Where are the opportunities in deep computing?
During the last week venture Capital investors Andreesson Horowitz shared a great article summarising the history of computing and where it appears to be heading as did the Economist. I’ve attempted to summarise some of the areas of opportunity that are highlighted in these articles.
With that out the way let’s explore a few application areas;
More human interfaces
Our ability to use computing power is still limited by awkward interactions using screens and keys. The application of deep learning is rapidly accelerating how well computers can understand human intent, but this is a tough area to play in as it involves a big element of human behavior and fashion (Google glass was just a bit too weird for most).
Opportunity: create alternative interfaces that don’t feel weird even if very application specific at first.
Most recently there has been a proliferation of chat based apps using Natural language Understanding within limited API driven domains to provide an easy mechanism to access services. As this begins to actually work reliably one could imagine it replacing a significant chunk of work that is currently handled over the phone which is pretty much the service offering behind anything that you can buy.
Opportunity: Look for service plays which could benefit from chat based interfaces or make it easy for others to add this facility to their service.
Voice communication is surely almost ready for prime-time, especially in limited domains like controlling devices (Vocal IQ was an interesting start-up in this space and recently acquired by Apple, another is of course Amazon Echo).
We might also start to see two way voice communication replacing human conversations but possibly more exciting is the ability to listen in and augment the human agent. For example a big investment house told me that one of their biggest challenges is that a client can phone up and expect an agent to have an opinion on an any one of hundreds of events that happened moments ago, which quite obviously requires a super human effort. Another clear area of opportunity is augmenting the process of creating code of complex documents, one could imagine the computer suggesting a template based on an initial conversation that is subsequently refined over further conversation.
Opportunity: Look for services where the agent could be vastly better if the machine could automatically pull information from the interaction.
Unfortunately we’re still a long way away from directly stimulating the brain but I wouldn’t be that surprised to see non-invasive highly specific stimulation of peripheral nerves (that run from existing senses) emerge within the next few years.
Self driving / flying
It’s pretty obvious that the days of driving or piloting are almost over as automated systems prove to be far better than humans. This opens up a raft of service opportunities alongside possibilities to improve the core tech (e.g. LIDAR is expensive, on board computing is power hungry). In may cases it also presents the possibility of replacing dangerous jobs such as power station inspection.
Opportunity: Look for dangerous or incredibly unpopular jobs or build kit that makes robots and drones more useful (such as the ability to perch temporarily).
Internet of Things
There seems to be a simple formula on this one: find things that can be optimised within industrial systems by actually knowing what’s going on in that system. However, it’s not necessarily so easy from a commercial perspective, platform plays already seem to be saturated (although it would be worth understanding whether current solutions are truly useful) whilst sensors are quickly commoditised. Where you can stand out is some real intelligence advantage that leads to significant cost savings. A good example is Inflowmatix (an Imperial Innovations investment) which helps to prevent damage from surges in the water pipe system. Another obvious area of interest is security around these connected devices.
Opportunity: Get to understand production systems and where there is room for optimisation — not easy, guest post coming soon.
Virtual / Augmented Reality
There’s some really obvious high value use cases and potential service businesses around AR / VR, from convincing remote presence to making humans more effective in industrial situations such as production or maintenance. From a technical perspective it’s still a bit odd, it requires a huge headset and often cameras and lights in the corners of the room not to mention the lack of physical feedback (although several experts have stated that this isn’t important I’m not convinced). The recent explosion in interest came about form Oculus’ ability to reduce the feeling of sickness and making it even more seamless could represent a wide array of opportunities.
There are still plenty of opportunities to make AR / VR feel more natural, especially around touch.
Quantum is clearly set to open an unimaginable array of applications as optimisation questions can be answered almost instantly. This is another field where progress has been incredibly fast, in 2012 superposition could be maintained for 2 seconds, now the record stands at 6 hours.
One of the most exciting areas is the potential to model extremely complex systems, particularly chemistry (which underpins pretty much everything from materials to food) and the wider chemistry of biology. Drug development is still a process of trial and error, the ability to properly model biological systems would be absolutely game changing. The use of quantum computers for deep learning is also interesting as training time could be reduced to near zero, although I’m not sure how that would influence real world applications. Of course it would be remiss not to mention the urgent need for new types of security once current encryption methods are made obsolete by quantum attacks.
We will do a guest post in the future on the specific hardware challenges involved in quantum computation.
We’re beginning to reach the limit of Moore’s Law. This opens up the opportunity for processors dedicated to specific roles jobs such as deep learning (much faster than GPUs) either in IoT applications, within vehicles or in centralised cloud resources. The next step could be chips that learn to redesign themselves on the fly to suit the application at hand. The finance industry already builds custom FPGAs for specific algorithms. As we begin to better understand how things like memory improve deep-learning we will likely see more processors with these features built in to hardware.
A couple of key opportunities for custom chips is in identifying applications where speed is essential, connectivity is challenging (e.g. underground or moving rapidly) or power is constrained.