The Democratisation of Artificial Intelligence

The true democratisation of artificial intelligence is coming. It is happening the same way the democratisation of all technology has happened in the last 15 years, through the medium of your mobile device. Through the last decade and a half, mobile technology has been at the forefront of making previously costly and hard to implement technology accessible. The reason for this explosion of accessibility has been the huge investment in sensor and chip miniaturisation that has been a consequence of the mobile revolution. As mobile devices got smaller and more capable, there was a large competitive drive by phone manufacturers and their chip suppliers, like intel and Qualcomm, to miniaturise previously large and complex systems and to manufacture these systems to be as power efficient as possible, due to the limitations in battery technology. This competition has resulted in all kinds of sensors and systems on a chip from air pressure sensors to motion and heart rate sensors being included in mobile and wearable technology. The miniaturisation and power efficiency drive has enabled manufacturers to build functionality into the small personal computers that we all carry (smartphones) that would have been pure science fiction only a decade ago. Now this same driving force is resulting in large advancements in on-device personal artificial intelligence.

The main difficulty with artificial intelligence is related to the way neural networks function. The truly useful artificial intelligence categories such as deep learning, pattern recognition, and classification require expensive hardware that is very power hungry to run. The reason for this is that neural networks mimic the way a brain works rather than the way a machine traditionally works. A traditional computer processor is able to do many tasks in a row very quickly and can usually switch between parallel tasks if it has multiple cores. However even with multiple cores a CPU is not truly parallel it is still doing sequential work very fast. The brain and neural networks on the other hand are good at tasks that are massively parallel. For example, tasks like face or image recognition require a neural network to make several parallel connections at the same time in order to do it quickly. Traditionally this has meant that neural networks are run on GPU (Graphical Processing Unit) hardware as it is built to perform massively parallel operations (which is needed for 3D graphics).

GPU Technology

GPU technology was initially used heavily for artificial intelligence, with racks of GPUs mounted in data centres and massive power sources to support them. As artificial intelligence research progressed and the uses for artificial intelligence started to grow, companies like Google that rely heavily on it started to develop specialised hardware based on GPUs. However, this hardware still runs in massive data centres with massive power supplies.

Meanwhile artificial intelligence functionality started to make it to mobile apps, for example face detection in Snapchat, dynamic artwork generation in Prisma, and person recognition in google photos. However, all of the actual processing was occurring in the data centres on this specialised hardware and the data was communicated to and from your mobile device over the internet. While this worked it meant that the speed of the functionality was dependent on the strength of your network and limited to what can be securely transmitted over that network. This is not ideal as it introduces latency and security issues. Some manufacturers like apple have managed to make neural networks work well on their devices by allowing developers to use the GPUs to run them, this works however because of power limitations it results in neural networks with very low mathematical precision as compared to those running in data centres and cannot be used for serious heavy lifting.

Recent Advances

In the last year the tech industry has seen the emergence of specialised miniature chips for running specialised artificial intelligence directly on a mobile device. One example of this is Apple’s NPU (Neural Processing Unit), this chip allows apple to quickly analyse the results from the iPhone X’s new face recognition sensor suite and very rapidly match it against the known user. The NPU is the cornerstone of Apple’s FaceID and is what enables its face recognition system to work so rapidly (takes 0.5-1 second on average). Apple keep your biometric data locked up safely on the device as a matter of security policy, so FaceID would have been impossible without this dedicated hardware. In 2017 as well, Huwei released the “Huwei Mate 10 pro” smartphone. This device comes with a built in AI chip, that handles the processing for a lot of artificial intelligence tasks that are supposed to make using the device better. For example it can recognise what kind of scene you are trying to take a picture of and apply its AI processing to adjust lighting and white balance for a better image. It also is able to learn your habit and load applications into memory before you use them for a more streamlined experience.

We are just starting to witness the emergence of dedicated mobile artificial intelligence hardware. As the micro-processors become more advanced and power management on mobiles improves we will see a boom in AI assisted functionality on our mobile devices. This will include everything from making the digital assistants on the phones smarter, to learning your daily patterns, to being able to do complex AI tasks on the device without the need to call out to the network. These are very exciting times for machine learning and AI and we at SmilePass cannot wait to see what 2018 will bring.