Yet another company has emerged, with the announcement of DeepMap, making huge strides in the world of self-driving cars. The web page and announcements mention that many of the engineers are former Google and Apple employees (both seem to make maps and directions a huge aspect of their products, with Google maps and Apple maps being used by a vast amount of people day-to-day), which is incredibly promising because these sorts of engineers are probably exquisitely insightful and have unique ideas that may not have been carried out at an already established company, or may have taken much longer to come to fruition.
From a basic standpoint, self-driving vehicles are built around taking whatever input the car has, and generating an output that relates to speed and direction (i.e. for a traditional driving image, what speed and direction should the car use to keep on a path to whatever destination and avoid crashing). Many of the startups doing this seem to be focused on using image-based deep learning systems as the underlying technology (typically, convolutional neural networks). These systems have made incredible strides in recent years along with the companies implementing these autonomous vehicles having made tremendous progress (think Google’s self driving car, or Tesla, who recently hired Andrej Karpathy as head of AI). There has also been a recent scramble for many other companies to enter the autonomous vehicle arena, and create competitive offerings (for instance the recent company Andrew Ng has been associated with) and even recent scandals such as the Google-Uber lawsuit. These events are signs that this technology is going to become incredibly commonplace very soon, and will be an integral technology that people will come to expect in day-to-day life, somewhat akin to smartphones.
One of the interesting things after looking into DeepLearning is that the company and it’s underlying technology seems to be heavily focused on LiDAR systems versus the approach that many other companies seem to be using with camera/image based mapping. LiDAR is different in that it depends on pulsing lights to create a representation of 3D surfaces around it (quite an oversimplification). While I’m not an expert in autonomous vehicles, I’m guessing that a combination of LiDAR-based and image-based approaches will make for the first true autonomous vehicle in that simply relying on one type of data is too dangerous when the stakes of self-driving cars carry huge implications for the technology companies behind them.
A continuously updating and dynamic system
After reading through the introduction post by the cofounders, James Wu and Mark Wheeler, I was stuck by the fact that the company raised a sizable amount of money for something in stealth, and also by the many novel explanations and ideas in the post. One of the ideas that struck me as incredibly profound is viewing maps not as a static image that may be useful to humans, but as a continuously updating and dynamic system that incorporates an entire data stack and is useful to a machine such as a self-driving car.
This may be incredibly obvious to people already in the autonomous vehicle industry, but as an outsider, it made me think that maps without a dynamic and huge data stack underlying them will not only be useless to autonomous vehicles, but perhaps even dangerous. The map that humans see and glean knowledge from compared to what would be useful to machines would be fundamentally different, and makes me curious about the implications in other realms that deep learning and “AI” will have (for instance in NLP and time series data).
Having actual competition among companies in the pursuit of a true self driving car (there is a 5 level SAE classification and while I am not entirely sure where Tesla or Google’s vehicles currently stand, no one has yet to achieve level 5, which is what would be necessary for autonomous cars to most likely replace large swaths of industry and have worldwide impact) is a revolutionary achievement from a technological standpoint as it creates and encourages companies to try new and novel approaches on a field that will likely never be fully ‘solved,’ but in need of constant and continuous improvement much like other technology fields. In conclusion, the technical ideas that DeepMap is bringing along with yet another company to push forward the prospects and chances of autonomous vehicles becoming commonplace is incredibly promising and something to keep an eye on. Hopefully the products and technology they claim to be working on will be as groundbreaking as they propose, and not just crash and burn out like many technology startups seem to do.
About the author
Graham Annett is an NLP Engineer at Kip (Kipthis.com). He has been interested in deep learning for a bit over a year and has worked with and contributed to Keras (https://github.com/fchollet/keras). He can be found on Github at http://github.com/grahamannett or via http://grahamannett.me.