How AI Can Save 35,000 Lives Every Year
Liran Zvibel. November 21, 2017
I just got back from attending the GPU Technology Conference in Washington DC where we heard about some of the latest innovations in NVIDIA performance GPU technology including smart cities, healthcare, virtual reality and finance predictive analytics. But probably the most moving and thought provoking speech came from US Senator Gary Peters, Michigan. In 2016 alone, 37,461 people died from automobile accidents and 94% of them caused by human error. So how come a death toll that has been at national epidemic level for decades does not have resolution?
Well the answer is obvious, if 94% of the problems are caused by human error then remove the human. Which leads to the next question – what is holding back immediate action to replace humans with autonomous cars? The challenge has been that while the technology for artificial intelligence was documented in the 1970’s under the research wing of luminaries like Marvin Minsky, the practical use of neural networks required some key industry developments to make it practical. You need massive computational power, massive data sets, a way to feed the data to the computational layer and a political will to make it all happen.
We solved the data crunching with massively parallel computation. Traditional CPU architecture has relied on very high performance cores to process data, but it has been prohibitively expensive to scale for general day to day problems. The development of GPU accelerated computing allowed scientists to offload compute intensive pieces of an application to low cost GPUs, where thousands of smaller, more efficient cores execute the application in a massively parallel way. The result has been dramatically faster time to results in an affordable cost model.
We solved the data volume problem by generating and storing massive training sets. Remember when storing data used to be expensive? In 1990 it cost $10,000 to store 1 GByte of data on a hard drive, today it cost less than 5 cents. When data was expensive to store, it simply was not stored. The amount of data stored has grown exponentially as the cost of storing has fallen exponentially. Artificial intelligence (AI) systems are only as smart as the amount of data that they have been trained on and it took a critical mass of data to reach a point where AI models can consistently outperform human. We saw this with Watson on the Jeopardy show.
We have the data and the GPUs that can crunch the data, the new challenge is one of data accessibility. Put simply, GPUs are expensive and you cannot afford to have them idling. The data used in training systems has to be sharable across the many GPU systems simultaneously and the computation requires a super fast way to ingest tens of terabytes of data at a time. Legacy storage systems are really good at doing one or the other task, but until WekaIO Matrix came on the market, no system could do both. The software has been written from scratch to run natively on flash providing very high read performance across a shared file system. It has an in-built low latency network stack ensuring that data from any node can reach its intended destination in a few hundred microseconds. The result is the highest performance, scalable file system for data intensive applications such as vehicle training systems.
WekaIO Matrix can deliver over 6GByes/second to a single GPU client, outperforming a local NVMe drive, because we parallelize the data ingest across multiple storage nodes. The solution will scale as the level of GPU processing scales with fastest time to results.
The last remaining obstacle to saving over 35,000 lives per year unfortunately is also in the hands of human beings, namely Government legislation to allow the deployment of self-driving vehicles in the USA. Unfortunately, that barrier may be far greater than any of the technology advances that have made the elimination of this epidemic a possibility.