With artificial intelligence at the core of software technology for autonomous vehicles, machine learning developers are focused on accelerating the pace of model creation and innovation.
For many machine learning experts, the data set is the first step in accelerating development when new outputs need to be added to the perception stack, such as the computational resources that enable artificial intelligence hardware and software of autonomous vehicle control systems to “ver.”
In the case of autonomous and autonomous vehicles, a key challenge for the design team is that the vehicle must recognize traffic cones or, for example, yellow lights for the first time.
Start With The Data
The first step in solving this problem is building a framework for continuous learning, said Sammy Omari, vice president of engineering and head of autonomy at Motional, an autonomous vehicle maker that is a joint venture with Hyundai Motor Group and Aptiv.
The framework begins with the tagging or detection of incidents and other incident scenarios and then transferring those scenarios to new training sets. Once done, developers need an effective training framework to train new models. He said the final step is to understand how the new release will affect the overall end-to-end performance of the autonomous vehicle system.
In the traffic light situation, crews can start with what Omari called “naive” data. This data defines six to 12 months of driving data and includes potential traffic light scenarios where an amber light may have been present.
Developers can run this data on an offline system such as the cloud. The next step is to send the results to human annotators to determine where there are still inconsistencies in the dataset.
Balance Automation and Focused Improvements
Since each step of the machine learning (ML) development process often involves multiple teams focusing on different parts of the workflow, the challenge is to balance automation of the whole process making sure to focus on targeted goals and improvements.
Building an effective mining system could solve this problem, Omari said. In Motional, this means creating a scenario mining and search framework. This framework enables developers to compute an extremely large set of attributes after each training mission for the autonomous vehicles they drive.
Accelerating ML production is also critical, said Yanbing Li, senior vice president of engineering at Aurora, a provider of autonomous vehicle control systems.
Li said his team is removing the friction of developing core ML technology by applying automation, making starting experiments, a real push of a button experience for ML developers who focus on making minor changes focus on your ML code.
But they get this automated experience of doing experiments and getting results, he continued. By making the validation process smoother and the AI infrastructure invisible and behind the scenes, his team at Aurora can reduce complexity and allow ML developers to focus on validating models.
Being able to change data in real time is another way to use automation while making targeted changes to the system, said Gonen Barkan, General Motors group director for autonomous vehicle radars.
When we look forward to the behavior of the sensors in the near future, this will not be fixed, said Barkan. For today’s radars, you can check their operation on the fly. He added that when a machine learning team is not flexible to change data, it ends up losing a lot of skills. Having a very flexible ML pipeline, to digest, train, adapt the noise modeling, adapt the way you process data, is extremely critical in order to be able to use the sensor effectively, he said.
The result of modifying the dataset, however, could bring engineers back into machine learning and interrupt the experiment.
One way to avoid this is to build a simulation system, Omari said. This allows machine learning teams to automate assessments of every large-scale dataset change and allows teams to get the same or a very similar signal from a human vehicle that the driver would receive. .
Omari thought, this is one of the biggest challenges for him, in the entire industry. At Aurora, machine learning teams focus on seamlessly managing the way they deal with ever-changing data by seamlessly automating and maintaining the ML side of their experiments.
He said his team is focused on managing the CI / CD cycle, a series of steps taken to successfully deploy a new version of the software pipeline. We really try to increase the time we spend driving autonomously every day because that’s what gives us the most feedback, said Li.
Avoiding Regressions
ML teams should also ensure that they create a validation model that doesn’t improve in one scenario and backs out in others.
According to Li, several test modalities can help solve this problem. It’s extremely important that we have a framework that allows us to test one thing at a time, he said.
Siva Gurumurthy, senior vice president of engineering at AI vehicle fleet management provider KeepTruckin, said the provider’s platform combats regressions by creating different versions of each model and data.
KeepTruckin then feeds the different versions into an automated motor that shows when the model performed well and when it didn’t. Another key for machine learning teams is to understand which results are false positives.
Gurumurthy noted that each machine learning team approaches testing and data differently. Everyone is trying to figure out what works best for their environment and there is no set of traditional practices like, Hey, this is how you do your model development, this is how you do your code reviews, model, Gurumurthy continued. Everyone comes up with a lot of ideas and sees what sticks to these ML engineers.
Source link
