I haven't been able to work on this competition at all but figured I'd at least share an idea I had a while ago. My idea is to utilize all of the many handwriting datasets available, normalize them in some consistent way (zero is black, similar resolution, etc.) and then train a multi-task network on them. We have access to quite a few such datasets already on Kaggle
\n\nUsing all of these datasets to train a model that captures all of the knowledge of handwritting \n and other languages it may be possible to make an even stronger generalizing model. All of these datasets could be combined and tasks trained in tandem a la Multi-Task Deep Neural Networks for Natural Language Understanding
\n\nThis could potentially be done as an initialization step to use instead of an imagenet pretrained or uninitialized CNN or done in a fashion similar to the paper linked above which was successful for language tasks.
", "messages": [ { "id": "767762", "postDate": "03/10/2020 04:18:06", "content": "I haven't been able to work on this competition at all but figured I'd at least share an idea I had a while ago. My idea is to utilize all of the many handwriting datasets available, normalize them in some consistent way (zero is black, similar resolution, etc.) and then train a multi-task network on them. We have access to quite a few such datasets already on Kaggle
\n\nUsing all of these datasets to train a model that captures all of the knowledge of handwritting \n and other languages it may be possible to make an even stronger generalizing model. All of these datasets could be combined and tasks trained in tandem a la Multi-Task Deep Neural Networks for Natural Language Understanding
\n\nThis could potentially be done as an initialization step to use instead of an imagenet pretrained or uninitialized CNN or done in a fashion similar to the paper linked above which was successful for language tasks.
", "rawMarkdown": "I haven't been able to work on this competition at all but figured I'd at least share an idea I had a while ago. My idea is to utilize all of the many handwriting datasets available, normalize them in some consistent way (zero is black, similar resolution, etc.) and then train a multi-task network on them. We have access to quite a few such datasets already on Kaggle\n\n1. [EMNIST](https://www.kaggle.com/crawford/emnist)\n2. [Kannada MNIST](https://www.kaggle.com/higgstachyon/kannada-mnist)\n3. [Kuzushiji-MNIST](https://www.kaggle.com/anokas/kuzushiji)\n4. [Handwritten math expression](https://www.kaggle.com/rtatman/handwritten-mathematical-expressions)\n5. [Handwritten Russian Letters](https://www.kaggle.com/olgabelitskaya/classification-of-handwritten-letters)\n6. [Numta DB (Bengali Digits)](https://www.kaggle.com/BengaliAI/numta/kernels)\n\nUsing all of these datasets to train a model that captures all of the knowledge of handwritting \n and other languages it may be possible to make an even stronger generalizing model. All of these datasets could be combined and tasks trained in tandem a la [Multi-Task Deep Neural Networks for Natural Language Understanding](https://arxiv.org/abs/1901.11504)\n\nThis could potentially be done as an initialization step to use instead of an imagenet pretrained or uninitialized CNN or done in a fashion similar to the paper linked above which was successful for language tasks.", "votes": null }, { "id": "767812", "postDate": "03/10/2020 05:46:25", "content": "Thanks for the useful list. Would like to add one more though its not for the exact same purpose
\n\n\n\nThis has few examples from a lot of handwritten languages, its u used a lot for few-shot learning literature.
", "rawMarkdown": "Thanks for the useful list. Would like to add one more though its not for the exact same purpose\n\n[Omniglot](https://github.com/brendenlake/omniglot)\n\nThis has few examples from a lot of handwritten languages, its u used a lot for few-shot learning literature.", "votes": null } ], "comments": [ { "id": 767812, "author_name": "dipamc77", "author_url": "", "post_date": "03/10/2020 05:46:25", "content": "Thanks for the useful list. Would like to add one more though its not for the exact same purpose
\n\n\n\nThis has few examples from a lot of handwritten languages, its u used a lot for few-shot learning literature.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "767762": "I haven't been able to work on this competition at all but figured I'd at least share an idea I had a while ago. My idea is to utilize all of the many handwriting datasets available, normalize them in some consistent way (zero is black, similar resolution, etc.) and then train a multi-task network on them. We have access to quite a few such datasets already on Kaggle\n\n1. [EMNIST](https://www.kaggle.com/crawford/emnist)\n2. [Kannada MNIST](https://www.kaggle.com/higgstachyon/kannada-mnist)\n3. [Kuzushiji-MNIST](https://www.kaggle.com/anokas/kuzushiji)\n4. [Handwritten math expression](https://www.kaggle.com/rtatman/handwritten-mathematical-expressions)\n5. [Handwritten Russian Letters](https://www.kaggle.com/olgabelitskaya/classification-of-handwritten-letters)\n6. [Numta DB (Bengali Digits)](https://www.kaggle.com/BengaliAI/numta/kernels)\n\nUsing all of these datasets to train a model that captures all of the knowledge of handwritting \n and other languages it may be possible to make an even stronger generalizing model. All of these datasets could be combined and tasks trained in tandem a la [Multi-Task Deep Neural Networks for Natural Language Understanding](https://arxiv.org/abs/1901.11504)\n\nThis could potentially be done as an initialization step to use instead of an imagenet pretrained or uninitialized CNN or done in a fashion similar to the paper linked above which was successful for language tasks.", "767812": "Thanks for the useful list. Would like to add one more though its not for the exact same purpose\n\n[Omniglot](https://github.com/brendenlake/omniglot)\n\nThis has few examples from a lot of handwritten languages, its u used a lot for few-shot learning literature." }, "source": "meta" }