Artificial Intelligence as the second phase for IoT – Part 2
The outburst of IoT demand over the last decade brought about technical challenges as explained in Part 1 of this blog:
- First, data scalability - the vast volume of devices is generating/streaming huge amounts of data permanently
- Then, latency due to data transmission time
- Finally, data protection with the increase of threats and vulnerabilities.
By shifting the intelligence closer to the devices, the complexity of the aforementioned problems can be dramatically alleviated, as the amount of information to be handled is much lighter (i.e. in the most optimistic case, each device might carry out their own AI process). On one hand, it is evident that offline model training is a complex and iterative process that requires powerful hardware resources, since it implies applying data engineering techniques to information coming from as many sources as possible. On the other hand, once the model is trained and satisfies our requirements in terms of accuracy, precision and recall, it can be deployed to an IoT device to provide dynamic inferencing or even to adapt its behaviour to the changing patterns of the input features with online training.
Following two example scenarios that are already being developed by the Internet of Everything lab at Atos Research and Innovation, using different IoT Technologies (Machine Learning, Neural Networks, Linear Regressions, etc).:
- Automotive scenario
The introduction of autonomous vehicles is going to deeply transform urban mobility in the next future. In a final and ideal situation, the fifth level of automation shall be reached and manual driving will be only an eccentricity behaviour for nostalgic people. Nevertheless, the adoption of this new generation of vehicles, which has already started with ADAS systems (Advanced Driver-Assistance Systems), is being progressive and is causing the installation of complex and powerful electronic systems in our cars; and their interconnection to the Internet (V2Internet), to other devices (V2I) and vehicles (V2V).Within SecureIoT (https://secureiot.eu/), our Internet of Everything lab is collaborating with a partner specialized in design, manufacturing and testing for automotive industry, which provides a powerful onboard electronic control unit (ECU) based on a heterogeneous multi-processing embedded system (6 x CPU + GPU). Using this IoT platform, vehicle information will be collected (i.e., speed, position, acceleration) from the internal buses in order to train and build TensorFlow machine learning models for two different applications: to assess the insurance risk considering the driving profile and to identify and predict traffic situations.
- Smart cities
One of the realms where the exploitation of data glimpses a largest potential (and impact) is Smart Cities. Moreover, the continuous evolution of IoT devices, protocols and technologies leads to a constant breakthrough, giving rise to brand new disruptive applications and services that were far from being doable not long ago. Together with this, IoT devices have also undergone a huge growth in terms of computational capacity and even battery life, thus opening the door to the direct execution of complex operations within the own devices; this means that, instead of having to stream the data onto the cloud (leading to a non-negligible latency and overhead), nodes can carry out these calculations by themselves. As a direct consequence, we can directly apply Machine-Learning/AI techniques over IoT devices in order to perform "atomic" tasks, whose outcome might either used in standalone applications (e.g. licence-plate recognition) or feed more ambitious systems (e.g. energy and grid management).Under the umbrella of the H2020's Synchronicity project (https://synchronicity-iot.eu/), the Internet of Everything Lab is collaborating with various city clusters (Santander, Milano, Porto, Carouge and Seoul). Technically speaking, the lab is harvesting both parking (availability of free parking lots throughout the city) and traffic flow intensity data as the input to train a Keras-based (https://keras.io/) neural network with Tensorflow as backend. The output of these two processes will allow us to estimate/predict: 1- the probability of finding a free parking lot in a particular area and, 2- the situation of traffic. On top of this, each city will leverage these developments and seamlessly integrate both services into their own (and tailored) multi-modal transportation apps.
The Internet of Everything Lab is currently preparing a webinar where it will be shared its vision and the progress achieved. Namely, a demonstration will be included about the use of these machine learning techniques to estimate the probability of finding a free parking spot based on the context information harvested in a smart city.