Artificial Intelligence is considered the modern era of innovation that has transformed the digital age. Smartphones are already an example of how our mobile devices have become more intelligent and how AI is powering the future intelligence of mobile devices is yet to be explored.
Today, businesses and mobile manufacturers are collaborating to build hardware capabilities of mobile devices to accommodate the machine learning applications for the devices. But, many enterprises and researchers are looking to develop an intelligent architecture for mobile devices and applications to be more intelligent than ever.
Mobile Intelligence is the power of mobile systems to learn, analyze, understand and resolve user queries through intelligent solutions. For this intelligent capabilities, AI-powered machine learning and other technologies are utilized.
Mobile architecture is divided into three major architectures:
- Partial offloading
Many developers are already using the cloud-based or local-based for application development, while partial offloading is relatively a newer approach. Though there have been approaches using these architectures in combinations and yet if all the three are used to form an architecture, it can help applications become more intelligent and in turn make our mobile devices more intelligent.
The mobile client can be considered a local-side as it receives input signals and preprocesses them locally. Then these mobiles send the signals either to the cloud’s Machine learning model or to the local ML model.
After the processing in the ML models, the results are shared with the users for intelligent solutions.
The cloud server has many computing and cognitive resources such as CPU, GPU, and TPU, These resources are utilized for ML model learning and to make the ML learning easier, cloud server should be provided with the learning data and configuration files of the related models. The cloud itself can continue carrying learned models and provide seamless web APIs for inference processing.
Image Source: Mobile Intelligence: Architecture
For cloud-based architecture as suggested in the above-given figure, the mobile client and cloud server work together to make predictions including learning and inferencing. When learning is finished on the server, the cloud server obtains the learned parameters.
Further, we can put the learned model on the server and publish web APIs which mobile devices can use. It is easy to port the application to different platforms with the model being on the server. The same can’t be done for the inferences as they require a network and can’t be done locally.
Join our weekly newsletter to receive:
- Latest articles & interviews
- AI events: updates, free passes and discount codes
- Opportunities to join AI Time Journal initiatives
Phone B is a locally based device, which means the predictions are made by mobile-only. We can put the learned model in the mobile device and inference locally on the device itself. There is no requirement of server over the network, but, the device needs higher CPU and RAM capabilities.
This architecture is more dynamic and flexible than the other two. This model is composed of many abstract layers. On one side the mobile client partitions the model according to variables like network condition, mobile capability, and server load. On the other side, it executes the model up to a specific layer and transfers the intermediate data to the cloud through a network.
The cloud server then executes the remaining layers and sends the prediction results back to the mobile client. These architectures are universal and can be applied to various machine learning models like Deep Neural Network (DNN), Reinforcement Learning (RL) models and Generative Adversarial Network (GAN).
Parameters for Architecture:
It is the time lapse between the user’s request and prediction results, including pre-processing, model operation and post-processing. Take an example of Virtual Reality applications, they require 14ms latency and 60FPS(mobile display refresh rate) and in applications like cloud-based gaming, interaction latency needs to be kept as short as possible.
The ratio of the number of samples that get the correct results, to the total number of samples. It is used to measure the performance of the model. Applications with a requirement of a high level of security and functions such as autonomous driving and navigation can require very high accuracy.
There is no room for error and that is why businesses look to hire android developers and other developers from various platforms to exploit the learned DNN(Deep Neural Networks). If we consider the application of these models in automobile automation, these self-learned DNN can predict the steering angle of a driver for road navigation and safety.
Mobile devices have a limited source of energy. However, running these complex machine learning models can introduce considerable computing and communication overhead.
Mobile intelligent applications are quite famous among the users. But, this can change with the applications using a large amount of energy. Thus, energy efficiency is a desired attribute.
Key Challenges for Mobile Intelligence:
- Network conditions are often sporadic, unstable and unpredictable.
- Lack of energy efficiency due to cloud computing and cloud collaborative operations.
- Cloud computing can lead to data risks and data privacy issues.
- The increase in model complexity and data size are increasing.
- Current partition of inference process is still experimental.
- Hardware capabilities issue for mobile devices.
- Enterprise infrastructure and organizational structures are not calibrated for ML adaptation.
- Lack of control and regulation over the data sources and data breach.
- Calibration of models to provide higher results and predictions.
- Latency issues for low latency demands.
As Artificial Intelligence technologies mature into a peak phase, we will see many such mobile intelligent applications that will explore the machine learning technologies. The parameters of required architecture and current architecture are not a match and there remains one of the most prominent challenges.
With the above-proposed architecture, the process of inference and learning becomes swift and smooth without the fear of data risks. Businesses around the world know the importance of mobile intelligence and intelligent applications and the same has been worked upon through many types of research and experiments today. Thus, the use of a combination of cloud, local and partial offloading architectures can help meet the requirements of the parameters of machine learning.
Manoj Rupareliya is a Marketing Consultant and blogger. who has been writing for various blogs. He has previously covered an extensive range of topics in his posts, including Business, Technology, Finance, Make Money, Cryptocurrency, and Start-ups.