Not only that, if another FSD crash was to happen and an investigation was to be looked into the system, can we still get the AI in FSD to explain itself thoroughly on the decisions made which lead it to say 'crash right into another car right in-front of it' [0][1] or malfunction when it 'confuses the moon with a traffic light' and slow down on the highway [1]?
I don't think you would want to sign up to be a crash dummy on beta quality safety critical software, especially when the AI is a complete black box?
I’m certainly not a fan of teslas marketing or approach to self-driving but I don’t believe their FSD is anywhere near an end-to-end black box model.
Given that it’s not end-to-end, it’s pretty simple to determine whether an object was incorrectly identified, missed, etc. Why the model made an incorrect prediction on one input vs another, that’s mainly speculation and testing.
This article makes a lot of pretty bold claims (not only pointing out a flaw but suggesting that there is a better alternative) without posting their own AI's image recognition results based on spike timing.
> One thing is certain; a new AI model based on spike timing will pop out of nowhere and displace the existing paradigm.
Ok this is one thing which admittedly annoys me: is this certain, as in 99.999% probability? And with what time frame? Would the author be willing to enter a bet where I wager $1 and win a billion dollars (since it has positive expected value for a "certain" event)?
From the link - "Isaac Newton had a cause for acceleration (force) but none for inertia."
Funny, because physics explains that all objects of mass have inertia. Heavy objects traveling at speed have inertia. Faster speeds? More Inertia. Heavier objects? More inertia, and we can calculate inertia : I = mr2.
I doubt that this author has any credibility for anything related to physics.
It’s also pretty bold to sell “full self driving”, claim that teslas increase in value because they’ll become robo-taxis and you’d be insane not to buy one, and claim that teslas can drive coast to coast autonomously, when the reality is that FSD isn’t even at the level of that faked “paint it black” sf drive video yet.
Jeff Hawkins work on predictive memory models the cerebral cortex and works based on comparing input signals that are close in time, which absolutely could be described as "spike timing."
> This is a criticism that the late philosopher Hubert Dreyfus made for half a century to no avail.
Whenever someone says a particular machine learning application is not going to work, and their justification is something a philosopher said, I always laugh.
I’m sure this philosopher was a great thinker with many interesting ideas about human nature, but likely they had little familiarity with convolution neural networks.
It’s unclear why the criticism in the paper does not apply to many existing “AI” technologies.
For example, clearly, video filters (like those developed by Apple, Meta, Snapchat, etc) that create a cartoon figure that mimics the movement of a users face are impossible since there is a temporal component and we cannot possibly perfectly represent every single face perfectly? Tesla self driving cars already work. Sure they might crash sometimes, but humans also crash sometimes. At this point is only a matter of time until the crash rate is deemed acceptable
Dreyfus was a very intelligent man. He knew the difference between representational and non-representational systems. As Yann LeCun has said many times, deep learning is the learning of representations. That's all anyone needs to know in order to understand Dreyfus's thesis. The brain does not need a prior representation of a bicycle to perceive a bicycle. A DNN would be blind to a bicycle without a prior representation. Did you read the article?
What exactly does it mean to "perceive a bicycle"? Noticing shapes and colours and recognising them as a distinct object? Recognising an obstacle? Noticing qualities like smoothness and straightness and associating the concept "man-made"? Being able to explain its purpose? Predicting how it might move, if ridden by a person?
Yes, pretty much. In my opinion, perception is generalization. To perceive a bicycle is to perceive many types of qualities or properties about it that can also be applied to a potentially infinite number of other objects. The brain can perform this generalization instantly without having stored previous representations (bicycle patterns) in memory.
A great example of non-representational intelligence is the honeybee's brain. It has less than 1 million neurons but it can handle zillions of patterns/objects in its 3D environment. It would be impossible for it to store all those zillions of patterns in its tiny brain. It uses the world itself as its own model.
For these reasons, deep learning is irrelevant to AGI.
I agree the criticism on the basis of philosophy is kind of silly. Elon Musk's recent description of how the thing is going in real life on Lex Fridman Podcast 28 dec 21 was interesting https://youtu.be/DxREm3s1scA?t=3972
The author of this post laments how deep learning works, but the suggestion is to solve the problem somehow by only reasoning about how other neurons are firing locally. Isn’t that exactly what deep neural networks already do? A basic feedforward layer has individual neurons that fire if their inputs exceed a threshold, which is based on the timing of the firing of the other neurons.
The article presents a vague non-solution.
Of course, deep learning has not (yet) cracked full self-driving or general cognition, but without providing more examples or differentiators, this post is somewhat lacking.
Every so often I'll be driving and see something kind of strange before my brain processes and disambiguates the situation. In those moments I think about how current technology could never process this situation or environment like my human brain just did. Current car and road infrastructure is made for humans and requires human intelligence to operate within. Without changing this infrastructure or vastly more capable AI self driving will never work.
Machines have a lot of advantages though like perfect, practically infinite memory, internet connectivity, the ability to learn from the cumulative experience of millions of humans, specialized sensors, omnidirectional vision, never getting tired, never being inattentive, never being intoxicated.
I wouldn't bet against machines' ability to brute force this. The failure modes will be different and scary, but much safer on average.
No, that's silly. That'd be like saying we won't know whether time-travel is possible without trying or won't know if drinking urine can cure Covid without trying. A feasible project is something that has a plausible path to finish line based on existing body of knowledge without breaking basic laws of physics or economics.
From what I know of Tesla AI efforts they are doing essentially something like this.
The AI isn't trained to recognize a bike, person, car etc it's trained to recognize where it can drive.
I also know that they have revised the neural architecture more than once and that what happened 6 years ago is not representative of what is happening now.
There are so many different ways that this can play out but assuming Tesla cannot figure it out or that everyone is doing the same thing is a mistake.
I don't really see how philosophy is all that relevant. Maybe somewhat relevant, but AIs are not brains. In fact we use computers specifically because they aren't brains, for things that would be expensive
What's the track record in general like for "philosophy-driven development"? It has big marketing success like Apple, and Chuck Moore's FORTH chips have interesting low power capabilities, but 99% of the world seems to run on tech that's built using the usual tech thought process.
Lots of cars can automatically stop to prevent a collision, using what's probably just a fancy version of if(goingToCrash){dont()}, based on LIDAR, Radar, and ultrasound.
I usually hear nothing but praise from safety researchers. Unless there's some coverup I'm not aware of, it seems like It Works.
Why could the Tesla in the mentioned accident not do that? Why does it need to recognize a person? It just needs to know that it's bad to hit stuff that reflects radar.
Sure, maybe a prankster could toss some corner reflectors on the road and jam up traffic or something, but that's the price of safety, and AI has adversarial inputs too. Just add some redundant sensor types to make it harder or a manual override.
Choosing to rely on brain-like AI is philosophy driven development, assuming the brain is always best at every task.
Instead, we can change the task itself. Not "drive a car as a human would" but "Transport passengers with as close to zero risk as we can, at all cost".
AFAIK every other industry with that goal is using deterministic code on the level of GOFAI or below, with the exception of a very few deep learning systems to detect certain things like fire, probably combined with deterministic code as a fallback.
AI only needs to be like 95% reliable. With some of the advanced routing stuff they could do, weird stuff like cars randomly stopping for 10s isn't going to bother anyone.
Even an occasional missed speed limit sign or running a red light or not seeing lane markings shouldn't matter, if there is trusted deterministic code that makes sure it never hits anything.
Current self driving cars are making progress. What's actually missing besides the extreme level of safety people expect? They seem to be slowly getting better at working in varied environments and the like.
When was the last time your toaster tried to kill you? Are you worried autopilot will land a plane on your house? Why are we not aiming for that level of reliability, instead of emulating humans that currently have accidents all the time?
Rule based systems don't usually fail catastrophically when there is no rule, as long as they are able to detect that there is no rule. Then they just safely shut down.
Current cars already fail in various mechanical ways. The important thing is making sure nobody dies, not guessing whether something is a bike or a dumpster. All you need to know is it's a thing to not hit.
>Using Heidegger as a starting point, he argued that the brain does not create internal representations of objects in the world. Rather, it learns how to see the world directly.
That's not how vision works. There's a small high resolution foveated area that is scanned to look at regions of interest, while the peripheral vision does a far lower resolution perception of the rest of the field of view. Everything that our brain does is a model of the world, not a direct perception.
If you're in a room, you can't look at all the corners at the same time, yet you know where they are, to a fair degree of precision, and can point to them with your eyes closed once you're oriented, even the corners behind you.
>Deep learning experts essentially found a way (backpropagation, gradient descent, fast computers and lots of labeled or pre-categorized data) to create the rules automatically. The rules are in the form, if A then B, where A is a pattern and B a label or symbol representing a category.
Not really, they are function approximators, very good ones. However... the next point is still true, none the less
>The problem with expert systems is that they are brittle. Presented with a situation for which there is no rule, they fail catastrophically.
Functions which work reliably over a given set of inputs, are only valid for inputs very near that range. They do fail when unexpected inputs are presented. This is evident to any parent upon reflection of their experience.
This is why I've suggested many times, here and elsewhere, that Tesla needs a team to just make up unusual circumstances and aggressively try to get the self driving network to fail, and add that to the training data.
>Why the Brain Does Not Model the World
Just pain wrong, in so many ways. It's not an exact model of the world, but the model we have is good enough to predict things that effect our survival, which is all that really matters to evolution. Surprise and humor are what happen when a model of the world meets an unexpected input.
>A New AI Paradigm Will leave Deep Learning in the Dust
Discrete signal timing should be the main focus of AI research, in our opinion. It is the key to generalization.
Spike timing might generate a better model of the human brain, but that doesn't imply it will make better function approximators. If it can some how reduce the computational complexity of training, I'm all for it.
Eventually, with enough time and training, a neural network learns how to efficiently represent the world, and anticipate what's next. It's a question of training data and computation. Consider it takes a human 16 years to be considered adequate to the task of learning to drive, and you get a rough approximation of the complexity involved.
Eventually, full self driving will happen, and it's likely to all be neural nets doing the job.
I don't think you would want to sign up to be a crash dummy on beta quality safety critical software, especially when the AI is a complete black box?
[0] https://twitter.com/greentheonly/status/1473307236952940548
[1] https://www.theverge.com/2021/11/12/22778135/tesla-full-self...
[2] https://twitter.com/JordanTeslaTech/status/14184133078625853...