4 ms·

With Neural Nets, training the AI to handle emergencies will be all about exposing as many emergency situations as possible. What's better about an AI powered

With Neural Nets, training the AI to handle emergencies will be all about exposing as many emergency situations as possible. What's better about an AI powered by neural nets is that you could train an AI to go offroading. Get enough data and you've got a model for dealing with a given situation. Google's biggest strides with OCR, Voice recognition, Spam filters and other AI tech early on came from its ability to gather a huge corpus of data. The real challenge is two fold. Gathering data, and feeding the AI with inputs with data actually matters. This is the secret sauce that Hotz refers to in the article as the information he's not willing to disclose. That information will become commoditized in due time (like low-latency optimization for HFT), but it will take plenty of institutional money & experience (Google, Apple, Tesla, Ford, etc) to get it there.

nradov 11y ago
The trouble with using neural nets for safety-critical real-time systems is that it's really hard to do the necessary level of validation. You can't accurately predict how the system might react in totally novel or unexpected situations. Which isn't to say that human drivers handle those situations well, but most of the time they don't do something totally bizarre or dangerous.
- analyst74 11y ago
  Humans totally do things that are bizarre or dangerous when in shock, but we've come to accept that as a personal responsibility and a price the society has to bear.
  - zymhan 11y ago
    Exactly. People may swerve and over-correct, causing their car to flip, for example.
    - nradov 11y ago
      Electronic stability control (standard in US passenger cars since 2012) has already mostly solved that problem. http://www.safercar.gov/Vehicle+Shoppers/Rollover/Electronic+Stability+Control http://www.safercar.gov/Vehicle+Shoppers/Rollover/Electronic...
  - yongjik 11y ago
    Because we can't re-engineer humans into rule-based automata. (And we probably shouldn't even if we could.) Machines, on the other hand, are a different story.
  - romaniv 11y ago
    We have millennia of experience in regards to estimating how people will react to various shock situations and what constitutes those situations. It's intuitive. The reactions of NN to unusual stimuli are likely to be counter-intuitive at best and unpredictable at worst.
- foota 11y ago
  Doesn't it seem possible that we could start testing the cars in a simulated environment?
  - nradov 11y ago
    Simulated environments aren't accurate enough (inputs are too clean, other drivers don't act real, etc) and would end up training the software to do the wrong things. A more reasonable approach would be to record the activities of multiple safe human drivers across a wide range of situations and then train the software to act like them.
    - Houshalter 11y ago
      He said for testing. Not for training the neural network. But just seeing how it behaves in various situations, to find it's flaws, and see if it's ready for the road.
      - argonaut 11y ago
        Sure, but the real world is 100x more complex than simulation.
- bparsons 11y ago
  Human error when driving a vehicle is one of the top causes of premature death globally. That is what we should be measuring the technology against, not perfection. It seems that the technology has already, or is very close to approaching human levels of proficiency on the road. If specific use cases (offroad, snowpacked road etc.) are problematic, they can be limited or prohibited in the mean time.
argonaut 11y ago
Have you built deep learning models before? Neural nets are not magical boxes that you stick data into and instantly get great, generalized, robust models at the other end. I have no inside knowledge, but I would be very surprised if Google's self-driving cars use neural nets as a significant component right now (which isn't to say there aren't people exploring its use).
vkou 11y ago
Using neural nets to deal with emergencies runs you into the Anna Karenina problem - "All happy families are alike; each unhappy family is unhappy in its own way." It's fairly easy to train, and verify a system for driving in well-behaved traffic. Unfortunately, the problem space of not-well-behaved traffic is far wider - and is very hard to gather enough data to train a system well. What you're going to get is self-driving cars which handle 99% of driving just fine - and when they end up in emergencies, find the human 'driver' to have dozed off at the wheel. (All-in-all, their safety record might end up better then the status quo - but that's not a certainty.)

nradov 11y ago

The trouble with using neural nets for safety-critical real-time systems is that it's really hard to do the necessary level of validation. You can't accurately predict how the system might react in totally novel or unexpected situations. Which isn't to say that human drivers handle those situations well, but most of the time they don't do something totally bizarre or dangerous.

analyst74 11y ago
Humans totally do things that are bizarre or dangerous when in shock, but we've come to accept that as a personal responsibility and a price the society has to bear.
- zymhan 11y ago
  Exactly. People may swerve and over-correct, causing their car to flip, for example.
  - nradov 11y ago
    Electronic stability control (standard in US passenger cars since 2012) has already mostly solved that problem. http://www.safercar.gov/Vehicle+Shoppers/Rollover/Electronic+Stability+Control http://www.safercar.gov/Vehicle+Shoppers/Rollover/Electronic...
- yongjik 11y ago
  Because we can't re-engineer humans into rule-based automata. (And we probably shouldn't even if we could.) Machines, on the other hand, are a different story.
- romaniv 11y ago
  We have millennia of experience in regards to estimating how people will react to various shock situations and what constitutes those situations. It's intuitive. The reactions of NN to unusual stimuli are likely to be counter-intuitive at best and unpredictable at worst.
foota 11y ago
Doesn't it seem possible that we could start testing the cars in a simulated environment?
- nradov 11y ago
  Simulated environments aren't accurate enough (inputs are too clean, other drivers don't act real, etc) and would end up training the software to do the wrong things. A more reasonable approach would be to record the activities of multiple safe human drivers across a wide range of situations and then train the software to act like them.
  - Houshalter 11y ago
    He said for testing. Not for training the neural network. But just seeing how it behaves in various situations, to find it's flaws, and see if it's ready for the road.
    - argonaut 11y ago
      Sure, but the real world is 100x more complex than simulation.
bparsons 11y ago
Human error when driving a vehicle is one of the top causes of premature death globally. That is what we should be measuring the technology against, not perfection. It seems that the technology has already, or is very close to approaching human levels of proficiency on the road. If specific use cases (offroad, snowpacked road etc.) are problematic, they can be limited or prohibited in the mean time.

argonaut 11y ago

Have you built deep learning models before? Neural nets are not magical boxes that you stick data into and instantly get great, generalized, robust models at the other end. I have no inside knowledge, but I would be very surprised if Google's self-driving cars use neural nets as a significant component right now (which isn't to say there aren't people exploring its use).

vkou 11y ago

Using neural nets to deal with emergencies runs you into the Anna Karenina problem - "All happy families are alike; each unhappy family is unhappy in its own way." It's fairly easy to train, and verify a system for driving in well-behaved traffic. Unfortunately, the problem space of not-well-behaved traffic is far wider - and is very hard to gather enough data to train a system well. What you're going to get is self-driving cars which handle 99% of driving just fine - and when they end up in emergencies, find the human 'driver' to have dozed off at the wheel. (All-in-all, their safety record might end up better then the status quo - but that's not a certainty.)