By Jos Martin, Senior Engineering Manager at MathWorks
As AI, deep learning, data analytics, IoT and other concepts intersect, applications that once seemed futuristic come closer to reality. However, the engineers and scientists tasked with researching, developing and launching these applications are still learning and adapting to new concepts, design workflows and skills for their evolving job roles.
With AI becoming more prevalent across industries, there is a growing need to make it broadly available, accessible, and applicable to engineers and scientists with varying specialisations. Engineers and scientists, not just data scientists, will drive the experimentation and adoption of deep learning in industrial applications. Complexity of larger datasets, embedded applications, and bigger development teams will drive solution providers towards interoperability, greater collaboration, reduced reliance on IT departments, and higher productivity workflows.
- AI is not just for data scientists
Engineers and scientists, not just data scientists, will drive the experimentation and adoption of deep learning. Technical curiosity, business imperatives to reap the promise of AI, and automation tools will empower more engineers and scientists to adopt AI. New workflow tools are simplifying and automating data synthesis, labelling, tuning, and deployment, thus making AI accessible beyond data scientists. These tools are also broadening the breadth of applications from image and computer vision to time-series data like audio, signal, and IoT that are common in numerous engineering domains. Example applications range from unmanned aerial vehicles (UAV) using AI for object detection in satellite imagery, to improved pathology diagnosis for early disease detection during cancer screenings.
- Application and domain specialisation
Industrial applications are becoming a major consumer of AI but bring new demands for specialisation. Smart cities, predictive maintenance, Industry 4.0 and other IoT and AI led applications demand a set of criteria be met as they move from visionary concepts to reality. Examples include safety critical applications that need increased reliability and verifiability, low-power, mass-produced and moving systems that require form factors, and advanced mechatronics design approaches that integrate mechanical, electrical and other components. A further challenge is these specialised applications are often developed and managed by decentralised development and service teams (not centralised under IT). Example applications range from agricultural equipment using AI for smart spraying and weed detection to overheating detection on aircraft engines.
- Interoperability
Interoperability will be critical to assembling a complete AI solution. The reality is there isn’t a single framework that can provide “best-in-class” for everything in AI. Currently, each deep learning framework tends to focus on a few applications and production platforms, while effective solutions require assembling pieces from several different workflows. This creates friction and reduces productivity. Organisations like ONNX.ai are focusing on addressing these interoperability challenges, which will enable developers to freely choose the best tool, more easily share their models, and deploy their solutions to a wider set of production platforms.
- Edge computing
Edge computing will enable AI applications in scenarios where processing must be local. Advances in sensors and low-power computing architectures will enable edge computing with high performance, real-time, and increasingly complex AI solutions. Edge computing will be critical to safety in autonomous vehicles that need to understand their local environment and assess driving options in real-time. This has promise to yield huge cost savings for remote locations with limited or expensive Internet connectivity, like deep sea oil platforms.
- Complexity necessitates greater collaboration
The increased use of machine learning and deep learning in complex systems will necessitate many more participants and greater collaboration. Data collection, synthesis and labelling are increasing the scope and complexity of deep learning projects, requiring larger and decentralised teams. Systems and embedded engineers will require flexibility to deploy inference models to data centres, cloud platforms, and embedded architectures such as FPGAs, ASICs, and microcontrollers. These teams will also need expertise in optimisation, power management and component reuse. Engineers at the center of collaboration, developing deep learning models, will need tools to experiment and manage the ever-growing volumes of training data and lifecycle management of the inference models they handoff to system engineers.
