By Balakrishna D. R. (Bali), Executive Vice President – Global Head, AI and Automation and ECS, Infosys
Artificial Intelligence (AI) has revolutionized human endeavor and augmented our capacity to perform tasks that require extremely high level of human cognition using powerful deep learning models. But for the state-of-the-art AI systems to achieve mainstream acceptance, it needs to enter into our factories, workplaces, and even our day-to-day gadgets and get embedded into the fabric of our existence.
One of the myriad factors helping this is the adoption of Edge-based AI systems, where the inferencing and other computations happen locally on devices at the edge of the network, with minimal involvement of the cloud. This is significantly less resource-intensive than typical AI systems that are heavily reliant on cloud-based data centers for their compute-intensive workloads. It enables AI to deliver insights in real time and consume significantly less energy, bandwidth, and allied infrastructure. Subsequently, this unlocks many novel AI use cases in autonomous vehicles, smart devices, security and surveillance, healthcare, manufacturing, and any other areas where real-time inferencing is a critical differentiator.
The right kind of models that give Edge AI it’s ‘edge’
Edge AI uses compressed (or tiny) versions of existing AI models for better performance in resource-limited areas. Through careful data selection, compact model architectures, and other techniques, developers can optimize and shrink the original AI models.
This shrinking of the models reduces complexity and makes it possible to run the inferencing on smart-edge devices without sending data to the cloud every time. The compression and shrinkage of the model comes with a trade-off in accuracy, which is acceptable in most industrial use-cases.
Choosing the appropriate model is of paramount importance in Edge AI as a heavy resource-intensive model will fail to deliver the performance, and finetuning and compressing it in a manner that is suited to the edge environment is needed. Model considerations also need to factor in the needed levels of explainability, the necessary levels of accuracy, and its trade-off with resources and various processes in the downstream logic.
For example, in a stock monitoring situation in retail, one may need an object counting model, an object detection model, or even an object tracking model based on the customer requirement. There are various open-source models available and careful consideration is needed during selection, based on the edge compute available.
Due to the reduction in workload of training and inferencing on cloud servers, the speed of execution also increases. Edge AI requires substantially less data to train; thus, the corresponding data management frameworks and infrastructure can be far simpler. Ultimately, since it consumes significantly lesser power, it is also more environment-friendly.
Edge AI can also leverage the acceleration offered by other emerging technologies like 5G and the Internet of Things (IoT). IoT devices can detect and capture the right kind of data for Edge AI, and 5G can offer unparallel network capabilities for Edge devices to run faster.
Edge AI has diverse applications
Smart consumer devices: Edge AI can fundamentally transform our interaction with a variety of common consumer devices by bringing powerful deep learning at an affordable price points to make a variety of context-aware devices. AI can potentially run our household operations by integrating natural language processing and computer vision capabilities into common household devices. In the workplace too, digital workers will be far more enabled now than they already are. The reduction in costs due to lighter models and low compute requirement will give rise to a series of responsive smart devices.
Highly regulated and privacy-conscious industries: Low data storage due to reduced data requirements will be appealing for sectors like healthcare and finance where data collection and storage is highly regulated. In each edge device, the models are customized to the specific edge environment and critical data never exits outside the edge network. Edge AI will resolve data concern related issues in this segment as it solves the problem by reducing the data infrastructure, also augmenting security.
Mobility solutions: For detecting and responding to obstructions and traffic situations in real time in autonomous vehicles, we would need AI systems that have extremely low levels of latency, but phenomenal accuracy. Compact models coupled with efficient high performance edge computing can solve such problems.
Healthcare and medical sciences: In the context of smart hospitals and responsive medical devices that track patient health conditions in real time and make decisions to adjust the dosage and other treatments in real time, this could potentially enable medical practitioners to save countless more lives by expending the same effort.
At present, AI’s carbon footprint is greater than the entire airline industry, and with further proliferation, this number is only expected to grow. By making more light-weight models that run on edge, we can achieve unparalleled efficiency gains and increased adoption, without comprising on our sustainability goals.
Very good blog post. I definitely love this website. Thanks!