4 ms·
(I gave the talk.) We are already starting to see the nature of data changing. Unsupervised learning is starting to work — see https://blog.openai.com/language
by gdb 8y ago
(I gave the talk.)
We are already starting to see the nature of data changing. Unsupervised learning is starting to work — see https://blog.openai.com/language-unsupervised/ https://blog.openai.com/language-unsupervised/ which learns from 7,000 books and then sets state-of-the-art across almost all relevant NLP datasets. With reinforcement learning, as you point out, you turn simulator compute into data. So even with today's models, it seems that the data bottleneck is much less significant than even two years ago.
The harder bottleneck is transfer. In most cases, we train a model on one domain at a time, and it can't use that knowledge for a new related task. To scale to the real world, we'll need to construct models that have "world knowledge" and are able to apply it to new situations.
Fortunately, we have lots of ideas about how this might work (e.g. using generative models to learn a world model, or applying energy-based models like https://blog.openai.com/learning-concepts-with-energy-functions/ https://blog.openai.com/learning-concepts-with-energy-functi...). The main limitation right now: the ideas are very computationally expensive. So we'll need engineers and researchers to help us to continue scaling our supercomputing clusters and build working systems to test our ideas.
- iotb 8y agoWhat are your thoughts on Starting over completely from scratch as Geoffrey Hinton has suggested? What are you doing as a group to attract and bring on such individuals? Does this occupy any portion of your efforts at OpenAI? If you were given a demo of an AI system that uses a completely new/revolutionary approach towards various different problems with success, how open would you be to rethinking your position on 'Optimization techniques'? Modeling seems as a stop-gap towards getting over the limitations of Weak AI. As I recall, this is what knowledge-based expert systems tried in a time's past and failed at because it's nothing but a glorified masking of the underlying problem with limited human inputted rulesets. I don't agree with Yann LeCun that the way forward to AGI is modeling. I feel like it's the best solution people worked up towards the limitations of Weak AI which were broadly and publicly acknowledged in 2017 and early 2018. > The main limitation right now: the ideas are very computationally expensive. This is because the fundamental core set of algorithms being used by the industry are fundamentally flawed yet favorable to big data/cloud computing.. A quite lucractive business model for currently entrenched tech companies. It's why they spend so much effort ensuring the broad range of AI techniques fundamentally stay the way they are.. because if they do, it means boat loads of money for them. > So we'll need engineers and researchers to help us to continue scaling our supercomputing clusters and build working systems to test our ideas. When you're attempting to resolve something and you are shown YoY that it isn't being resolved and requires even more massive amounts of compute, it means you're doing something wrong. It will be better to take a step back an re-evaluate your approach fundamentally. Again, what is the willingness you have to do so if shown something far more novel?