A talk from the Inspire Track at AWE USA 2017 - the largest conference for AR+VR in Santa Clara, California May 31- June 2, 2017.
Leila Janah (Samasource): Better Algorithms, Better Lives; Alleviating Poverty Through Training Data
Think that tech advances will always result in job automation and leaving behind those without opportunity? People may not realize that artificial intelligence is still powered by human intelligence- advances in AI can actually create jobs: think of training data. Producing quality training data is critical in ensuring the AI experience works, but expertly annotating hundreds of thousands of images necessary to create quality algorithms requires massive human effort. However, outsourcing this work through traditional channels misses a major opportunity-- the chance to impact lives. Instead, let's connect the 1.3 billion people living in extreme poverty around the world to cutting-edge innovation in the AR space. Let’s leverage the talent pool of global human intelligence to advance technology sustainably. Let’s ensure our automated cars and VR headsets don’t isolate but instead connect us more deeply to each other.
http://AugmentedWorldExpo.com
Let me take you on a little journey that seems like it has nothing to do with AR… but you’ll soon see how it does
This is Ken Kihara
Ken’s life pre-Samasource [Mathare]
Like Ken, 50% of the African population lives on less than $2/day [50%] - that’s a lot of people, and a lot of human talent that goes wasted
But, things are changing: In 2002, the first fiber optic cable came to Africa, improving the infrastructure and connectivity.
By 2009, east and southern africa have over 10k miles of cable (stat from here: http://www.nytimes.com/2009/08/10/business/global/10cable.html)
Facebook announced earlier this year plans to lay 500miles of cable in Uganda. The largest fiber optic provider across Africa, Liquid Telecom, has alone laid over 18,000km of cable across the continent in 2017.
The technology sector benefits: Services like call centers can offer more competitive rates because of lower operating costs, and technology companies can communicate better with clients and partners overseas. Now these people living in poverty have an opportunity to work. And luckily, there’s a lot of work that needs to be done:
CV requires accurate, high-quality data… and lots of it… to properly train the deep learning neural network for the specific application. Capture and especially annotation is done manually,
14M+ images in the open source ImageNet database. Target 100s – 1000 images per class…. Could be 100K – 1M+ images for the target application
** So… how is this annotation done?
We need lots of talented humans to develop and manually annotate large sets of training data
An image like this took 2 hours for 1 person to annotate
Each dot also has labels behind it.
4 of these images is 1 day’s wage.
1,500 images (365x4) = 1 year’s worth of work for someone living in poverty.
Imagine what 1 million annotated images could do. And that is just per CV application
So, what is this training data actually powering? What applications use them?
What does this combination of effort and cutting edge technology bring… let’s look at some use cases:
Augmented Driving
An area of incredible recent investment and growth has been in Mobility. While autonomous vehicles are grabbling headlines, they are still many years away from widespread availability. In the meanwhile, helping drivers get to their destinations more safely is paramount to both car providers and legislators. Next gen cockpits will include intelligent displays that combine imaging, computer vision, and HD maps to recognize and communicate directions and hazards to the driver in a seamless and intuitive way.
Per Samasource value prop, the same requirements for accurate tagging of training data in automotive apply here, as the ADAS/autonomy system is being leveraged to properly overlay the data on the windscreen.
Augmented surgery:
AR is starting to have an major impact in medicine, providing students with powerful training tools and aid doctors with an unprecedented awareness of the body for more effective surgical procedures. These systems combine highly accurate tagged medical imaging data with 3D computer vision to precisely overlay information on the body so the surgeon can navigate to the right spot for treatment in real time. Even more so than Mobility, there is no room for error as lives are at stake.
Per Samasource value prop, high accuracy of image annotation is critical, so well defined processes and trained personnel is key
Who would have thought that the Hololens was partly built by people living in the slums of Nairobi? [Hololens]
Training data is the cornerstone of these CV applications, but also can be leveraged to lift people out of poverty.
That’s what we do at Samasource….we provide high quality ground truth training data, while also lifting people out of poverty (by giving work)
We’re a sustainable (you heard that right...sustainable. One of the few in the Valley ;)) social enterprise who believes in the power of training data to revolutionize not only our technology, but the lives of people living in poverty.
We’ve completed over 500 projects for over 100 clients, and paid out $10M in wages to people who who were previously living below the poverty line. [stats]
And, our quality rates are actually higher than the industry average (95%). [98%]
Since 2008 we have lifted over 36k ppl out of poverty by providing them with stable employment and living wages