Dr. Heloise Stevance
@sydonahi.bsky.social
Astrophysicist - Schmidt A.I. in Science Fellow - Oxford Uni - she/they
Sky surveys
Exploding stars
Hungry black holes
AI Literacy
Beatrice Tinsley Lecturer 2021
Caroline Herschel Prize 2024
hfstevance.com
Sky surveys
Exploding stars
Hungry black holes
AI Literacy
Beatrice Tinsley Lecturer 2021
Caroline Herschel Prize 2024
hfstevance.com
September 10, 2025 at 9:13 AM
August 11, 2025 at 8:59 AM
@eso.org has this really cool paper model of the Extremely Large Telescope and I just finished mine! I got the kit at EAS but you can get it for free with instructions online! Thanks ESO ❤️🙏
www.eso.org/public/produ...
🔭🧪
www.eso.org/public/produ...
🔭🧪
June 30, 2025 at 6:29 PM
@eso.org has this really cool paper model of the Extremely Large Telescope and I just finished mine! I got the kit at EAS but you can get it for free with instructions online! Thanks ESO ❤️🙏
www.eso.org/public/produ...
🔭🧪
www.eso.org/public/produ...
🔭🧪
We would need 400 high res televisions to properly display the @vrubinobs.bsky.social images for our eyes to see the details
June 23, 2025 at 3:32 PM
We would need 400 high res televisions to properly display the @vrubinobs.bsky.social images for our eyes to see the details
June 23, 2025 at 3:29 PM
LET'S GO! @lsstuk.bsky.social @vrubinobs.bsky.social
June 23, 2025 at 3:00 PM
LET'S GO! @lsstuk.bsky.social @vrubinobs.bsky.social
My face when I'm trying to juggle two conference parallel sessions AND the vera rubin first look party happening this afternoon
🔭
🔭
June 23, 2025 at 11:54 AM
My face when I'm trying to juggle two conference parallel sessions AND the vera rubin first look party happening this afternoon
🔭
🔭
@eas-meeting.bsky.social here we gooo!
June 22, 2025 at 7:36 PM
@eas-meeting.bsky.social here we gooo!
Landscape pictures, I found, are better when I perfectly time a selfie with my mum taking a picture of the view in the background.
I've found I rarely keep "perfect" pictures.
I always keep the ones that have life.
(Also thank god for my nose which is truly weight bearing massive shades)
I've found I rarely keep "perfect" pictures.
I always keep the ones that have life.
(Also thank god for my nose which is truly weight bearing massive shades)
June 18, 2025 at 4:20 PM
Landscape pictures, I found, are better when I perfectly time a selfie with my mum taking a picture of the view in the background.
I've found I rarely keep "perfect" pictures.
I always keep the ones that have life.
(Also thank god for my nose which is truly weight bearing massive shades)
I've found I rarely keep "perfect" pictures.
I always keep the ones that have life.
(Also thank god for my nose which is truly weight bearing massive shades)
Early bus to Heathrow to head to cork for @eas-meeting.bsky.social ! Heading a few days early to visit a bit of Ireland with my mom before hand ☘️
Best thing is I managed to clear my to-do list before the holidays ✨
Can't wait for all the science next week!
Best thing is I managed to clear my to-do list before the holidays ✨
Can't wait for all the science next week!
June 18, 2025 at 4:09 AM
Early bus to Heathrow to head to cork for @eas-meeting.bsky.social ! Heading a few days early to visit a bit of Ireland with my mom before hand ☘️
Best thing is I managed to clear my to-do list before the holidays ✨
Can't wait for all the science next week!
Best thing is I managed to clear my to-do list before the holidays ✨
Can't wait for all the science next week!
Hope the #nokings protest go well today if you're going 🤞
The UK was sunny even in the north and we marched as well but to protest another authoritarian government.
Hundreds of people, including families, chanted for #Palestine 🇵🇸
Stay safe
The UK was sunny even in the north and we marched as well but to protest another authoritarian government.
Hundreds of people, including families, chanted for #Palestine 🇵🇸
Stay safe
June 14, 2025 at 3:29 PM
Hope the #nokings protest go well today if you're going 🤞
The UK was sunny even in the north and we marched as well but to protest another authoritarian government.
Hundreds of people, including families, chanted for #Palestine 🇵🇸
Stay safe
The UK was sunny even in the north and we marched as well but to protest another authoritarian government.
Hundreds of people, including families, chanted for #Palestine 🇵🇸
Stay safe
Yes thanks for noticing🤩!
I made a color palette for my plots which works for as many colorblindness filters I could find and both with light and dark backgrounds!
Here's what they look like in different filters:
I made a color palette for my plots which works for as many colorblindness filters I could find and both with light and dark backgrounds!
Here's what they look like in different filters:
June 13, 2025 at 4:39 PM
Yes thanks for noticing🤩!
I made a color palette for my plots which works for as many colorblindness filters I could find and both with light and dark backgrounds!
Here's what they look like in different filters:
I made a color palette for my plots which works for as many colorblindness filters I could find and both with light and dark backgrounds!
Here's what they look like in different filters:
One way to test this meaningfully is to show the VRA weird/rare transients we CARE A LOT ABOUT: That's the green stars there - AT 2018cow, SN 2023zaw, SN 2023ufx - things you DON'T WANT YOUR BOT TO MISS!
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
June 13, 2025 at 2:37 PM
One way to test this meaningfully is to show the VRA weird/rare transients we CARE A LOT ABOUT: That's the green stars there - AT 2018cow, SN 2023zaw, SN 2023ufx - things you DON'T WANT YOUR BOT TO MISS!
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
Another day another brag about how I tested the crap out of my ATLAS Virtual Research Assistant.
Here' my FAVOURITE PLOT of the past 2 years.
So, the VRA has a threshold beyond which it doesn't show alerts to humans - but ML can be bad at what it's not seen before. What if it messes up!??
🔭
1/3
Here' my FAVOURITE PLOT of the past 2 years.
So, the VRA has a threshold beyond which it doesn't show alerts to humans - but ML can be bad at what it's not seen before. What if it messes up!??
🔭
1/3
June 13, 2025 at 2:37 PM
Another day another brag about how I tested the crap out of my ATLAS Virtual Research Assistant.
Here' my FAVOURITE PLOT of the past 2 years.
So, the VRA has a threshold beyond which it doesn't show alerts to humans - but ML can be bad at what it's not seen before. What if it messes up!??
🔭
1/3
Here' my FAVOURITE PLOT of the past 2 years.
So, the VRA has a threshold beyond which it doesn't show alerts to humans - but ML can be bad at what it's not seen before. What if it messes up!??
🔭
1/3
One way to test this meaningfully is to show the VRA weird/rare transients we CARE A LOT ABOUT: That's the green stars there - AT 2018cow, SN 2023zaw, SN 2023ufx - things you DON'T WANT YOUR BOT TO MISS!
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
June 13, 2025 at 2:32 PM
One way to test this meaningfully is to show the VRA weird/rare transients we CARE A LOT ABOUT: That's the green stars there - AT 2018cow, SN 2023zaw, SN 2023ufx - things you DON'T WANT YOUR BOT TO MISS!
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
They're all* WELL within the threshold!
This plot is both a mix pride and RELIEF🥳
2/3
LESSON 3
Histogram Based Gradient Boosted Classifiers are the goat.
This algorithm creates your scoring function by starting with a 💩 guess and progressively learning corrections.
It's quite robust to over fitting and the histogram version natively handles NaN values and categories!
11/14
Histogram Based Gradient Boosted Classifiers are the goat.
This algorithm creates your scoring function by starting with a 💩 guess and progressively learning corrections.
It's quite robust to over fitting and the histogram version natively handles NaN values and categories!
11/14
June 12, 2025 at 7:11 AM
LESSON 3
Histogram Based Gradient Boosted Classifiers are the goat.
This algorithm creates your scoring function by starting with a 💩 guess and progressively learning corrections.
It's quite robust to over fitting and the histogram version natively handles NaN values and categories!
11/14
Histogram Based Gradient Boosted Classifiers are the goat.
This algorithm creates your scoring function by starting with a 💩 guess and progressively learning corrections.
It's quite robust to over fitting and the histogram version natively handles NaN values and categories!
11/14
LESSON 2
Use metrics that match your science case!
My goal is to rank the extra galactic alerts at the TOP of the list, so I'm using something called the Recall and rank k.
The plot below shows we recover 90% of the good alerts in the top 10% of the ordered list!
9/14
Use metrics that match your science case!
My goal is to rank the extra galactic alerts at the TOP of the list, so I'm using something called the Recall and rank k.
The plot below shows we recover 90% of the good alerts in the top 10% of the ordered list!
9/14
June 12, 2025 at 7:11 AM
LESSON 2
Use metrics that match your science case!
My goal is to rank the extra galactic alerts at the TOP of the list, so I'm using something called the Recall and rank k.
The plot below shows we recover 90% of the good alerts in the top 10% of the ordered list!
9/14
Use metrics that match your science case!
My goal is to rank the extra galactic alerts at the TOP of the list, so I'm using something called the Recall and rank k.
The plot below shows we recover 90% of the good alerts in the top 10% of the ordered list!
9/14
So the models are two binary classifiers that ask the same questions as the Eyeballers:
1. Is this real?
2. Is this Galactic?
Once the alert is scored in that space I use the new fangled technology of the PYTHAGORAS THEOREM to rank the alerts! The closer to the "good" corner the higher!
6/14
1. Is this real?
2. Is this Galactic?
Once the alert is scored in that space I use the new fangled technology of the PYTHAGORAS THEOREM to rank the alerts! The closer to the "good" corner the higher!
6/14
June 12, 2025 at 7:11 AM
So the models are two binary classifiers that ask the same questions as the Eyeballers:
1. Is this real?
2. Is this Galactic?
Once the alert is scored in that space I use the new fangled technology of the PYTHAGORAS THEOREM to rank the alerts! The closer to the "good" corner the higher!
6/14
1. Is this real?
2. Is this Galactic?
Once the alert is scored in that space I use the new fangled technology of the PYTHAGORAS THEOREM to rank the alerts! The closer to the "good" corner the higher!
6/14
Spoiler: I did it!
The in prod results of the VRA since the start of April are below. The auto garbage is what the VRA labeled as garbage automatically!
85% of the work is GONE!
And the rest is made easier by ranking the alerts for the Eyeballers. Let's get to the models.
5/14
The in prod results of the VRA since the start of April are below. The auto garbage is what the VRA labeled as garbage automatically!
85% of the work is GONE!
And the rest is made easier by ranking the alerts for the Eyeballers. Let's get to the models.
5/14
June 12, 2025 at 7:11 AM
Spoiler: I did it!
The in prod results of the VRA since the start of April are below. The auto garbage is what the VRA labeled as garbage automatically!
85% of the work is GONE!
And the rest is made easier by ranking the alerts for the Eyeballers. Let's get to the models.
5/14
The in prod results of the VRA since the start of April are below. The auto garbage is what the VRA labeled as garbage automatically!
85% of the work is GONE!
And the rest is made easier by ranking the alerts for the Eyeballers. Let's get to the models.
5/14
200 a day is manageable and so people would eyeball the data (we literally call these people Eyeballers) and classify it into bogus events🟥🟧, or real galactic event 🟨 or real EXTRA GALACTIC event 🟦 (good events for us).
My job? Automate as much of that eyeballing for us!
4/14
My job? Automate as much of that eyeballing for us!
4/14
June 12, 2025 at 7:11 AM
200 a day is manageable and so people would eyeball the data (we literally call these people Eyeballers) and classify it into bogus events🟥🟧, or real galactic event 🟨 or real EXTRA GALACTIC event 🟦 (good events for us).
My job? Automate as much of that eyeballing for us!
4/14
My job? Automate as much of that eyeballing for us!
4/14
Just gonna leave this gem here:
"Feed your family for a fiver! (Only £8!)"
"Feed your family for a fiver! (Only £8!)"
May 11, 2025 at 8:20 AM
Just gonna leave this gem here:
"Feed your family for a fiver! (Only £8!)"
"Feed your family for a fiver! (Only £8!)"
I hope we're about the agree that the guardian isn't that left and that the liberal left isnt radical enough BUT I SUSPECT WE'LL BE GOING THE WRONG WAY.
This is p4 of chapter 1. It's gonna be a long thread.
This is p4 of chapter 1. It's gonna be a long thread.
May 6, 2025 at 7:36 AM
I hope we're about the agree that the guardian isn't that left and that the liberal left isnt radical enough BUT I SUSPECT WE'LL BE GOING THE WRONG WAY.
This is p4 of chapter 1. It's gonna be a long thread.
This is p4 of chapter 1. It's gonna be a long thread.
This is the bottom of page 4 of the first chapter....
I messed up....
🤦🤦🤦🤦🤦
I messed up....
🤦🤦🤦🤦🤦
May 6, 2025 at 7:34 AM
This is the bottom of page 4 of the first chapter....
I messed up....
🤦🤦🤦🤦🤦
I messed up....
🤦🤦🤦🤦🤦