{
  "id": 542331,
  "title": "Language model is all you need?",
  "url": "/competitions/applications-of-deep-learning-wustl-fall-2024/discussion/542331",
  "author_name": "",
  "post_date": "2024-10-24T08:13:51.133100900Z",
  "votes": 4,
  "comment_count": 11,
  "views": 0,
  "content": "<p>So far, CNN-based models is almost trivial in their performance compared to language based methods for me, even on the image data. It kind of makes sense since CNN cannot capture the atmosphere the generated images is trying to convey. We will see how it goes…</p>",
  "messages": [
    {
      "id": "3026778",
      "postDate": "10/24/2024 08:13:51",
      "content": "<p>So far, CNN-based models is almost trivial in their performance compared to language based methods for me, even on the image data. It kind of makes sense since CNN cannot capture the atmosphere the generated images is trying to convey. We will see how it goes…</p>",
      "rawMarkdown": "So far, CNN-based models is almost trivial in their performance compared to language based methods for me, even on the image data. It kind of makes sense since CNN cannot capture the atmosphere the generated images is trying to convey. We will see how it goes...",
      "votes": null
    },
    {
      "id": "3030817",
      "postDate": "10/28/2024 23:24:04",
      "content": "<p>You are on the right track, but remember each dataset has a unique set of information.The Kaggle powerpoint slides outlines multiple ways you can use the images to extract otherwise hidden features about the homes.</p>\n<p>If your goal is get the most accurate solution possible, you may benefit from using both LLM and CNN models, focus on integrating multiple methods and models.</p>",
      "rawMarkdown": "You are on the right track, but remember each dataset has a unique set of information.The Kaggle powerpoint slides outlines multiple ways you can use the images to extract otherwise hidden features about the homes.\n\nIf your goal is get the most accurate solution possible, you may benefit from using both LLM and CNN models, focus on integrating multiple methods and models.",
      "votes": null
    },
    {
      "id": "3033511",
      "postDate": "11/01/2024 08:27:46",
      "content": "<p>I have now made use of everything hinted in the powerpoint slides, but the improvement doesn't seem to be much. Probably next step will be to revisit CNN. </p>",
      "rawMarkdown": "I have now made use of everything hinted in the powerpoint slides, but the improvement doesn't seem to be much. Probably next step will be to revisit CNN.",
      "votes": null
    },
    {
      "id": "3034265",
      "postDate": "11/02/2024 00:08:15",
      "content": "<p>I just finished my first try based on LLM, and the result is actually far better than expected. I'm actually thinking of using other models rather than CNN, but I'm afraid that it's too much computing work for my free colab account. Anyway, please keep this discussion alive, I'm curious about your tries!! Thanks again!!</p>",
      "rawMarkdown": "I just finished my first try based on LLM, and the result is actually far better than expected. I'm actually thinking of using other models rather than CNN, but I'm afraid that it's too much computing work for my free colab account. Anyway, please keep this discussion alive, I'm curious about your tries!! Thanks again!!",
      "votes": null
    },
    {
      "id": "3035050",
      "postDate": "11/03/2024 00:00:23",
      "content": "<p>What methods are you thinking? So far my model is based on ensemble of Ilm and vision language. Tried 3 different CNN methods and none were working</p>",
      "rawMarkdown": "What methods are you thinking? So far my model is based on ensemble of Ilm and vision language. Tried 3 different CNN methods and none were working",
      "votes": null
    },
    {
      "id": "3035052",
      "postDate": "11/03/2024 00:06:27",
      "content": "<p>If your method is based on Ilm and no vision language, it is probably better than my llm model. My single llm model is around 160000</p>",
      "rawMarkdown": "If your method is based on Ilm and no vision language, it is probably better than my llm model. My single llm model is around 160000",
      "votes": null
    },
    {
      "id": "3035396",
      "postDate": "11/03/2024 11:59:29",
      "content": "<p>I'm also trying to emsemble llm and vision language, but somehow my ResNet50 returns only three-digits value… Still trying to figure out what's going wrong there… I'm not familiar with vision language</p>",
      "rawMarkdown": "I'm also trying to emsemble llm and vision language, but somehow my ResNet50 returns only three-digits value... Still trying to figure out what's going wrong there... I'm not familiar with vision language",
      "votes": null
    },
    {
      "id": "3035397",
      "postDate": "11/03/2024 12:00:08",
      "content": "<p>Yep, my mthod is based on a single llm model.</p>",
      "rawMarkdown": "Yep, my mthod is based on a single llm model.",
      "votes": null
    },
    {
      "id": "3037453",
      "postDate": "11/05/2024 18:12:44",
      "content": "<p>My resnet also does not work well. I think it's because CNN is generally for visual recognition</p>",
      "rawMarkdown": "My resnet also does not work well. I think it's because CNN is generally for visual recognition",
      "votes": null
    },
    {
      "id": "3041816",
      "postDate": "11/10/2024 22:41:46",
      "content": "<p>Thanks. I'll try again recently.</p>",
      "rawMarkdown": "Thanks. I'll try again recently.",
      "votes": null
    },
    {
      "id": "3055652",
      "postDate": "11/26/2024 01:08:32",
      "content": "<p>From the final touch it seems like image-language model and language model are indeed all you need. </p>",
      "rawMarkdown": "From the final touch it seems like image-language model and language model are indeed all you need.",
      "votes": null
    },
    {
      "id": "3055653",
      "postDate": "11/26/2024 01:09:10",
      "content": "<p>I will make that the title of presentation :)</p>",
      "rawMarkdown": "I will make that the title of presentation :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3030817,
      "author_name": "robertjirsaraie",
      "author_url": "",
      "post_date": "10/28/2024 23:24:04",
      "content": "<p>You are on the right track, but remember each dataset has a unique set of information.The Kaggle powerpoint slides outlines multiple ways you can use the images to extract otherwise hidden features about the homes.</p>\n<p>If your goal is get the most accurate solution possible, you may benefit from using both LLM and CNN models, focus on integrating multiple methods and models.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3033511,
          "author_name": "dantingzeng",
          "author_url": "",
          "post_date": "11/01/2024 08:27:46",
          "content": "<p>I have now made use of everything hinted in the powerpoint slides, but the improvement doesn't seem to be much. Probably next step will be to revisit CNN. </p>",
          "votes": null,
          "replies": [
            {
              "id": 3034265,
              "author_name": "kaiyisung",
              "author_url": "",
              "post_date": "11/02/2024 00:08:15",
              "content": "<p>I just finished my first try based on LLM, and the result is actually far better than expected. I'm actually thinking of using other models rather than CNN, but I'm afraid that it's too much computing work for my free colab account. Anyway, please keep this discussion alive, I'm curious about your tries!! Thanks again!!</p>",
              "votes": null,
              "replies": [
                {
                  "id": 3035050,
                  "author_name": "dantingzeng",
                  "author_url": "",
                  "post_date": "11/03/2024 00:00:23",
                  "content": "<p>What methods are you thinking? So far my model is based on ensemble of Ilm and vision language. Tried 3 different CNN methods and none were working</p>",
                  "votes": null,
                  "replies": [
                    {
                      "id": 3035396,
                      "author_name": "kaiyisung",
                      "author_url": "",
                      "post_date": "11/03/2024 11:59:29",
                      "content": "<p>I'm also trying to emsemble llm and vision language, but somehow my ResNet50 returns only three-digits value… Still trying to figure out what's going wrong there… I'm not familiar with vision language</p>",
                      "votes": null,
                      "replies": [
                        {
                          "id": 3037453,
                          "author_name": "dantingzeng",
                          "author_url": "",
                          "post_date": "11/05/2024 18:12:44",
                          "content": "<p>My resnet also does not work well. I think it's because CNN is generally for visual recognition</p>",
                          "votes": null,
                          "replies": [
                            {
                              "id": 3041816,
                              "author_name": "kaiyisung",
                              "author_url": "",
                              "post_date": "11/10/2024 22:41:46",
                              "content": "<p>Thanks. I'll try again recently.</p>",
                              "votes": null,
                              "replies": []
                            }
                          ]
                        }
                      ]
                    }
                  ]
                },
                {
                  "id": 3035052,
                  "author_name": "dantingzeng",
                  "author_url": "",
                  "post_date": "11/03/2024 00:06:27",
                  "content": "<p>If your method is based on Ilm and no vision language, it is probably better than my llm model. My single llm model is around 160000</p>",
                  "votes": null,
                  "replies": [
                    {
                      "id": 3035397,
                      "author_name": "kaiyisung",
                      "author_url": "",
                      "post_date": "11/03/2024 12:00:08",
                      "content": "<p>Yep, my mthod is based on a single llm model.</p>",
                      "votes": null,
                      "replies": []
                    }
                  ]
                }
              ]
            }
          ]
        }
      ]
    },
    {
      "id": 3055652,
      "author_name": "dantingzeng",
      "author_url": "",
      "post_date": "11/26/2024 01:08:32",
      "content": "<p>From the final touch it seems like image-language model and language model are indeed all you need. </p>",
      "votes": null,
      "replies": [
        {
          "id": 3055653,
          "author_name": "dantingzeng",
          "author_url": "",
          "post_date": "11/26/2024 01:09:10",
          "content": "<p>I will make that the title of presentation :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3026778": "So far, CNN-based models is almost trivial in their performance compared to language based methods for me, even on the image data. It kind of makes sense since CNN cannot capture the atmosphere the generated images is trying to convey. We will see how it goes...",
    "3030817": "You are on the right track, but remember each dataset has a unique set of information.The Kaggle powerpoint slides outlines multiple ways you can use the images to extract otherwise hidden features about the homes.\n\nIf your goal is get the most accurate solution possible, you may benefit from using both LLM and CNN models, focus on integrating multiple methods and models.",
    "3033511": "I have now made use of everything hinted in the powerpoint slides, but the improvement doesn't seem to be much. Probably next step will be to revisit CNN.",
    "3034265": "I just finished my first try based on LLM, and the result is actually far better than expected. I'm actually thinking of using other models rather than CNN, but I'm afraid that it's too much computing work for my free colab account. Anyway, please keep this discussion alive, I'm curious about your tries!! Thanks again!!",
    "3035050": "What methods are you thinking? So far my model is based on ensemble of Ilm and vision language. Tried 3 different CNN methods and none were working",
    "3035052": "If your method is based on Ilm and no vision language, it is probably better than my llm model. My single llm model is around 160000",
    "3035396": "I'm also trying to emsemble llm and vision language, but somehow my ResNet50 returns only three-digits value... Still trying to figure out what's going wrong there... I'm not familiar with vision language",
    "3035397": "Yep, my mthod is based on a single llm model.",
    "3037453": "My resnet also does not work well. I think it's because CNN is generally for visual recognition",
    "3041816": "Thanks. I'll try again recently.",
    "3055652": "From the final touch it seems like image-language model and language model are indeed all you need.",
    "3055653": "I will make that the title of presentation :)"
  },
  "source": "meta"
}