Jeremy Howard Profile picture
🇦🇺 Distinguished research scientist: @usfca; Co-founder: https://t.co/ZvDGNlehRt; Chair: https://t.co/M3OlNFAfuY. Masks4All(retired)
Tricia Profile picture Ross Grayson, MPH, CIH Profile picture Remco Frijling Profile picture Angela Rogers Profile picture Aswath Rao Profile picture 6 added to My Authors
22 Jul
I couldn't be happier with my little standing desk setup. I have tried far to many products over the years, and here's what I highly recommend:
1/ Image
Ergotron standing desk. It's a simple mechanical design, but it is really easy to move up and down.
The new Sony ZV-1 vlogging camera. It's connected to the computer via an Ansten HD capture card.

It has a nice color balance and good bokeh. It's pretty much unique as a zoom and vlogging camera at this point. The screen flips so you can see yourself. Image
Read 10 tweets
21 Jul
Last week I presented to the WHO Guidelines Development Group. It was targeted towards infectious disease epidemiologists, but many folks have asked me to share the presentation.

So, here's a video discussion. Read the thread below to see each slide. 1/
The evidence that WHO relies on for its mask guidelines has many problems. For instance, the meta-analysis they sponsored only has 3 non-healthcare studies, 1 of which is wrongly categorized, and one of which is under-powered. None studied COVID-19.
The review had three studies of SARS-CoV-2 in healthcare, but largely ignored them, by using inappropriate statistical techniques. (See the tiny weights in the right-most column.)
Read 27 tweets
15 Jul
I was invited to present the paper we've worked on for four months to @WHO's Guideline Development Group yesterday.

Now that the world is wearing masks, I'm retiring from mask advocacy.

Here's a thread on how #Masks4All got to this point...
preprints.org/manuscript/202…
In early March, our community in SF wasn't, on the whole, taking COVID-19 at all seriously. @math_rachel & I looked at the data, and we were worried it was about to blow up.

So we wrote an article saying so.
fast.ai/2020/03/09/cor…
We didn't expect it our article make much of a difference, frankly. We're just data scientists, in our niche area of deep learning. Our most popular post, introducing @math_rachel's Natural Language Processing course, is our only post ever that has topped 100k unique readers
Read 70 tweets
12 Jul
"Face Masks Against COVID-19: An Evidence Review", the most viewed paper on preprints.org of all time, is now even better, thanks to our wonderful reviewers and editor.

With 146 references, this is the most comprehensive study of masks today.
preprints.org/manuscript/202…
Huge thanks too all my great co-authors, @arimoin, @larrychu, @zeynep, @lexfridman, @austinvhuang, @hernandez_danny, @arnedelft, @HeleneMarivdW, @AmyPricePhD, @reshamas, Z. Li, C. Ramirez, L. Tang, V. Zdimal, C. Bax, G. Watson & V. Tang.
The extensive coverage of so many topics in this paper is only possible because of the breadth of expertise of our international author team, including epidemiology, biostatistics, aerosol science, sociology, infectious disease, computational modeling, data science, and more.
Read 7 tweets
19 Jun
So do these non-medical masks we're all meant to be wearing actually do anything?

Well yes, yes they almost certainly do. Here's a thread about the current science.

Let's start by explaining this striking picture of a researcher speaking in a laser scattering chamber.
1/n
The picture above shows speech droplets, which are believed to be the key transmission vector for COVID-19, with vs without a paper towel face cover. As you see, nearly all droplets are blocked.

Here's a video with a cloth covering, from this NEJM paper: nejm.org/doi/full/10.10…
The reason we look at speech is because we now know that COVID-19 results in viral shedding in the upper respiratory tract, where speech can result in droplet formation.

Few asymptomatic patients show significant lower respiratory tract shedding
nature.com/articles/s4158…
Read 22 tweets
8 Jun
New analysis.

Texas re-opened early, without requiring masks. Things are looking bad.

NB: chart shows hospitalizations, not cases, so this is not a side-effect of increased testing.
Source data from Texas DSHS.
dshs.texas.gov/coronavirus/ad…
This chart is provided under creative commons attribution 3.0 license. So feel free to use and share! :)
creativecommons.org/licenses/by/3.…
Read 5 tweets
17 May
What's the best way to animate in real time a plot of a hundred thousand points or more in jupyter with python? Is there anything with a WebGL backend or similar? Or at least canvas?

(I'm looking to visualize a large epidemiological model.)
Only 2d is needed BTW
Read 2 tweets
23 Apr
Want to take your mask wearing to the next level, and maximize the protection to yourself, in addition to protecting those around you? Here's some tips... 1/
Protecting yourself is harder than protecting those around you, because the droplets which are ejected when unmasked people speak evaporate quickly and are hard to filter.

Here's a mask my mother-in-law made with some great protective features.
There's a small long pocket at the back which is open at both ends. I use a small piece of rolled up aluminum foil in it. This is used to provide a moldable nose clip which helps minimize the gaps around my nose, like a surgical mask.
Read 8 tweets
20 Apr
I've studied over a hundred research and policy papers regarding masks, and used them to answer the most common 17 questions I've seen asked by curious skeptics. Here's the result!

See below for a quick skim of a few to whet your appetite.
1/
fast.ai/2020/04/20/ske…
Why should most people wear masks?
Shouldn’t only sick people wear masks?
Read 16 tweets
19 Apr
A paper in @AnnalsofIM claimed "cotton masks seem to be ineffective in preventing" COVID-19 transmission.

However, the paper made a serious data analysis mistake, which, when corrected, shows cotton masks blocked >95% of the viral load!

I'll explain. 1/
annals.org/aim/fullarticl…
The paper is the first to report results of real source control experiments with COVID-19 patients, so it's important. It only used 4 patients, which isn't great, but it's a fairly simple mechanical test so hopefully the results will hold in larger samples.
Here's the key table. Note that these are logarithms (base 10), so we need to "undo" that with 10^x to get the actual viral load. "ND" means "not detected"; ie the viral load was zero, or too small to be detectable. As you see, 2 times the cotton mask blocked so much it was "ND"
Read 11 tweets
13 Apr
We've just completed a 19-author analysis of the effectiveness of mask wearing, with 84 references. To explain what the science shows, I teamed up with the wonderful Prof @trishgreenhalgh CBE, who just led a British Medical Journal study on this. (thread)

fast.ai/2020/04/13/mas…
Our team's review of the literature found substantial evidence in favor of widespread mask use to reduce community transmission, based on droplet dynamics, mask material analysis, efficacy studies, and behavioral studies. Here's our paper: up.fm/masks
The key insight is that most discussions assume that the purpose of the mask is to protect the wearer, since this is what all doctors learn about in medical school. But actually masks work *far* better at blocking the infection at the source. This is called "source control"
Read 15 tweets
1 Apr
Guess which of these countries instituted #Masks4All recently.

Wear a home made mask whenever you go out. Get everyone you know to do the same. Protect each other. See Masks4all.co for science and guides.
h/t @ducha_aiki for this chart.
Which country do you want to be?
Read 3 tweets
31 Mar
We now have a website for #Masks4All!

Get the scientific evidence, quotes from experts, DIY recipes for homemade masks, media coverage, and everything you need to know about why we should all make a mask, and wear it when in public.
Masks4All.co
Huge thanks to @CamWoodsum and @mishachellam for their amazing work on this.
And most importantly, thank you to @PetrLudwig, without whom none of this would have been possible.
Read 3 tweets
25 Mar
The Czech Republic went from zero mask usage to 100% in 10 days, and in the process they halted the growth of new covid-19 cases.

How? They made their own! They didn't need government help; they did it themselves.

It's time for #masks4all. See why:

1/
Please watch the video above for the full explanation. Here's a quick summary.

See the light turquoise lines in this chart? See how they're all pretty flat? Know what they have in common? These folks all wear masks in public!
Masks aren't enough on their own. But they are a part of the solution in *every* region that's figuring out how to beat this thing!

(Of course, you can't even *begin* to solve this until you've done a lock-down long enough to buy your hospital system time.)
Read 24 tweets
24 Mar
This is one of the most amazing and hopeful things I've read. How the Czech Republic, through community effort, got to 100% mask usage in 10 days, nearly all thru home DIY effort.

They only have 2 deaths and have no growth in daily new covid-19 cases.
docs.google.com/document/d/1EW… Image
C'mon everyone, let's do this! Image
I'm going to record a video on this issue tonight. Will release tomorrow. Hope you all can help me share it then.

I'm getting increasingly convinced this is the most important thing that we can do directly as a community. (But we also need test&trace!)
Read 5 tweets
19 Mar
Folks are sharing this chart, from the paper "High Temperature and High Humidity Reduce the Transmission of COVID-19", & are saying that the claims of the paper are rubbish, because this isn't showing a strong correlation.

They're wrong. I'll explain
1/
papers.ssrn.com/sol3/papers.cf…
The chart does, indeed, not show a strong correlation. If this chart was the basis for the conclusion of the paper ("High Temperature Humidity Reduces Transmission") then it would, indeed, be rubbish.

But that's not the basis. The paper has bad visualization, not bad modeling.
The paper is actually based on a rather heroic effort of careful contact tracing and analysis of chains of 105 contacts across each of 100 cities. It's thanks to China's massive investment in tracing, starting Jan 20.
Read 14 tweets
18 Mar
Last night we started teaching our new deep learning course (virtually). We included a section about covid-19, since it's on all our minds.

The course won't be released until July. The class asked if we could release the covid-19 part now. Yes, we can!
1/
See the video for the 45 min version. But here's a quick summary of what we discussed. We focused on a practical, data-driven approach, since that's what we do.

We started with a summary of our article - so take a look at this first.
fast.ai/2020/03/09/cor…
There's also a short version as a tweet storm, if you're in a hurry.
Read 29 tweets
12 Mar
covid-19 is spreading exponentially. It takes weeks to find out if someone will recover. That means that at any point the vast majority of infected people are classified "unknown outcome".

To handle this, we need survival analysis - let's dive in!

1/
Survival analysis is used to handle "censored data". That means: data where we sometimes don't know the label yet.

Most data scientists aren't familiar with this technique. That's understandable, because there's lots of fields where it doesn't really come up. But we need it now!
Censored data can include analyzing life expectancy, since by definition anyone who is still alive is just "not dead yet"!

It can also be used for marketing conversion analysis, since prospects can be still in the pipeline. Here's a great intro
erikbern.com/2019/08/05/mod…
Read 14 tweets
10 Mar
I've seen many people claim that concerns about covid-19 are overblown, because the death rate in S Korea is "only" 0.7%. This number is statistically meaningless.

You see, if the true death rate was 5.0%, and cases increased 3x per week, S Korea would now show only 0.6%!

1/
The 0.7% comes from March 9th deaths so far (51) vs cases so far (7382). But remember, epidemics initially grow exponentially, and it takes a couple of weeks to find out the outcome for a patient.

Because of this, currently 98% of outcomes in S Korea are *unknown* !
2/
Therefore, the 0.7% rate only makes sense if we assume all 98% of the "unknown" group recover. This is, of course, absurd.

Here's what it actually looks like when we put the numbers into a spreadsheet.
3/
Read 9 tweets
9 Mar
As data scientists, we (with @math_rachel) have done our best to look at the data around covid-19, and what it means to you and your community.

Our view: it is appropriate to be very concerned, and significantly change your lifestyle, right now.
1/13

fast.ai/2020/03/09/cor…
The impact is already clear. In Italy, 10 days ago, all was fine. Now it's not. 432 medical tents have been set up. 16 million people are on lock-down.

By the time the impact in your community is clearly visible, you've missed your best opportunity.
2/
If you are in a position of authority, you should be doing everything you can to avoid the need for groups of people to get together (provide sick leave; make meetings virtual; cancel events or make them online; etc...)

For everyone else, here's a list of things to do:
3/
Read 13 tweets
27 Feb
Ever wondered what @PyTorch nn.Module and nn.Parameter do really? And how hooks actually work? Here's a working implementation from scratch of their key functionality, in one tweet!

From our upcoming book and course:
- usfca.edu/data-institute…
- amazon.com/Deep-Learning-…
Here's an example of using it to define a convolution layer:
Technical note: PyTorch tensors don't yet support sub-classing using standard Python (hopefully that'll be fixed soon!) So we have to use fastai's `TensorBase`, which does support that.
Read 6 tweets