3 ms·

All deep learning implementations to date fail on the last 2% of real world cases. Why would deep learning be the solution for its own shortcomings?

ethbr0 6y ago
Failing on the last 2% (or whatever small number we want to set) isn't a shortcoming of deep learning: it's a shortcoming of all prior knowledge systems. Humans included. And the barrier to surpassing that problem has been hardware (now approaching solving) and calendar- & person-time. The problem with 1970s AI (ingnoring hardware) isn't that we couldn't build useful AI, but that we couldn't afford to build useful AI in most domains. Calendar time costs money. People time costs money. Deep learning and derivates finally provide a solution where we can substitute machine time and data volume for calendar and people time. As a result, we can afford to do things in domains we couldn't before.
- klyrs 6y ago
  Forget this 2% business. How does it do in snowy conditions at night? Because in some places, that's around 12.5%.
- justapassenger 6y ago
  > Failing on the last 2% (or whatever small number we want to set) isn't a shortcoming of deep learning Citation needed. Deep learning is used to power some of the biggest money making machines in the world (like google or facebook ads) and they're still bad. It's also used to power products like voice recognition and it'll still add dog to my shopping list instead of soap.
  - ethbr0 6y ago
    I would counter that those systems are as good as they need to be. Google and Facebook have realized advertisers are dumb. So why waste a ton of people time optimizing past what moves the revenue needle? Where I'd expect to see interesting developments (hopefully published openly) is in sentiment analysis / disinformation identification, as companies begin to see it as an existential threat to the unregulated nature of their businesses and fund accordingly.

ethbr0 6y ago

Failing on the last 2% (or whatever small number we want to set) isn't a shortcoming of deep learning: it's a shortcoming of all prior knowledge systems. Humans included. And the barrier to surpassing that problem has been hardware (now approaching solving) and calendar- & person-time. The problem with 1970s AI (ingnoring hardware) isn't that we couldn't build useful AI, but that we couldn't afford to build useful AI in most domains. Calendar time costs money. People time costs money. Deep learning and derivates finally provide a solution where we can substitute machine time and data volume for calendar and people time. As a result, we can afford to do things in domains we couldn't before.

klyrs 6y ago
Forget this 2% business. How does it do in snowy conditions at night? Because in some places, that's around 12.5%.
justapassenger 6y ago
> Failing on the last 2% (or whatever small number we want to set) isn't a shortcoming of deep learning Citation needed. Deep learning is used to power some of the biggest money making machines in the world (like google or facebook ads) and they're still bad. It's also used to power products like voice recognition and it'll still add dog to my shopping list instead of soap.
- ethbr0 6y ago
  I would counter that those systems are as good as they need to be. Google and Facebook have realized advertisers are dumb. So why waste a ton of people time optimizing past what moves the revenue needle? Where I'd expect to see interesting developments (hopefully published openly) is in sentiment analysis / disinformation identification, as companies begin to see it as an existential threat to the unregulated nature of their businesses and fund accordingly.