CTO Tamar Shoham on AutoSens TV

Beamr CTO Tamar Shoham spoke with Carl Anthony on AutoSens TV about ML-Safe compression:

-> Why we can't ignore the elephant sitting on the servers, and how to identify the information that matters least to the model.
-> Why lossless and content-adaptive compression must be safe for machine vision models.
-> How we benchmarked ML-Safe compression across 2D and 3D object detection, depth estimation and reasoning models.

AutoSens is the global community of engineers for solving the challenges shaping automotive safety.

Read the full transcript:

Carl Anthony (AutoSense):

Welcome to AutoSense, the global community shaping the future of ADAS and autonomous vehicle perception technology. We are joined by Tamar Shoham, CTO at Beamer, to talk about the technology they're building and some of the challenges they're solving in the automotive industry right now, including a solution called ML-Safe compression. Tamar, welcome to AutoSense. Good to see you, friend.

Tamar Shoham (Beamer):

Thanks so much for having me.

Carl Anthony:

Indeed. So let's begin with autonomous vehicles and next-generation ADAS vehicles. We know that they generate enormous amounts of camera data. What are the biggest challenges that this is creating right now for engineering and machine learning teams?

Tamar Shoham:

A typical autonomous vehicle will have anywhere between 5 to 29 cameras (like in the new Waymo automobiles), each accumulating high-resolution video and creating massive amounts—terabytes—of raw video, even for a typical length trip or commute. This means even the simple action of uploading the trip data to a server becomes a challenging task. It can rarely be accommodated in real time and often requires the actual physical dismount of disks or lengthy overnight uploads.

When we look at AV companies running fleets of cars to collect data, the challenge becomes even more insurmountable. They could be storing these trips for legal reasons, offline analysis, or future improvements of their solutions. This creates two main problems. One is storage—these are huge and exponentially increasing datasets that are expensive to store. But possibly a larger concern is accessing this data and the incurred overhead.

For example, if I want to collect new training data to improve one of my models to create the next generation of my product, and I need to access this huge dataset, there are large time and cost overheads in access. This adds significant friction to the workflow and essentially slows down industry progress. That's why we believe compression is needed to reduce video bitrates and file sizes—but in a way that keeps functionality intact.

Carl Anthony:

Right. Just as I mentioned in the beginning, MLSafe compression is something that Beamer talks about. What is ML-Safe compression?

Tamar Shoham:

As we all know, machine learning models are the beating heart of AV solutions. These range from simple models like lane-keeping assist all the way to full autonomous driving models that make all driving decisions based on sensor inputs. Video data specifically plays a huge part in these inputs.

In order to make the video manageable, we want to compress it, but in a way that won't jeopardize model accuracy, outcomes, or the decisions the model is making. That's exactly what we call ML-Safe compression or MLSafe encoding.

Traditionally, video encoding primarily targeted humans, and for many years the goal was retaining as much perceptual quality as possible—at Beamer, that was a lot of what we did in our first decade. For AV, while perceptual quality is still relevant, the primary goal we're targeting is preserving the quality needed for our ML tasks. That is MLSafe encoding: compressing video in a way that is safe for the machine learning models using the data.

Carl Anthony:

Staying with the ML-Safe compression conversation for a moment, how do you validate it? Walk us through some of the tests and benchmarks you've performed on various computer vision models.

Tamar Shoham:

That's a really good question and one that's hard to answer in a couple of minutes—we have spent endless hours measuring and evaluating the impact of compression on different tasks, ranging from lane detection through 2D/3D object detection, depth estimation, and sophisticated VLMs that do reasoning on video.

We found that each model type requires a slightly different methodology to understand how much compression can be introduced without any negative impact on performance. We look at the behavior under compression of model statistics—things like mean average precision (mAP), detection rates, and semantic similarity of results—and compare these to various noise floors.

We also need to look at the behavior of outliers or edge cases, because those can be mission-critical and must be examined. Our research shows that some models are more cooperative; they are relatively robust to mild changes in video, so as long as visual cues and semantics are retained, the model is happy. Other tasks may be very sensitive to any modification—even moving something one pixel left or right—due to overfitting in training or strict data requirements. These require either more conservative compression or additional actions in the pipeline to reach the right ML-Safe compression solution.

That's just the tip of the iceberg. We've published quite a few blog posts on these case studies on the Beamer website, which also link to technical Hugging Face white papers where you can learn more.

Carl Anthony:

Looking ahead, what role do you think ML-Safe compression could play as autonomous vehicle datasets continue to grow?

Tamar Shoham:

The massive amounts of video data in this world dictate that compression must be used. So the question isn't if we should compress the data, but how. We believe the answer is doing it with awareness of ML safety so that we remove data without harming downstream processes.

Some data may require lossless compression where any manipulation cannot be tolerated for that specific workflow. Other data can be compressed with a lossy algorithm, provided it preserves the data needed for the task. In the more mature world of media and entertainment, the industry learned to discard information with the lowest perceptual importance and spend bits where needed. We need to do the exact same thing for AV video.

If we ignore the literal elephant on our servers, we'll have no choice but to start throwing away meaningful data—something industry leaders are already admitting to doing out of necessity. As an industry, we need cost-effective solutions that remove non-critical data while preserving what's important, and applying MLSafe compression at scale is a key part of that solution.

Carl Anthony:

Tamar Shoham, CTO at Beamer, joining us to talk about MLSafe compression and its benefits. You and your colleagues will be joining us in Barcelona for AutoSense Europe 2026, September 22nd through the 24th. We look forward to seeing you then!

For more information on Beamer and ML-Safe compression, see the links in the description and visit auto-sense.com. For more in-depth interviews like this, subscribe to the AutoSense YouTube channel and follow AutoSense on LinkedIn. On behalf of AutoSense in Detroit, I'm Carl Anthony.