3 ms·

I’d be interested in where to look for more info on that. Do you think it could work on anything more complex? As I said in comment below I reckon I could mak

I’d be interested in where to look for more info on that. Do you think it could work on anything more complex? As I said in comment below I reckon I could make a much more elaborate explanation and still have the kid get it.

ShamelessC 5y ago
https://openai.com/blog/dall-e/ https://openai.com/blog/dall-e/ shows a decent ability to generalize to previously unseen concepts. You're correct to consider the complexity of the phrase and just how good humans are at this sort of thing without needing much "training". For now, concepts that aren't explicitly in the training set are effectively composed from those which are. This can lead to some bizarre and outright incorrect results, particularly when it comes to counting objects in a scene or with relative positioning between objects (e.g. a blue box on top of a red rectangle to the left of a green triangle) but it's early days and there's lots of progress happening all the time.
- jonplackett 5y ago
  Thanks. this is interesting. In a way it's opposite thought-direction of what we are talking about. eg. can it look at an avocado shaped chair and recognise it as a chair in the shape of an avocado - for me that would display a lot more understanding of the concept of 'chair' and 'avocado' than being able to produce an image of the phrase 'a chair shaped like an avocado' - but maybe the same process must be happening in there somewhere to make this possible? What do you think?
  - ShamelessC 5y ago
    https://openai.com/blog/clip/ https://openai.com/blog/clip/ CLIP is the corrolary model created purely for classification purposes rather than generative as in DALL-E and is quite impressive across a range of tasks. Give it an image and a caption, and in return you get a score (0.0 to 1.0) telling you how much they match. I think it is more in line with your premise. Others have taken CLIP and combined it with frozen language models (GPT2) to create automatic captioning models that are very impressive. edit: To try to address your question about whether or not actual semantic composition occurs, I think the answer is "yes" but it would be challenging to convince you this is true without going into details of the "self attention" mechanism which allows both methods to work. The short version is that these networks are able to find meaning in extremely high-dimensional problems by having a mechanism specifically tasked with learning positional statistics of the training data. In language this refers to e.g. how often the word "pillow" is directly next to the word "fort". In vision, this similarly refers to how often e.g. trees are positioned next to gift-wrapped presents. That's quite simplified but I hope that makes sense to a degree!
    - jonplackett 5y ago
      Thanks for writing such detailed reply! Very helpful. It’s exciting that this is happening. Do these models still have to be trained in the same way and then become essentially static? What I also find very different about teaching a child is how dynamically their understanding can shift based on explanations and observations. That’s probably the another hallmark of understanding.
      - ShamelessC 5y ago
        In statistics and machine learning this problem is called concept drift and it's an open research problem. Indeed all the models I listed are already out-of-date with the present state of the world; unless they are fine-tuned on new data. The regime for the above networks is called "pre-training". The idea here is that rather than training _directly_ on some challenging, specific task, you instead train a more generic task on a _lot_ of data. This gives you a "backbone model" that winds up being very strong on more specific tasks as well. In many cases, the ability to (cheaply) curate or create enough accurate data for specific sub-tasks might not even be possible. It's easier to scrape 400 million captioned images from the internet than it is to find/create millions of visual Q&A prompts/images. CLIP is a great example of this. While it was trained explicitly "just" to compare images and captions and to output a score of the cosine similarity between the features of the two - I have seen it approach effective state of the art on text-to-image generation tasks, image-captioning, Q&A, etc. This style of training needs to be updated after it is trained, typically because of the way these datasets are curated. Automatic curation via pretrained models is one option. Another option is to give the model direct access to the internet which is starting to be explored. Pre-training helps a lot with general distribution shift; but it's _probably_ not going to be able predict memes before they happen anytime soon. Reinforcement learning, on the other hand, requires an agent to learn in real and simulated environments. This obviously lends itself to re-training on-the-fly (and indeed, self-play and on-the-fly retraining are used heavily in practice). https://youtu.be/kopoLzvh5jY https://youtu.be/kopoLzvh5jY I have not yet fully grokked reinforcement learning but it is incredibly exciting research and is the correct direction towards making effective use of machine learning in robotics. Note that reinforcement learning and pre-training are not mutually exclusive and may be used in tandem.

ShamelessC 5y ago

https://openai.com/blog/dall-e/ https://openai.com/blog/dall-e/ shows a decent ability to generalize to previously unseen concepts. You're correct to consider the complexity of the phrase and just how good humans are at this sort of thing without needing much "training". For now, concepts that aren't explicitly in the training set are effectively composed from those which are. This can lead to some bizarre and outright incorrect results, particularly when it comes to counting objects in a scene or with relative positioning between objects (e.g. a blue box on top of a red rectangle to the left of a green triangle) but it's early days and there's lots of progress happening all the time.

jonplackett 5y ago
Thanks. this is interesting. In a way it's opposite thought-direction of what we are talking about. eg. can it look at an avocado shaped chair and recognise it as a chair in the shape of an avocado - for me that would display a lot more understanding of the concept of 'chair' and 'avocado' than being able to produce an image of the phrase 'a chair shaped like an avocado' - but maybe the same process must be happening in there somewhere to make this possible? What do you think?
- ShamelessC 5y ago
  https://openai.com/blog/clip/ https://openai.com/blog/clip/ CLIP is the corrolary model created purely for classification purposes rather than generative as in DALL-E and is quite impressive across a range of tasks. Give it an image and a caption, and in return you get a score (0.0 to 1.0) telling you how much they match. I think it is more in line with your premise. Others have taken CLIP and combined it with frozen language models (GPT2) to create automatic captioning models that are very impressive. edit: To try to address your question about whether or not actual semantic composition occurs, I think the answer is "yes" but it would be challenging to convince you this is true without going into details of the "self attention" mechanism which allows both methods to work. The short version is that these networks are able to find meaning in extremely high-dimensional problems by having a mechanism specifically tasked with learning positional statistics of the training data. In language this refers to e.g. how often the word "pillow" is directly next to the word "fort". In vision, this similarly refers to how often e.g. trees are positioned next to gift-wrapped presents. That's quite simplified but I hope that makes sense to a degree!
  - jonplackett 5y ago
    Thanks for writing such detailed reply! Very helpful. It’s exciting that this is happening. Do these models still have to be trained in the same way and then become essentially static? What I also find very different about teaching a child is how dynamically their understanding can shift based on explanations and observations. That’s probably the another hallmark of understanding.
    - ShamelessC 5y ago
      In statistics and machine learning this problem is called concept drift and it's an open research problem. Indeed all the models I listed are already out-of-date with the present state of the world; unless they are fine-tuned on new data. The regime for the above networks is called "pre-training". The idea here is that rather than training _directly_ on some challenging, specific task, you instead train a more generic task on a _lot_ of data. This gives you a "backbone model" that winds up being very strong on more specific tasks as well. In many cases, the ability to (cheaply) curate or create enough accurate data for specific sub-tasks might not even be possible. It's easier to scrape 400 million captioned images from the internet than it is to find/create millions of visual Q&A prompts/images. CLIP is a great example of this. While it was trained explicitly "just" to compare images and captions and to output a score of the cosine similarity between the features of the two - I have seen it approach effective state of the art on text-to-image generation tasks, image-captioning, Q&A, etc. This style of training needs to be updated after it is trained, typically because of the way these datasets are curated. Automatic curation via pretrained models is one option. Another option is to give the model direct access to the internet which is starting to be explored. Pre-training helps a lot with general distribution shift; but it's _probably_ not going to be able predict memes before they happen anytime soon. Reinforcement learning, on the other hand, requires an agent to learn in real and simulated environments. This obviously lends itself to re-training on-the-fly (and indeed, self-play and on-the-fly retraining are used heavily in practice). https://youtu.be/kopoLzvh5jY https://youtu.be/kopoLzvh5jY I have not yet fully grokked reinforcement learning but it is incredibly exciting research and is the correct direction towards making effective use of machine learning in robotics. Note that reinforcement learning and pre-training are not mutually exclusive and may be used in tandem.