As illustrated by your examples, if all you are interested in is noticing a red flag, then the problem is simple, just notice any red flag. But, if the two systems are giving you entirely different views of reality, then it is a whole different class of problem.
Yeah, we had discussion on camera-only vs multi-sensor upthread.
Tesla seems pretty alone in camera-only, except when they do their “side work” with lidar.
Here’s another take:
Conclusion:
All of the above leads us to the following summary:
- Both approaches will be successful;
- The Waymo approach is much more expensive today than Tesla’s approach. The former is only affordable for robotaxi fleets, the latter for all cars (private cars);
- In the short term, the Waymo approach is safer and a geofenced region can be put into operation more quickly. In the medium and long term, the Tesla approach will also be safe enough thanks to ever-advancing algorithms;
- Waymo is targeting robotaxi fleets, Tesla is targeting both private cars and robotaxi fleets;
Who can capture more market share depends above all on who wins the race to answer the following question: Will a Waymo sensor stack become cheaper more quickly and therefore ready for use in many vehicles, or will the algorithms in the Tesla approach become better and safer more quickly?
1 Like
OK, here’s a hypothetical. Your car monitors tire pressure. If it slips by a pound or two, should the car immediately pull over and stop? Why or why not? Suppose the pressure suddenly drops by 25 PSI. Same question: why or why not?
These sorts of decisions are already embedded in the software, and there is some point at which it says “red flag.” Here’s another: my “auto” wiper setting goes slow for a drizzle, fast for a heavier rain, so presumably it’s a sensing monitor for both “the condition” and “the state of condition.” In self-drive mode, should the car slow down for drizzle? How about for downpour? Another decision which the software is entirely capable of making, even as it is watching for street lights, stop signs, speed limit signs and more.
There’s lots of examples, they’re all over the place of “conditional condition”. It isn’t as though your only choices are “stop” or “go”.
3 Likes
Yeah, but this is what machine learning models are designed to solve.
Fundamentally, AV driving is an optimization problem (as is machine learning).
If the multi-sensor model is no better than a single sensor model, then go single.
But at least considering multi-sensor model gives you a larger solution set to explore, and you should do no worse than any single sensor solution and you might do better, of course.
I gave some explanation way upthread:
1 Like
I don’t know. I just said it could be the more difficult task…not impossible. And more difficult could be solved with just more HW and/or more expensive HW. Waymo doesn’t disclose details on their HW, but does say some of it is custom. Of course Tesla has custom ASIC for inference – but then they make millions of them.
Mike
2 Likes
Again, these are simple binary decisions, not differences in world view.
I agree that software is important, but until optimal hardware configuration(s) is/are identified, optimizing and iterating on hardware is part of the r&d process - and that’s what we are seeing in the solutions of different companies: different sensor configurations.
So I believe @albaby1 is framing the state of the problem correctly.
In general, the AI driving problem is to “measure the environment (using data from some sensor configuration), process the environmental measures from analog to digital data, and calculate the driving decision with a mathematical formula using the digital environmental data”
The machine learning problem is an optimization problem to “choose the sensor data and an associated data processing and mathematical formula for driving decisions so as to meet minimum safety criteria and satisfy minimum driving performance criteria while possibly minimizing economic costs.”
Hardware design decisions and how to collect and process data from various sensor designs are very much a part of solving the above problems.
1 Like
Pretty much every company has moved from mathematical formula/algorithmic based driving policy to AI / neural net / transformer based driving policy.
Instead of programmatic if-then-else, there are weights assigned to parameters. And, there are a ton of parameters for autonomy.
No, absolutely not so (except for maybe how you are thinking about “formula”).
This is why I gave a definition of the AI driving problem above, including framing it as an optimization problem and representing the AI calculation as just a formula, so we can all be talking about the same things in simple terms (the importance of defining terms and not using too much technical jargon).
Companies that use neural networks (NN) and any flavor of machine learning absolutely have not stopped (“moved from”) using formulas.
Let’s briefly explain NNs.
A NN is a collection of mathematical formulas that takes input and returns output.
The parameters are part of the formulas, like slope and intercept are parameters for a straight line.
A neural net is absolutely a mathematical formula, a very complex one (with many parameters) that isn’t simple to write down with paper and pencil, but it is just a mathematical function with an input and an output (which can all be multivariate).
I probably should have written function instead of formula to be more technically precise, but I was trying to use everyday language.
The AI driving data flow is:
- Data on the environment is captured from sensors in digital form.
- A mathematical function (NN) operates on data as input.
- The output of the function is a driving decision.
Your knee-jerk reaction to disagree with me aside, what I said:
Groups “formula” in with algorithmic expression, which AI certainly is not. And with which you actually agree:
I think everyone here is as comfortable with “function” as they are with “formula,” maybe even more so. You should have just led with this, but whatever.
As for what’s involved in AI driving flow, I won’t bother correcting you in detail, but will say that’s far too simplistic and even on a basic level overlooks other important data inputs.
Well, I’m not super clear on what exactly your point is about formulas, functions and algorithms, and companies moving from them, but neural networks and other machine learning calculations are just functions that take input and return output.
Learning can be done on any function with parameters (or hyperparameters) that can be adjusted, including sequences of calculations in formulas, which could be easily called and considered and described as algorithms.
Complex models like NNs are just many calculations, some sequential, like what one might call an algorithm.
I see machine learning and the word algorithm used together all the time.
I think the important distinction is the learning step, not the name for the calculation (algorithm, transformer, whatever).
The learning step is in the optimization step I described above:
These comments are not helpful to furthering the discussion and they are not necessary.
1 Like
I think the distinction which is being attempted here is that the word formula implies the sort of thing that one might write out and twiddle with the factors until the desired result was achieved, i.e., one could actually see and understand its components. Whereas, the NN is thousands of parameters which no human can hope to understand, much less intelligently tweak.
1 Like
Right, agreed.
I tried this wording upthread, that it wouldn’t be simple to write down:
There’s a famous (in AI circles anyway) white paper by Rich Sutton called “The Bitter Lesson
.” Here’s a quote:
The bitter lesson is based on the historical observations that
- AI researchers have often tried to build knowledge into their agents,
- this always helps in the short term, and is personally satisfying to the researcher, but
- in the long run it plateaus and even inhibits further progress, and
- breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning.
The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.
Algorithms are how computer scientists express human knowledge in human-written machine code. My short-cut for that was the “if-then-else” expression, and an example might be “if the car is to make a right hand turn, then change into the right lane, else stay in the middle lane.” Another use would be for obeying rules of the road - there are thousands of such possible conditionals.
Mobileye wrote a paper claiming the exact opposite of The Bitter Lesson:
The premise is to switch from a well-engineered system comprised of data-driven components interconnected by many lines of codes to a pure data-driven approach comprised of a single end-to-end neural network. …In summary, we argue that an end-to-end approach is neither necessary nor sufficient for self-driving systems.
Rather than trying to conflate human-created algorithms and Machine-Learned parameters, it’s thus important to understand just how differently these are created and improved. Saying one can “learn” how to adjust co-efficients in human-created mathematical algorithms is the same as the training phase of AI which automatically adjusts the parmeters is not helpful and hides just how different these operations are.
A neural-net system is not a “complex mathematical formula.” Sure, like any computer program, it has inputs and outputs, but they are more different than alike, and those differences matter in any discussion of advancing the technology of autonomy.
Well, I think we might have irreconcilable differences.
And I understand (I think) what you mean by what some might call a “rules based” algorithm.
Nothing I wrote upthread was intended to say that a rules-based algorithm, in and of itself, is equivalent to machine learning in any sense. In fact, I was not making any statements about algorithms until you introduced them.
My main points:
Assuming you mean function and not formula (I am using the word function to be more precise, as mentioned upthread), this is a false statement, period.
I’m surprised you would write this.
Do you really mean it?
The following statements are true.
And these are not controversial or extreme statements, by the way.
It’s borderline silly to be writing these down, just ask the internet or consult some introductory books.
- A neural network is a mathematical function, albeit complex, that takes input and returns output.
- A sequence of calculations with parameters and hyperparameters that takes input and returns output, which can be named an algorithm, can be trained using machine learning.
One source (https://www.cs.princeton.edu/courses/archive/fall19/cos484/lectures/lec6.pdf), it has lots of keywords like neural network, input, output, function, algorithm, optimization, and there are a ton of formulas.
Other points:
I wouldn’t conflate the two. These are your words.
Addressing these in a separate reply, maybe we can better understand each other.
It doesn’t hide anything from me. It doesn’t make it less helpful to me.
So we have 1 vote less helpful and hiding stuff and 1 vote not less helpful and not hiding stuff.
We are tied.
In your view, and specifically with respect to machine learning, what are the most important ways in which these operations are different?
(Above is your phrasing, I would use the word function instead of formula.)
In your view, which differences matter the most in advancing autonomy technology?
Agreed. This is over.