5 ms·

Many of these top tier students are moving to the big companies not just because of the salary, but also because they often have resources that is better than a

by k2xl 9y ago

Many of these top tier students are moving to the big companies not just because of the salary, but also because they often have resources that is better than any university can offer. I don't think innovation is being stifled. Every week there is a new amazing publication on AI/ML, but instead of coming from University it sometimes comes from DeepMind , Google, FB, etc... If you want to research self driving cars, wouldn't you want to work at Google which has tons of data?

CJefferson 9y ago
The problem is, all this company research is locked behind secrets and patents. No-one outside Google knows how deepmind works. No-one can reproduce it.
- mathgenius 9y ago
  They published their Go algorithms in Nature. Also they put stuff into NIPS. So I don't think this qualifies as "No-one outside Google knows how deepmind works." If their stuff can't be reproduced, then it isn't science and the journals shouldn't be accepting papers from them. Or am I not understanding something?
  - kuschku 9y ago
    Try getting the training dataset they used. For anything, actually. They train all their networks on data taken from users, and publish neither the data nor the models. And the scientific world doesn't have the resources to replicate anything close to it.
    - mathgenius 9y ago
      Alpha Go (the earlier versions) was trained on publicly available games.
      - kuschku 9y ago
        That is a lucky case then, because for their NLP projects nothing is publicly available. It's quite annoying if you can't get anything.
    - Ao7bei3s 9y ago
      So? Try getting any raw data and/or code from academics. In my experience, often all you ever get are the papers. Now the tables are flipped and academics are the ones not getting access. A culture of real open access would be nice.
  - CJefferson 9y ago
    In my opinion, it's not reproducible. The stuff in Nature is no-where near enough to reproduce it. They won't give out their trained networks, so we can check how they perform. We have to dig around for scraps of games AlphaGo has played. As to why it got published? I don't think it should have been.

by k2xl 9y ago

CJefferson 9y ago

The problem is, all this company research is locked behind secrets and patents. No-one outside Google knows how deepmind works. No-one can reproduce it.

mathgenius 9y ago
They published their Go algorithms in Nature. Also they put stuff into NIPS. So I don't think this qualifies as "No-one outside Google knows how deepmind works." If their stuff can't be reproduced, then it isn't science and the journals shouldn't be accepting papers from them. Or am I not understanding something?
- kuschku 9y ago
  Try getting the training dataset they used. For anything, actually. They train all their networks on data taken from users, and publish neither the data nor the models. And the scientific world doesn't have the resources to replicate anything close to it.
  - mathgenius 9y ago
    Alpha Go (the earlier versions) was trained on publicly available games.
    - kuschku 9y ago
      That is a lucky case then, because for their NLP projects nothing is publicly available. It's quite annoying if you can't get anything.
  - Ao7bei3s 9y ago
    So? Try getting any raw data and/or code from academics. In my experience, often all you ever get are the papers. Now the tables are flipped and academics are the ones not getting access. A culture of real open access would be nice.
- CJefferson 9y ago
  In my opinion, it's not reproducible. The stuff in Nature is no-where near enough to reproduce it. They won't give out their trained networks, so we can check how they perform. We have to dig around for scraps of games AlphaGo has played. As to why it got published? I don't think it should have been.