о машинном обучении
Jun. 12th, 2017 07:44 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
avva(для программистов в основном)
Райан Даль, известный в качестве создателя технологии Node.js, недавноушел в монастырь ушел на год заниматься глубинным обучением в Гугл, и по итогам этого стажерства написал довольно интересное резюме:
http://tinyclouds.org/residency/
Цитата: "I remain bullish that machine learning will transform essentially all industries and eventually improve the lives of every human. There are many industrial processes that can benefit from the smart guesses that ML provides. I believe my motivating demo will be achieved some day soon—you will watch Charlie Chaplin in 4K resolution and it will be indistinguishable from a modern movie.
That said, I've found it very difficult to build, train, and debug models. Certainly much of that difficulty is just my own inexperience, but that itself points to how much experience is needed to effectively train these beasts. My work has been focused on the easiest branch of ML: supervised learning. Even with perfect labels, developing models can be quite difficult. [...] If I use the word "working" in a subjective, gut-reaction way of describing software: Image classification seems to work robustly. Generative models barely work and are not well understood. GANs have great images, but are almost impossible to build—my experience has been that any small change to the architecture and it will just stop working. I've heard reinforcement learning is even more difficult. I can't speak to recurrent networks. [...]
The signal-to-noise ratio in papers is low. There's too much volume to keep up with. People are often not upfront about the failures of their models because conferences prefer accuracy over transparency. [...] It's an exciting time for ML. There is ample work to be done at all levels: from the theory end to the framework end, much can be improved. It's almost as exciting as the creation of the internet. Grab a shovel!"
Райан Даль, известный в качестве создателя технологии Node.js, недавно
http://tinyclouds.org/residency/
Цитата: "I remain bullish that machine learning will transform essentially all industries and eventually improve the lives of every human. There are many industrial processes that can benefit from the smart guesses that ML provides. I believe my motivating demo will be achieved some day soon—you will watch Charlie Chaplin in 4K resolution and it will be indistinguishable from a modern movie.
That said, I've found it very difficult to build, train, and debug models. Certainly much of that difficulty is just my own inexperience, but that itself points to how much experience is needed to effectively train these beasts. My work has been focused on the easiest branch of ML: supervised learning. Even with perfect labels, developing models can be quite difficult. [...] If I use the word "working" in a subjective, gut-reaction way of describing software: Image classification seems to work robustly. Generative models barely work and are not well understood. GANs have great images, but are almost impossible to build—my experience has been that any small change to the architecture and it will just stop working. I've heard reinforcement learning is even more difficult. I can't speak to recurrent networks. [...]
The signal-to-noise ratio in papers is low. There's too much volume to keep up with. People are often not upfront about the failures of their models because conferences prefer accuracy over transparency. [...] It's an exciting time for ML. There is ample work to be done at all levels: from the theory end to the framework end, much can be improved. It's almost as exciting as the creation of the internet. Grab a shovel!"
