{
  "id": 635197,
  "title": "Pretrain with the dataset of Brain-to-Text '24",
  "url": "/competitions/brain-to-text-25/discussion/635197",
  "author_name": "",
  "post_date": "2025-11-20T06:54:31.990820600Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I attempted to pretrain the RNN decoder using the Brain-to-Text '24 dataset and then fine-tuned it with the Brain-to-Text '25 dataset. The final result was slightly worse than that of the RNN trained only on the Brain-to-Text '25 dataset. Has anyone else attempted to pretrain with the Brain-to-Text '24 dataset? Were there any improvements in the results?</p>",
  "messages": [
    {
      "id": "3341425",
      "postDate": "11/20/2025 06:54:31",
      "content": "<p>I attempted to pretrain the RNN decoder using the Brain-to-Text '24 dataset and then fine-tuned it with the Brain-to-Text '25 dataset. The final result was slightly worse than that of the RNN trained only on the Brain-to-Text '25 dataset. Has anyone else attempted to pretrain with the Brain-to-Text '24 dataset? Were there any improvements in the results?</p>",
      "rawMarkdown": "I attempted to pretrain the RNN decoder using the Brain-to-Text '24 dataset and then fine-tuned it with the Brain-to-Text '25 dataset. The final result was slightly worse than that of the RNN trained only on the Brain-to-Text '25 dataset. Has anyone else attempted to pretrain with the Brain-to-Text '24 dataset? Were there any improvements in the results?",
      "votes": null
    },
    {
      "id": "3341571",
      "postDate": "11/20/2025 09:15:43",
      "content": "<p>Didn't tried that but am curious how do you account for </p>\n<ol>\n<li>electrode differences ( 128 vs 256), and </li>\n<li>subject differences ( T12 vs T15) ? </li>\n</ol>\n<p>If you can get a better model by pretraining on Brain-to-Text '24 dataset, that means your model learns a robust general activation pattern for phonemes across different patients, which is quite cool !</p>",
      "rawMarkdown": "Didn't tried that but am curious how do you account for \n1. electrode differences ( 128 vs 256), and \n2. subject differences ( T12 vs T15) ? \n\nIf you can get a better model by pretraining on Brain-to-Text '24 dataset, that means your model learns a robust general activation pattern for phonemes across different patients, which is quite cool !",
      "votes": null
    },
    {
      "id": "3341757",
      "postDate": "11/20/2025 12:07:12",
      "content": "<ol>\n<li>The raw data of the Brain-to-Text '24 dataset contains 5 features [-3.5,-4.5,-5.5,-6.5 RMS threshold crossings and spike band power] of each electrode. </li>\n<li>I haven't fully addressed the second issue. I attempted to mitigate the impact of subject differences by resetting the linear day-specific input layers during fine-tuning.</li>\n</ol>",
      "rawMarkdown": "1. The raw data of the Brain-to-Text '24 dataset contains 5 features [-3.5,-4.5,-5.5,-6.5 RMS threshold crossings and spike band power] of each electrode. \n2. I haven't fully addressed the second issue. I attempted to mitigate the impact of subject differences by resetting the linear day-specific input layers during fine-tuning.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3341571,
      "author_name": "heyyousum",
      "author_url": "",
      "post_date": "11/20/2025 09:15:43",
      "content": "<p>Didn't tried that but am curious how do you account for </p>\n<ol>\n<li>electrode differences ( 128 vs 256), and </li>\n<li>subject differences ( T12 vs T15) ? </li>\n</ol>\n<p>If you can get a better model by pretraining on Brain-to-Text '24 dataset, that means your model learns a robust general activation pattern for phonemes across different patients, which is quite cool !</p>",
      "votes": null,
      "replies": [
        {
          "id": 3341757,
          "author_name": "pandadapan",
          "author_url": "",
          "post_date": "11/20/2025 12:07:12",
          "content": "<ol>\n<li>The raw data of the Brain-to-Text '24 dataset contains 5 features [-3.5,-4.5,-5.5,-6.5 RMS threshold crossings and spike band power] of each electrode. </li>\n<li>I haven't fully addressed the second issue. I attempted to mitigate the impact of subject differences by resetting the linear day-specific input layers during fine-tuning.</li>\n</ol>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3341425": "I attempted to pretrain the RNN decoder using the Brain-to-Text '24 dataset and then fine-tuned it with the Brain-to-Text '25 dataset. The final result was slightly worse than that of the RNN trained only on the Brain-to-Text '25 dataset. Has anyone else attempted to pretrain with the Brain-to-Text '24 dataset? Were there any improvements in the results?",
    "3341571": "Didn't tried that but am curious how do you account for \n1. electrode differences ( 128 vs 256), and \n2. subject differences ( T12 vs T15) ? \n\nIf you can get a better model by pretraining on Brain-to-Text '24 dataset, that means your model learns a robust general activation pattern for phonemes across different patients, which is quite cool !",
    "3341757": "1. The raw data of the Brain-to-Text '24 dataset contains 5 features [-3.5,-4.5,-5.5,-6.5 RMS threshold crossings and spike band power] of each electrode. \n2. I haven't fully addressed the second issue. I attempted to mitigate the impact of subject differences by resetting the linear day-specific input layers during fine-tuning."
  },
  "source": "meta"
}