theoretical details and to know how the data was gathered. openai.com/blog/dall-e/
1
(3) InnerEye has been deployed in a Cambridge Hospital to help speed up cancer treatment. The Deep Learning model takes a CT scan and is able to perform object segmentation and classification for tumors and healthy neighboring tissue. news.microsoft.com/en-gb/202…
Do large language models just memorize the data they’re trained on? 📚 A new paper evaluates adversarial methods that are able to extract information from the training data! youtu.be/plK2WVdLTOY
2
(2/3) The article above explains that there still exists a lack of transparency in AI research. A number of big corporations such as Google, DeepMind, and OpenAI are publishing ground-breaking papers using their private access to world-class hardware and proprietary data.
1
(3/3) This makes it difficult for researchers to reproduce and build upon these papers—going against the groundwork of scientific research. While advances have pushed back against this issue—such as arXiv's collaboration with PapersWithCode—there is still a long way to go.