Tesla acquires deep neural network start up DeepScale to help with FSD

[diplomat33]

"Tesla has acquired a small artificial intelligence startup called DeepScale in order to help build its self-driving technology and reach its goal of building a giant fleet of robotaxis."

DeepScale focuses on Deep Neural Networks (DNN).

Tesla acquires AI startup to help build self-driving robotaxis - Electrek

Interesting. I hope this bears fruit.

 

[cucubits]

This feels like a sign that they're getting desperate and slowly realizing this dream is not going to happen with the current hardware. I fear that Elon's hate of LIDAR will be their downfall.

I still hope they make it, just seems less and less likely as time and updates go by...

 

[electronblue]

Seems like a team acquisition but the more pressing question is if they are weeks away from feature complete Level 5 no geofence, where is this talent needed?

 

[Rockster]

Acquiring new talent is never a bad idea. Even if Tesla is weeks away from FSD (which I'm sure they're not), they wouldn't stop there. There's always innovation to be had. Between this post and the comments on Electrek, I'm seeing terms like "desperate" bandied about. Of course Tesla isn't weeks away from FSD but to go from there to "desperate" just because they acquire a juicy pool of talent seems a bit extreme.

 

[larmor]

"Tesla has acquired a small artificial intelligence startup called DeepScale in order to help build its self-driving technology and reach its goal of building a giant fleet of robotaxis."

DeepScale focuses on Deep Neural Networks (DNN).

Tesla acquires AI startup to help build self-driving robotaxis - Electrek

Interesting. I hope this bears fruit.

And if it doesn't bear fruit maybe DeepScale can pivot shift to assist in removing that pesky hard water buildup on water features, common in southern california...

 

[heltok]

Karpathy talks about the network architecture search here:
Andrej Karpathy | Multi-Task Learning in the Wilderness

He has also tweeted about it here:

[URL='https://twitter.com/karpathy']Andrej Karpathy
@karpathy[/URL]

We see more significant improvements from training data distribution search (data splits + oversampling factor ratios) than neural architecture search. The latter is so overrated

My take is that the slides indicates that Tesla has been using DeepScale already to optimize their depth/width of their neural network. As Tesla is using a multitask learning they will also have to optimize which tasks shares what resources, maybe this was done using intuition but could be searched over. There are many more parameters Tesla can optimize such as multi-frame-input choices and other design choices. And you have to search over what data to balance and how much of each etc.

My guess is that Tesla was happy with using DeepScale for neural architecture search but realized that wanted even more things searched over and wanted to be able to direct the future work of the company to better fit their own needs and to speed up progress.

I think this bodes well for Tesla’s vision team and also for all other teams using neural networks. A lot of hand tuning will be replaced by search letting engineers focus on data and domain knowledge rather than deep learning specific work.

 

[diplomat33]

"Carver21 is purpose built to scale to your perception needs whether enabling safety features or delivering autonomous driving functions."

This should greatly help Tesla get to autonomous driving.

 

[electronblue]

"Carver21 is purpose built to scale to your perception needs whether enabling safety features or delivering autonomous driving functions."

This should greatly help Tesla get to autonomous driving.

Why... why... why do you say such things!?! You pick a vague talking point from the materials of a startup company and extrapolate from that wishful thinking on how it should ”greatly help” Tesla’s autonomous project?

Anyway, here’s my take on this acquisition: It is possible DeepScale has some technology component beneficial to Tesla. More likely they are a talent acquisition purchase that allows Tesla to lock-down key developers for a period of time. It could be both.

However, when adding new talent to software projects, that does not usually amount to instant acceleration since making software (let alone deep learning) doesn’t work like that. So it is unknown when the impact of this acquisition could affect Tesla’s products.

More likely it is a longer-term play rather than anything that will show its hand in the next quarter or two.

 

[electronblue]

Acquiring new talent is never a bad idea. Even if Tesla is weeks away from FSD (which I'm sure they're not), they wouldn't stop there. There's always innovation to be had. Between this post and the comments on Electrek, I'm seeing terms like "desperate" bandied about. Of course Tesla isn't weeks away from FSD but to go from there to "desperate" just because they acquire a juicy pool of talent seems a bit extreme.

I agree this acquisition could be unrelated to anything going on currently. It may be a longer-term play.

That said, it certainly is possible that Tesla has bitten more than they can chew with the current Autopilot suite and the quest to be Level 5 no geofence feature complete in 2019 and Level 5 robotaxi operational somewhere in the U.S. in 2020. Those are some lofty goals and some level of desperation at this stage does not sound impossible.

 

[diplomat33]

Why... why... why do you say such things!?! You pick a vague talking point from the materials of a startup company and extrapolate from that wishful thinking on how it should ”greatly help” Tesla’s autonomous project?

It's called expressing an opinion. And it's not based on nothing. A company that has DNN's for FSD would be useful for a company like Tesla that needs DNN's for FSD.

You do the same just in the opposite direction. You extrapolate that the news is bad for Tesla. Why? You have no evidence that it is bad for Tesla. But somehow any optimistic opinion is seen as hubris, but any pessimistic opinion is seen as "keeping it real". Both are opinions. Your negative opinion is not more valid than my positive opinion.

Like this:

That said, it certainly is possible that Tesla has bitten more than they can chew with the current Autopilot suite and the quest to be Level 5 no geofence feature complete in 2019 and Level 5 robotaxi operational somewhere in the U.S. in 2020. Those are some lofty goals and some level of desperation at this stage does not sound impossible.

You are putting a very negative spin on the article with no evidence whatsoever. But that's ok. It's your opinion.

 

[heltok]

After having watched all deepscales videos, another interesting aspect comes to my mind. It seems that their main product is neural architecture search. But to develop this as a product they had to implement almost an entire deep learning FSD stack. Most work seems to be done with public datasets and also some proprietary datasets from customers(read Tesla). They would make a great acquisition for auto manufacturers trying to catch up with Tesla. As the robotaxi is so valuable, increasing your chance of winning by taking away a few months for the competition might be worth it.

But mostly I think it’s their energy efficient neural architecture search that Tesla wants to have in-house!

 

[emmz0r]

It seems like DeepScale is about improving the NN performance, not drastically change the way they actually develop the features with some fancy new tech that will magically make everything perfect.

 

[electronblue]

This should greatly help Tesla get to autonomous driving.

It's called expressing an opinion. And it's not based on nothing. A company that has DNN's for FSD would be useful for a company like Tesla that needs DNN's for FSD.

You are certainly free to express your opinion, nothing against that. But I am genuinely curious — well, aghast really — why you would have such an opinion and thought I’d express my puzzlement.

I mean, little as we know of this acquisition and its reasoning, and you come up with the conclusion that ”this should greatly help Tesla get to autonomous driving.” To me that is hyperbole to the extreme, but more so I don’t get quite why you’d say that.

You do the same just in the opposite direction. You extrapolate that the news is bad for Tesla. Why? You have no evidence that it is bad for Tesla. But somehow any optimistic opinion is seen as hubris, but any pessimistic opinion is seen as "keeping it real". Both are opinions. Your negative opinion is not more valid than my positive opinion.

Where did I do that?

My points in this thread were the following:

1. This may be a talent acquisition. In Silicon Valley many startup or small company purchases are, very plausible. I find this likely.
2. This may be a component acquisition. That certainly is a plausible possibility, but I clearly left it as a possibility. Could be 1. and 2.
3. Talent purchased today is likely not helpful in achieving the aggressive timeline Tesla announced for robotaxis earlier this year. This is just basic software development math given the feature complete goal is less than three months away.

All of these are perfectly plausible scenarios — and more to the point very specific opinions that cover small areas of the news. None of them are necessarily bad scenarios. In fact, the news may well be good news, we just don’t know yet.

If I had done what you did, in the opposite direction, I would have simply quoted a bit of the PR and said something like this:

This should greatly hurt Tesla’s chances of getting to autonomous driving.

The hyperbole in my view is not helping your case.

 

[diplomat33]

You are certainly free to express your opinion, nothing against that. But I am genuinely curious — well, aghast really — why you would have such an opinion and thought I’d express my puzzlement.

I probably should have provided more detail and explanation to support my statement. I can see how just saying "this will be great for Tesla" without explaining why I feel that way, could be puzzling.

The reason I quote this one line, ""Carver21 is purpose built to scale to your perception needs whether enabling safety features or delivering autonomous driving functions." is because it highlights what DeepScale does as it relates to what Tesla is doing. DeepScale is doing work in the very area that Tesla needs. Specifically, DeepScale has expertise in DNN's that relate to autonomous features which matches with Tesla needs if they want to make their vision based NN approach work. DeepScale should, in principle, help Tesla finish their NN that they are developing on AP3 which is key if Tesla ever hopes to do any kind of autonomous driving using their vision based approach. No, I am not suggesting that DeepScale will magically make FSD happen. But their expertise in more efficient NN can certainly help Tesla perfect their NN and get them to work better, especially on AP3, which will certainly be useful IMO to help Tesla make progress. The reason I said "greatly helps" and not just "might help" or "should help" is because we all know that Tesla has some struggles with their vision approach. It's taken them awhile just to perfect a few pieces of the NN. They have an ambitious goal of robotaxis by end of 2020 and they need to put together the whole NN just to finish the perception part for "automatic city driving". So, DeepScale's expertise in DNN's might be just what Tesla needs to.

I hope that clarifies things.

 

[diplomat33]

Here's Teslarati's take on the acquisition:
Tesla's DeepScale acquisition is a play for efficient neural networks, faster OTA updates

Let me quote from the article the specific arguments for why this acquisition is good for Tesla:

1) Better optimization
"DeepScale’s tech could then help optimize Tesla’s hardware better, allowing the company’s vehicles to perform autonomous driving functions in a smoother and seamless manner."

2) Faster and more frequent OTA updates
"Forrest Iandola’s paper on SqueezeNet even mentioned Tesla’s over-the-air updates as an area where the hyper-efficient solution could make a difference. In the paper’s Introduction, the DeepScale CEO stated that SqueezeNet could provide less overhead when exporting new models to clients. "

"Smaller models require less communication, making frequent updates more feasible."

3) Carver21 is tailored fit for processing data from cameras like what Tesla does.
"DeepScale’s first product, a perception software called Carver21, is specifically designed to optimize the processing of data from a Full Self-Driving vehicle’s forward-facing cameras. This is pretty much tailor-fit for Tesla’s approach to autonomy, which relies on cameras, radar, and ultrasonic sensors to detect and respond to vehicles and obstacles on the road. "

@electronblue Those are 3 specific areas where DeepScale will be a big help to Tesla. That is why I stick by my statement that it will greatly help Tesla. There was a reason for my statement. I hope that clarifies things again.

 

[mongo]

Seems like they will be helpful in cramming as much feature set as possible into AP2.x HW.
May also have ways of compensating for the camera difference between 2.0 and 2.5+ (2.5 has blue).

 

[heltok]

2 seems very unlikely. OTA data is not that limited and Tesla will likely use the biggest neural network they can give hardware, not make it 10x smaller to save OTA bandwidth... Maybe they can have some smaller internal builds to iterate their datasets faster, but for customers they will want to max accuracy...

 

[AlMc]

I do not have the intellectual capacity to understand the implications of this acquisition but a friend broke it down for me. Hope this add (and does not detract) from the discussion:

They specialize in fusion of sensor data (for contextual AI-NN). Probably top of heap there or at least top 3. Don’t have time to detail currently, but think of it as taking ‘pixels’ from multiple cameras (and/or other sensors) and merging them (correlating differences) BEFORE making object recognition and other higher level determinations (which can bias NN decision making when made too early).
There’s not a perfect analogy, but it may help to think about multiple humans looking at the same scene, and instead of comparing each interpretation of what is seen, a single cortex combines each human’s retina data before deciding what the objects are (or are doing). It attacks not only edge cases but mimics our own brains BELOW the synaptic level where weighted chemical reaction influences when and if synopsis actually fire. The analogies aren’t perfect, but gives some contextual perspective.
As of a few years ago fusion of sensory data was impractical at the application level of higher order NN. Because the local processing had insufficient bandwidth and the power requirements to do so were prohibitive (hence this company’s expertise in processing under lower power and chip design to accomplish etc).
Important take-aways here:
1) This is a level of integration nobody else is even thinking of much less in a position to utilize, as only Tesla has a chip capable of this, and only Tesla is approaching L4/5 edge case and super intelligent decision making. This is 2nd and 3rd order (think differential equations) vertical integration
2) The knee jerk reports this is some sort of replacement of Tesla AI people who left is WRONG WRONG WRONG. This is a completely new level of skill set not related to what those other folks were even qualified to do.

 

[electronblue]

I probably should have provided more detail and explanation to support my statement. I can see how just saying "this will be great for Tesla" without explaining why I feel that way, could be puzzling.

The reason I quote this one line, ""Carver21 is purpose built to scale to your perception needs whether enabling safety features or delivering autonomous driving functions." is because it highlights what DeepScale does as it relates to what Tesla is doing. DeepScale is doing work in the very area that Tesla needs. Specifically, DeepScale has expertise in DNN's that relate to autonomous features which matches with Tesla needs if they want to make their vision based NN approach work. DeepScale should, in principle, help Tesla finish their NN that they are developing on AP3 which is key if Tesla ever hopes to do any kind of autonomous driving using their vision based approach. No, I am not suggesting that DeepScale will magically make FSD happen. But their expertise in more efficient NN can certainly help Tesla perfect their NN and get them to work better, especially on AP3, which will certainly be useful IMO to help Tesla make progress. The reason I said "greatly helps" and not just "might help" or "should help" is because we all know that Tesla has some struggles with their vision approach. It's taken them awhile just to perfect a few pieces of the NN. They have an ambitious goal of robotaxis by end of 2020 and they need to put together the whole NN just to finish the perception part for "automatic city driving". So, DeepScale's expertise in DNN's might be just what Tesla needs to.

I hope that clarifies things.

Thank you, it does. I would not have been puzzled by that.

One detail where I am not quite as convinced about DeepScale is whether or not they are very advanced in their computer vision. For example the video only showed simplistic 2D car recognition. I am thus a bit hesitant to assume they can bring anything helpful to Tesla’s table in that area... necessarily. They might. They might not.

The NN ”efficiency” work, if you will, as well as the team — additional talent — they bring might be more plausible reasons for the purchase rather than any ready-made vision components Tesla could just drop in. Just speculating based on very limited info of course.

 

[mswlogo]

"Tesla has acquired a small artificial intelligence startup called DeepScale in order to help build its self-driving technology and reach its goal of building a giant fleet of robotaxis."

DeepScale focuses on Deep Neural Networks (DNN).

Tesla acquires AI startup to help build self-driving robotaxis - Electrek

Interesting. I hope this bears fruit.

FSD Feature complete in 4 months !!!

It will take a minimum of a year before this bears fruit. And that's if "integration" goes smooth (which rarely does between software groups).

The fact that they needed to outsource (buy) this tells me, they don't feel they already have a good handle on it.

Perhaps robotaxis have a whole new set of problems to solve and they can be focused on that.