Deep Learning

Understanding and using deep learning networks

Five AI Trends for 2021

It’s that time of year again, when many people look forward to a new year and new trends! At MathWorks, we put a lot of energy and focus on trends for AI to inform product direction that influences capabilities we add to products, and see how well we can anticipate customers’ needs so that when the trend is realized, the tools are already there.
First, before diving into the trends, a quick note on COVID-19: I’m pretty sure no one could have accurately predicted this huge event and the impact it has had on all of us. For all the changes that happened, many are using this time to invest in upskilling remote learning with AI-themed courses among the most sought after by the engineering and scientific community.

2021 Trends


AI becomes mainstream for engineers and scientists
AI will be one of the first tools engineers and scientists look to for innovative solutions in solving problems and building applications in 2021. This means it will become common to replace or augment traditional techniques in math, physics, controls and signal processing. AI will continue to help bring answers to previously unsolved problems and enhance existing solutions, such as Reduced Order Modeling, in which adding AI can offer a more efficient solution.
See related: This trend can also be foreseen with the next generation of engineers and scientists learning AI that is integrated into traditional courses.


AI aligns engineering, computer science, data science, and IT direction
Engineers are already working with data scientists to build AI models that innovate on new solutions to their projects. However, creating a successful AI-based system is more than just developing a model: it requires model lifecycle management to ensure models will operate in real-world environments over years or decades. In 2021, engineers will augment their workflows to include development best practices, such as model and data versioning.
See related: Experiment Manager


Model Explainability will reduce aversion to AI within safety-critical systems
As more explainability methods are produced by researchers, industry practitioners will become less hesitant to adopt AI innovations within their workflows, because the reasoning behind AI will be clearer. Engineers and scientists are incorporating new visualizations to understand why a model is making certain decisions and the limits at which a model can operate safely. You can see these techniques already being adopted in AI practices, such as LIME and occlusion mapping.
See related: Deep Learning Visualizations

Dr. Barath Narayanan produced visualizations of Covid-19 predictions using AI and Grad-CAM

This work is driving innovation in the verification and validation of AI for safety-critical systems, with automotive, aerospace, and medical standards committees, such as EUROCAE and the FDA, working on the levels needed for certification. Expect to continue hearing a lot about “Explainability” as a key topic for 2021.


Simulation and testing will go 3D and become more realistic
In 2021, engineers will look to leverage recent advances in software tools with 3D simulation that eliminates physical testing. They will integrate their AI models with traditional methods such as Model-Based Design for modelling physical systems, then perform automated testing against a variety of simulated 3D scenarios.
See related:


More AI models deploy to more low-power, low-cost embedded devices
The options to incorporate AI into more edge-based systems are increasing, and engineers are taking advantage of expanded hardware support for more low-cost, low-power devices including FPGAs, ECUs, and MCUs. Techniques such as quantization and pruning, and efficient pretrained models being available in the deep learning community will enable efficient deployment of AI.
See more on quantization, ECUs, and FPGAs.
Those are the predictions for 2021, with a few links to ways you can already take advantage through code, videos, and blog posts. Expect more content to follow on these topics throughout the year.

2020 Trends Response

I also think it’s helpful to look back at the previous year and take stock of how well MathWorks responded to last year’s trends. Here are 3 highlights:
Trend: Workforce skills start to abate
Engineers and scientists who weren’t formally trained in deep learning feel empowered to begin using deep learning in their applications. Free Onramp trainings saw a significant increase in participation. MathWorks continues to identify new materials - like a new data science Coursera course - to enable customers to be successful with their deep learning applications.
How MathWorks responded: We built these learning tools to help engineers and scientists to learn deep learning, and it turned out (due to Covid-19) these became even more important, so expect continued emphasis on virtual events and content.
Trend: Reinforcement Learning moves from gaming to real-world industrial applications
Overall, we are certainly seeing more customers looking into RL, and major research groups are applying RL to real-world problems. However, people are still exploring the technology though, and it’s still quite new.
How MathWorks responded: There are plenty of new resources for those interested in this space.
Trend: Simulation lowers a primary barrier to successful AI adoption – lack of data quality
There are three distinct areas of note to talk about simulation in response to this trend:
  • Simscape continued to create capabilities for modeling faults in simulation. Simscape Driveline and Simscape Electrical can be used to model component faults in system level simulations to create synthetic training data for development of predictive maintenance algorithms.
  • RoadRunner was added to MathWorks offerings in 2020, enabling design of 3D scenes for simulating and testing automated driving systems. RoadRunner photorealistic scenes can be integrated into AI algorithms that can be trained using simulation data.
  • Predictive Maintenance using simulation can be seen in this example on grid fault location detection, and a full series on model-based design for predictive maintenance here.
So that’s all for today. We covered the top trends coming in 2021 and took a look back at 2020 to see MathWorks response. Do you see any upcoming trends missing from the list? Comment below!

  • print
  • send email


To leave a comment, please click here to sign in to your MathWorks Account or create a new one.