Most routine tasks that everyone can do are impossible to automate with current ML technology, even after spending millions of dollars to collect training data
Show this thread
-
An example of this is Facebook discarding its M assistant project. The plan was to start with a Wizard of Oz, and collect tons of queries (with the steps to answer them), to eventually train a supervised model to answer the most common queries. Sounds clever, but it doesn't work
Show this thread
Another example is self-driving. You can automate driving, but not with an end-to-end ML system. The variability of real-world situations just cannot be handled by a system that merely pattern-matches with the past
-
-
At least not until some/most cars are networked and connected to cloud. This would provide a metric variable that would essentially guarantee autonomy.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Rule based systems are adequate for most decision support/automation. Advanced methods are useful esp an ensemble of different learners for the rest.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Very true. The brain does not learn complex patterns. It can instantly see and understand a new pattern that it has never seen before. In fact, almost everything it sees is new. It will not remember a pattern unless it is rehearsed many times.
-
This is why the late Hubert Dreyfus was fond of saying that the brain does not model the world. It senses it directly. The world is its own model. Unless our machines can do the same, we will not solve the AGI problem.
- Show replies
New conversation -
-
-
i agree that its often hard to do e2e, but often you can break the problem down into a set of roughly independent simpler tasks (simple enough for ML) that in aggregate give you something close enough to the e2e
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Not sure if I agree with the self-driving. Given a sufficiently good simulator, which I don’t think is impossible to build, I could imagine generating almost infinite amounts of data not seen in the past that cover the whole space.
-
Otavio Good showed off a remarkably impressive simulator at the recent O'Reilly AI conference in SF. There's a video available online if you're interested. Used a clever adversarial modeling approach for data augmentation. It covered just one track, but the idea could be expanded
End of conversation
New conversation -
-
-
Little reality check is healthy and long overdue.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
In the end, ML just arbitrages computation and data cost against developer cost and in many cases developers are still cheaper than the necessary computation and data.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
it is a good thing then that the methods which will win out cannot be characterized as supervised pattern-matches.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.