Applications for Neural Networks in the IoT
Neural Networks are a computing system made up of a number of simple, highly interconnected processing elements [4]. It has been described as the 3rd machine revolution after the industrial revolution and computing revolutions.
They are universal approximators ideally suited to situations with a number of characteristics. They do have caveats, they require careful training to avoid local minima and becoming too specific, and can be unpredictable in edge cases, this coupled with their lack of transparency makes them challenging to use in safety critical or heavily regulated environments:
Variety: Varied data can require more complex processing to account for the changes.
Volume: High volumes of data call for automated processing and are required to training neural networks.
Unstructured Data: Data without clear semantic labels or simple patterns, such as computer vision, speech recognition, or natural language processing.
Complex Variables: When there are many, complex variables that define a pass or fail criteria.
The Internet of Things is a technology trend with disruptive potential to create a paradigm shift in the way all physical products are designed, used and sold. It creates a bridge between the digital and physical worlds. In doing so it will provide new & better user value, create new products, open new markets and enable new business models. The expectation of exponential growth comes from the depth of value this connectivity can offer combined with the breadth of applications. As technologies standardise, this is the best way to describe the trend for B2C applications. The advantages of the IoT stem from several inherent characteristics:
Scale: The more connections made, the greater potential value a network can provide [9].
Automation: Complex decision making by machine processes improves quality, consistency, reliability, efficiency and throughput. Bob Metcalfe, “the IoT should disappear into the woodwork, even faster than Ethernet has” [2]. Mark Weiser, chief technologist at the Palo Alto Research Centre, 1991, “the most profound technologies … those that disappear“ [3].
Social: Connecting with real people will be easier than ever before. Mark Zuckerberg, Facebook CEO, envisions an “internet of people”[1].
Physicality: Devices are able to alter our reality through actuators, improving quality of life and quality of an environment.
Personal: Forming a semantic understanding of our world by leveraging data, connected devices will have much deeper interactions with people [10].
IoT operations can be divided into three functions, each of which is vital to conducting value adding activities.
Data Creation – reducing cost of sensors provides greater granularity of data due to quantity, advanced types of sensors, grouped sensors or advanced local processing provides better quality of data for processing.
Data Analysis – refers to the processing of data to answer specific questions. This creates useful information from data with which to derive a decision.
Actuation – devices which are able to act on the environment in useful ways are obviously instrumental to creating value from the IoT. The more capable an actuator is, the more value it can provide. Like data creation, the comes down to increasing quantity or increasing quality.
The role of neural networks is clearly in data analytics. It can be applied at an number of different places in the IoT technology stack.
At the Edge – Processing on a sensor offers maximum security by never saving or broadcasting raw data. Google Clip uses AI to only take photogenic snaps.
In the Fog – Fog computing refers to processing data near the source rather than in a remote cloud [12]. This reduces latency, conserves bandwidth and increases reliability while enabling more complex processing.
On the Cloud – The most complex processing takes place in the cloud. Natural language processing is undertaken here for Microsoft Cortana, Amazon Alexa, Samsung Bixby, Apple Seri, Baidu Duer and Google Home. Abstraction allows collating of many different data sources to generate the highest level insights. Neura:AI reveals semantic user insights by combining many data sources.
Neural Networks are already being applied to a number of applications in the IoT:
Device Security – Pattern based security spots unusual behaviour, it is light as it doesn’t rely on databases and scalable playing nice with any number of connections or traffic routes unlike a traditional firewall. Cylance’s Protect software for the IoT “will use less than 1% of CPU resources, according to Stuart McClure, the company’s CEO and founder – and does not require an Internet connection”[11]. This leaves the door open to being embedded on devices, Dark Trace uses a similar approach.
Network Security – The same pattern recognition approach can monitor network traffic, F-Secure’s Sense home router, Cojo, Dojo and Luma devices do just that [5]. Another approach by Israeli start up, Cyactive, bought by PayPal in 2015, used machine learning to build predictive malware databases as over 90% of malware is evolutionary [5].
Home Security – Security cameras from Ring, Nest, Hive, Netamo, Canary now flag unfamiliar faces and can avoid flagging pets or animals to users.
Voice Assistants – Microsoft Cortana, Amazon Alexa, Apple Seri, Baidu Duer, IBM Watson and Google Home all use neural networks for natural language processing.
Semantic Analytics Software – Neura:AI reveals semantic user insights from many granular data sources and IBM provide a flexible neural network computing platform through Watson.
In the future, I expect to see more applications for neural networks in the IoT. Though there is a risk they are held back by basic actuators. Without being able to do more, the power of neural networks will not be experienced in the physical world.
In 1 – 2 years – Machine learning will be applied at the edge to get even more out of a camera, edge detection of surfaces, semantic recognising of objects, activity detection will all be built into products.
In 1 – 2 years – The building of usage profiles will go beyond ‘mimic modes’ and basic memory like those of Nest’s thermostat and start to attach personal preferences not only to individuals, times and locations but to social contexts, moods and other complexities of human life. E.g. Dad likes his radio on classic fm, but flicks it to Radio 1 when the kids are in the car, unless he is in a bad mood when its on classic rock, loud.
In 2 – 5 years – More complex daily tasks will be undertaken by machine learning leveraging data and interconnectivity of the IoT, health care and insurance providers are already linking their systems with wearables. Vitality Health Insurance are already offering Apple watches as a sign up reward [6], allowing users to earn points via their app for staying active. Future collaborations can be expected to help diagnose patients in the home, leveraging multiple sensors alongside thousands of patient records and indirect data like how many times a toilet was flushed, and contact your doctor or even prescribe action accordingly. Google’s Deep Mind are working with the National Health Service in the UK [7] and Researchers have already demonstrated these systems are already capable of outperforming doctors in the diagnosis of some diseases [8].
In 5 – 10 years – Neural Networks will be used to build machine emotion into products Expect to see digital assistants start to bond with users through inside jokes or offering emotional support. This will be mirrored in hardware as robots Baidu’s recently announced home assistant the Raven R is designed to do just this, much like the Pixar Lamp.
References
www.stanforddaily.com/2014/02/07/mark-zuckerberg-and-the-internet-of-people
www.forbes.com/sites/gilpress/2014/09/21/the-new-apple-wristop-computer-a-missed-opportunity-to-defin e-the-internet-of-things/
Mark Weisner, The Computer for the 21st Century, 1991
“Neural Network Primer: Part I” by Maureen Caudill, AI Expert, Feb. 1989
www.thenextweb.com/insider/2016/01/04/4-devices-that-can-help-secure-your-homes-iot/
www.vitality.co.uk/rewards/partners/active-rewards/apple-watch/
www.deepmind.com/applied/deepmind-health/working-nhs/
www.sciencemag.org/news/2017/04/self-taught-artificial-intelligence-beats-doctors-predicting-heart-at tacks
www.p2pfoundation.net/Reed’s_Law
Stephen P. Anderson, Seductive Interaction Design, 2011
www.networkworld.com/article/3014499/internet-of-things/the-iot-calls-for-an-ai-based-security-a pproach.html
”Fog Computing and the Internet of Things: Extend the Cloud to Where the Things Are”, Cisco White Paper 2015