Geek blog series - part 1
Our SAS experts have a treasure of insights and knowledge on the latest techniques and developments in AI and Analytics topics like computer vision, the interpretability of algorithms, the monitoring of model performance, the governance of AI models and much more. In this ‘Geek blog series’ these ‘Geeks’ share their expertise and insights. A must-read for everyone who likes to learn more about data science! Below you find the first two topics of this series.
Legal doping for your AI model training
By Jos van Dongen
Deep Learning, computer games and bitcoin mining have one thing in common: they all benefit from a beefy GPU in your computer. If you’re not familiar with the acronym: it’s short for Graphical Processing Unit, as opposed to the CPU (Central Processing Unit) that’s more common. Make no mistake: every computer has a GPU, but most regular computers have a GPU embedded in the CPU which is fine for displaying websites, watching cat video’s on YouTube or using general office tools. Dedicated GPU’s however are essential if you like to play modern video games. What they are particularly good at is performing extremely fast matrix calculations; the kind of processing required for rendering video and images. And it’s these same matrix calculation capabilities that turn out to speed up training your Deep Learning models as well!
To give you an impression of the kind of improvements dedicated GPU cards can deliver we built three different servers on Amazon Web Services (AWS). One with 16 regular (but fast) CPU cores, one with 16 CPU cores and an additional Nvidia Tesla M60 GPU (nickname ‘Laurel’), and a big machine with 32 CPU cores and no less than 4 Nvidia Tesla V100 GPU’s (nickname ‘Hardy’). Then we started training the model used for the Geek/No Geek app (only works on a smart phone!) presented at the Dutch and Belgian SAS Forums.
The regular machine was no match for both Laurel and Hardy. Laurel, with just one GPU and a cost of about $1,27 per hour gave a tenfold increase in performance. Hardy on his turn again gave a tenfold increase in performance compared to Laurel, but, at a cost of $12,37 per hour, was also 10 times as expensive. Still, being able to train a model in only 1% of the time it would normally take is well worth the extra outlay, especially given the current hourly rates of AI specialists.
The conclusion? If you want to get started with image recognition and AI, don’t scrimp on GPU’s.
Privacy proof biometrics: user in control
By Berno Bucker
The use of biometrics has become very popular in a short time. This is not surprising: being recognized on the basis of your facial characteristics and thus gaining access to an office or stadium without a card, pass or tag; participating in a loyalty program or even making payments just by facial recognition. Who doesn't want that? Well, everyone who is afraid that their privacy is impacted by that.
That is why 20face has developed a method for privacy proof facial recognition. Although 20face has only been around for two years, the research into AI and privacy is already more than twenty years old. The company was founded by researchers from the University of Twente, who know all about this research and have sometimes even contributed to it. Using all this knowledge, we came to a system that can recognize people on camera images with 99.67 percent certainty, even in situations where the light is not optimal, the images move quickly or the person is rather far away from the camera. Our software can even distinguish fake faces from real ones.
How do we create a privacy proof model? By putting the user at the center. The user takes a photo of his face with our app. The app saves the facial features but deletes the picture. The user then indicates where he wants to be recognized, for example in his favorite store, in the football stadium of his club or in front of the entrance gate of his employer. The app makes contact with the access system and provides the facial features. In this way the visitor can be recognized without needing an admission ticket, tag or pass.
One of our first customers was the Amsterdam Arena, which recognizes season ticket holders via the app. We also work for Heracles (ticketless stadium), AS Watson Benelux (access to the head office), ICI Paris XL (loyalty system), the museum Beeld en Geluid in Hilversum (enhanced customer experience) and Technology Base (facility management).
Want to know more? Reach out to me via LinkedIn. We’re happy to come by for a demo.
SAS overtreft opnieuw alle advanced en predictive analytics leveranciers, volgens het IDC-rapport.
Ziekenhuis Gelderse Vallei maakt al ruim twaalf jaar gebruik van de analytische software van SAS. SAS Visual Analytics helpt met geavanceerde analyses waarmee het ziekenhu [...]
Digitale transformatie, wat is dat nu eigenlijk? De woorden worden te pas en te onpas in de mond genomen zonder het concreet te maken.
We praten veel over vaardigheden, en met name over het belang van het ontwikkelen van nieuwe vaardigheden om breder en beter gebruik van analytics mogelijk te maken. [...]
Tijdens de eerste SAS Data Science & Analytics Day met als thema ‘Welcome to the Analytics Economy’ werd duidelijk dat data inmiddels onmisbaar is in onze samenl [...]
Slim en strategisch asset management inrichten bij spoorbedrijven? Lees hoe DEKRA Rail datavisualisatie toepast.
SAS heeft nieuwe functionaliteiten en producten toegevoegd aan het SAS Platform. Suez, Euramax en Notilyze delen de eerste ervaringen met h [...]
Maar liefst 380 journalisten hebben sinds begin dit jaar 13 miljoen documenten uit de Paradise Papers onderzocht. Lees hoe de inzet van data science en machine learn [...]
De civiele infrastructuur in Nederland komt steeds verder onder druk te staan. We moeten daarom slim omgaan met ruimte en assets efficiënt inzetten. Maa [...]