Verifying Mission-Critical AI Programs | Two Minute Papers #179 financial deepmind

by Krabihometownbt1
banner



The paper “Reluplex: An Efficient SMT Solver for Verifying
Deep Neural Networks” is available here:

Out Patreon page:

Earlier episodes that were showcased:
pix2pix –
Breaking DeepMind’s Game AI System –

WE WOULD LIKE TO THANK OUR GENEROUS PATREON SUPPORTERS WHO MAKE TWO MINUTE PAPERS POSSIBLE:
Andrew Melnychuk, Christian Lawson, Dave Rushton-Smith, Dennis Abts, e, Eric Swenson, Esa Turkulainen, Kaben Gabriel Nanlohy, Michael Albrecht, Michael Orenstein, Steef, Sunil Kim, Torsten Reil.

Two Minute Papers Merch:
US:
EU/Worldwide:

Music: Antarctica by Audionautix is licensed under a Creative Commons Attribution license (
Artist:

Thumbnail background image credit:
Splash screen/thumbnail design: Felícia Fehér –

Károly Zsolnai-Fehér’s links:
Facebook →
Twitter →
Web →

Images related to the topic financial deepmind

Verifying Mission-Critical AI Programs | Two Minute Papers #179

Verifying Mission-Critical AI Programs | Two Minute Papers #179

Search related to the topic Verifying Mission-Critical AI Programs | Two Minute Papers #179

#Verifying #MissionCritical #Programs #Minute #Papers
Verifying Mission-Critical AI Programs | Two Minute Papers #179
financial deepmind
See all the latest ways to make money online: See more here
See all the latest ways to make money online: See more here

You may also like

24 comments

Jacobus Strydom 09/09/2021 - 7:17 Chiều

O wow thanks for this man, it's clear you do this for the love of science

Reply
Arjun 09/09/2021 - 7:17 Chiều

Thanks this is amazing.

Reply
o 09/09/2021 - 7:17 Chiều

A neved hallattáig azt hittem, indiai akcentust hallok, bocs.

Reply
Andreas Lindhé 09/09/2021 - 7:17 Chiều

Great video! I enjoyed this much more than all the new graphics rendering engines.

Reply
smuecke 09/09/2021 - 7:17 Chiều

The problem with Neural Networks in critical systems is that humans can't comprehend how they make their decisions. If a network makes a mistake, it is very hard to determine what exactly caused the problem because there is no fixed set of rules by which decisions are made. This is the real issue with NNs, and IMO this is why NNs are not the ultimate solution to all optimization problems, which seems to be the general opinion in recent years.

Reply
Tristan Wegner 09/09/2021 - 7:17 Chiều

I found the non relevant video very distracting. I get that youtbe rewards great thumbnails, but for the content just a static screenshot of the paper would be better.

Reply
Julien Scholz 09/09/2021 - 7:17 Chiều

Could you perhaps talk about the StarCraft 2 AI API that DeepMind has been working on? Great video as always!

Reply
TWPO 09/09/2021 - 7:17 Chiều

I love this channel so much. Thank you for providing such quality content

Reply
oguretsagressive 09/09/2021 - 7:17 Chiều

Why are these noise attacks even possible? I mean, humans can see false patterns in the noise too, but these tiny features don't have a higher priority over bigger ones. What are we doing wrong?

Reply
daniel f 09/09/2021 - 7:17 Chiều

This is a great one, I know that alot of autonomous vehicle companies cannot really benifit from ML because of unpredictability, this can be a game changer in the autonomous vehicle industry in the future

Reply
Mopic3d 09/09/2021 - 7:17 Chiều

So far all of the classifiers operate on tiny chunks of the images and their relative differences. We can easily see a school bus, with or without the adversarial input (which to us, humans looks like noise by the way). I don't know if it's even possible, based on 2D images alone, to create a system that would actually "understand" the basic structure and model of objects (just like we have a 'model' of a bus in our minds) and would be able to recognize them on this macro-level. I can recognize a cat because I know it has four legs, a tail and a head with pointy ears, I don't stop at the level of characteristic gradients and edges.

Reply
Harley Mckee 09/09/2021 - 7:17 Chiều

super cool

Reply
killroy42 09/09/2021 - 7:17 Chiều

Even my earliest ANNs, more then 20 years ago were trained with added random noise. Would that not harden them against adversarial inputs?

Reply
h0lyRS 09/09/2021 - 7:17 Chiều

Could an AI be trained to translate any given neuron network to a fixed mathematical function?

Reply
Abe Dillon 09/09/2021 - 7:17 Chiều

This is amazing! Just yesterday, I was discussing the problem of unpredictable failure modes in deep learning aircraft navigation systems in another comment thread on Youtube: https://www.youtube.com/watch?v=gB_-LabED68&lc=z12giviohl3hvlusb22agplbaxivip1ax04.1502588991481293

Another fantastic video! Thank you so much!

Edit: I love the internet!

Reply
Ram Krishan Charan 09/09/2021 - 7:17 Chiều

Please make another video about other YouTube channel and website suggestions by you

Reply
Emile 09/09/2021 - 7:17 Chiều

love your videos!

Reply
Mel Florance 09/09/2021 - 7:17 Chiều

Awesome to see that Neural Network will become more stable !

Reply
Bronn 09/09/2021 - 7:17 Chiều

Thanks for this one

Reply
foobargorch 09/09/2021 - 7:17 Chiều

watch?v=5WVu798pmRc

Reply
Zhengqun Koo 09/09/2021 - 7:17 Chiều

This is the kind of papers that I like!

Reply
Downstream01 09/09/2021 - 7:17 Chiều

Would selective Gaussian blur, that you can find in GIMP for example, help against that noise attack?

Reply
Einstein Cross 09/09/2021 - 7:17 Chiều

Amazing as always!

Reply
Wellington Boobs 09/09/2021 - 7:17 Chiều

Microsoft's AI apparently spoke in public like donald trump does in private.

Reply

Leave a Comment