{
  "id": 614169,
  "title": "About TTA (Test-Time Adaptation)",
  "url": "/competitions/brain-to-text-25/discussion/614169",
  "author_name": "",
  "post_date": "2025-11-01T17:33:43.424421700Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I saw that in the leaderboard, \n\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" got test PER = 0.04424 , while \n\"Stanford-NPTL causal RNN Ensemble + 5gram\" got test PER = 0.03090,</p>\n<p>I want to know the difference bewteen the 2 baseline model, are the differences solely on using TTA or not?\nIf that is the case, does that means applying TTA with the baseline model doesn't improve the score? </p>",
  "messages": [
    {
      "id": "3309906",
      "postDate": "11/01/2025 17:33:43",
      "content": "<p>I saw that in the leaderboard, \n\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" got test PER = 0.04424 , while \n\"Stanford-NPTL causal RNN Ensemble + 5gram\" got test PER = 0.03090,</p>\n<p>I want to know the difference bewteen the 2 baseline model, are the differences solely on using TTA or not?\nIf that is the case, does that means applying TTA with the baseline model doesn't improve the score? </p>",
      "rawMarkdown": "I saw that in the leaderboard, \n\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" got test PER = 0.04424 , while \n\"Stanford-NPTL causal RNN Ensemble + 5gram\" got test PER = 0.03090,\n\nI want to know the difference bewteen the 2 baseline model, are the differences solely on using TTA or not?\nIf that is the case, does that means applying TTA with the baseline model doesn't improve the score?",
      "votes": null
    },
    {
      "id": "3312025",
      "postDate": "11/06/2025 08:33:00",
      "content": "<p>I would like to know this too, <a href=\"https://www.kaggle.com/notnickc\" target=\"_blank\">@notnickc</a> </p>",
      "rawMarkdown": "I would like to know this too, @notnickc",
      "votes": null
    },
    {
      "id": "3312706",
      "postDate": "11/07/2025 18:45:53",
      "content": "<p>\"Stanford-NPTL causal RNN Ensemble + 5gram\" uses 10 RNN models to make phoneme sequence predictions for each val/test trial, then uses a joiner to make a final prediction form those</p>\n<p>\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" is similar, but instead of using 10 RNN models, it uses 1, and each val/test trial is copies 10 times and augmented with white noise 10 unique ways before being passed through that model to generate slightly different phoneme sequence predictions.</p>\n<p>See <a href=\"https://arxiv.org/abs/2412.17227\" target=\"_blank\">https://arxiv.org/abs/2412.17227</a> for general details about ensembling &amp; the joiner.</p>",
      "rawMarkdown": "\"Stanford-NPTL causal RNN Ensemble + 5gram\" uses 10 RNN models to make phoneme sequence predictions for each val/test trial, then uses a joiner to make a final prediction form those\n\n\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" is similar, but instead of using 10 RNN models, it uses 1, and each val/test trial is copies 10 times and augmented with white noise 10 unique ways before being passed through that model to generate slightly different phoneme sequence predictions.\n\nSee https://arxiv.org/abs/2412.17227 for general details about ensembling & the joiner.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3312025,
      "author_name": "heyyousum",
      "author_url": "",
      "post_date": "11/06/2025 08:33:00",
      "content": "<p>I would like to know this too, <a href=\"https://www.kaggle.com/notnickc\" target=\"_blank\">@notnickc</a> </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3312706,
      "author_name": "notnickc",
      "author_url": "",
      "post_date": "11/07/2025 18:45:53",
      "content": "<p>\"Stanford-NPTL causal RNN Ensemble + 5gram\" uses 10 RNN models to make phoneme sequence predictions for each val/test trial, then uses a joiner to make a final prediction form those</p>\n<p>\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" is similar, but instead of using 10 RNN models, it uses 1, and each val/test trial is copies 10 times and augmented with white noise 10 unique ways before being passed through that model to generate slightly different phoneme sequence predictions.</p>\n<p>See <a href=\"https://arxiv.org/abs/2412.17227\" target=\"_blank\">https://arxiv.org/abs/2412.17227</a> for general details about ensembling &amp; the joiner.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3309906": "I saw that in the leaderboard, \n\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" got test PER = 0.04424 , while \n\"Stanford-NPTL causal RNN Ensemble + 5gram\" got test PER = 0.03090,\n\nI want to know the difference bewteen the 2 baseline model, are the differences solely on using TTA or not?\nIf that is the case, does that means applying TTA with the baseline model doesn't improve the score?",
    "3312025": "I would like to know this too, @notnickc",
    "3312706": "\"Stanford-NPTL causal RNN Ensemble + 5gram\" uses 10 RNN models to make phoneme sequence predictions for each val/test trial, then uses a joiner to make a final prediction form those\n\n\"Stanford-NPTL causal RNN TTA-Ensemble + 5gram\" is similar, but instead of using 10 RNN models, it uses 1, and each val/test trial is copies 10 times and augmented with white noise 10 unique ways before being passed through that model to generate slightly different phoneme sequence predictions.\n\nSee https://arxiv.org/abs/2412.17227 for general details about ensembling & the joiner."
  },
  "source": "meta"
}