In the race to bring self-driving cars to market, the automotive industry is putting the pedal to the metal in the field of deep learning. Along the way, they’re navigating and overtaking workflow and infrastructure challenges to turn their research into real-world applications.
But the obstacles they face aren’t limited to the world of autonomous driving.
As they look to incorporate deep learning into their businesses, here are the top five things companies from just about any industry can learn from the automotive market’s experience.
1. Embrace the new.
We’re at the early stages of a new computing era — AI computing, which is fueled by GPU deep learning. Anything new is often met with trepidation or caution, but those who embrace change early on will reap the benefits.
Not long ago, self-driving vehicles were the stuff of science fiction. Now, autonomous racecars, like Robocar, are taking to the track. Autonomous trucks are poised to revolutionize long-distance freight delivery. And cars have the smarts to not only navigate traffic jams, but also to watch over you while you drive.
The automotive industry has been quick to break with tradition and embrace the newest technologies. Early adopters of AI have positioned themselves at the forefront of a competitive market, and are ready to reap the benefits.
Now is the time for enterprises to grab the opportunities AI provides with both hands.
2. Get good quality data. And lots of it.
The success of deep learning systems is measured by the level of accuracy they achieve. This requires a huge amount of training data. And complex neural networks need to be developed to learn from these vast datasets without forgetting past experience.
In a recent blog, Adam Grzywaczewski, a deep learning solutions architect at NVIDIA, explains how autonomous vehicle developers are achieving the level of performance demanded by safety-critical systems. Neural networks need to be trained on datasets that incorporate all possible driving variables — from weather to situational changes. These datasets also need to contain high-quality, labeled data.
Cutting corners on data collection is unwise. Implementing a clear workflow is crucial for enterprises to capture the kind and amount of data they need.
3. Hire and retain good talent.
The demand for skilled staff far outweighs supply in what is becoming an increasingly competitive job market for deep learning developers. But building the right team is critical to a successful long-term AI strategy.
The automotive industry understands this and has invested in hiring and developing talent. Investing in a balanced, experienced, talented team allows a business to stand out from the crowd, and draw still more recruits.
According to Forbes, some of the big players have turned to offering impressive salaries to attract talent.
To retain the talent you attract, you’ll also need to provide them the right space to grow and develop as well as the tools they need to do their best work. Your research team will have to execute thousands of training jobs as they build your systems. Some jobs will be run on small datasets for debugging and similarly small tasks. Volkswagen, for example, has set up its Data:Lab initiative to support startups and talented individuals who are working on innovative automotive solutions.
4. Make use of the latest deep learning development platforms.
Deep learning algorithms and frameworks are evolving at an unprecedented rate. In the last six months alone, one framework, TensorFlow, evolved through six different versions (from version 0.12 to 1.4).
Making use of the latest deep learning platforms helps ensure you’re always up to date and can focus on solving problems, rather than over-engineering and re-engineering.
Large automotive companies, such as Volvo, VW, ZF, Autoliv and HELLA, as well as 145 automotive startups from around the world have turned to NVIDIA’s AI platforms. These platforms offer a cloud-to-car solution, with NVIDIA DGX systems training deep neural networks in the data center, and NVIDIA DRIVE PX delivering real-time, low-latency inferencing in the vehicle to drive safely.
NVIDIA DGX systems are fully integrated with GPU-optimized deep learning frameworks, tools and libraries. Paired with the NVIDIA GPU Cloud, the powerful system makes experimentation, scaling and collaboration fast and easy. Plus, NVIDIA offers regular updates and enterprise-level support.
5. Success requires both flexibility and stability.
Creating a stable platform for developers in the midst of an extremely dynamic deep learning ecosystem is crucial. Developing an end-to-end production pipeline can help ensure you nurture innovation, while not losing sight of the aims of the project.
Establishing a baseline performance model and useful metrics can provide insights into project performance and allow for easy comparisons between models. Testing against these metrics regularly will make it easier to measure progress and reveal where improvements can be made.
Remember that AI algorithms rarely live in isolation. When deployed as products, they will have to meet a number of non-functional requirements. Whether your system is constrained by power consumption, latency, memory or something else, make sure that you understand, control and can react to those requirements from day one.
In Grzywaczewski’s blog, he explores further the need for establishing metrics and how they can help you to plan for scale in for automotive vehicle development. The same key themes he explores are applicable to enterprise challenges, in particular the message that complicated challenges require a combination of solutions.
Find Out More
To evaluate the complexity of your problem, discuss the challenges of distributed training and large-scale model validation, or get support in organizing your own artificial intelligence team, please contact our enterprise support team.