{
  "id": 665515,
  "title": "Transformer-based Decoding with 1D-Upsampling ",
  "url": "/competitions/brain-to-text-25/discussion/665515",
  "author_name": "Skala Hamasaed Fatah",
  "post_date": "2026-01-01T23:46:07.556000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n<p>I wanted to share my approach for this competition, which focuses on achieving robust decoding using a \"<strong>Transformer architecture</strong>\" combined with \"<strong>1D-Upsampling</strong>\".</p>\n<p>In my solution, I implemented a \"<strong>4-layer Transformer Encoder</strong>\" with \"<strong>8-head Multi-Head Attention</strong>\". To bridge the gap between high-dimensional neural signals and character sequences, I utilized a \"<strong>ConvTranspose1d</strong>\" layer for \"<strong>temporal upsampling</strong>\". This ensured that the output sequence length was optimal for \"<strong>CTC Loss</strong>\" decoding.</p>\n<p>Key Highlights of my submission:</p>\n<p>Innovation: Used \"1<strong>D-Upsampling</strong>\" to maintain \"<strong>spatial-temporal alignment</strong>\".</p>\n<p>Robustness &amp; Stability: My model showed significant stability on unseen data. <strong>My rank improved by 33 places</strong>  (from <strong>389</strong> on the <strong>Public Leaderboard</strong> to <strong>356</strong> on the \"<strong>Private Leaderboard</strong>\") with a score of <strong>0.9</strong> . This improvement highlights the model's ability to generalize across different neural recording sessions.</p>\n<p>Efficiency: Optimized for stable gradient flow using \"<strong>Z-score normalization</strong>\" and \"<strong>AdamW</strong>\".\nI have made my notebook \"<strong>Public</strong>\" for the community and the organizers to review. I would love to hear feedback from the \"<strong>Blackrock Neurotech</strong>\" team and the organizers!</p>\n<p>Link to Notebook: [<a href=\"https://www.kaggle.com/code/skalahamasaedfatah/brain-to-text-skala\" target=\"_blank\">https://www.kaggle.com/code/skalahamasaedfatah/brain-to-text-skala</a>]</p>\n<p>Best regards, Skala Hamasaed Fatah</p>",
  "messages": [
    {
      "id": 3384779,
      "postDate": "2026-01-01T23:46:07.557Z",
      "content": "<p>Hi everyone,</p>\n<p>I wanted to share my approach for this competition, which focuses on achieving robust decoding using a \"<strong>Transformer architecture</strong>\" combined with \"<strong>1D-Upsampling</strong>\".</p>\n<p>In my solution, I implemented a \"<strong>4-layer Transformer Encoder</strong>\" with \"<strong>8-head Multi-Head Attention</strong>\". To bridge the gap between high-dimensional neural signals and character sequences, I utilized a \"<strong>ConvTranspose1d</strong>\" layer for \"<strong>temporal upsampling</strong>\". This ensured that the output sequence length was optimal for \"<strong>CTC Loss</strong>\" decoding.</p>\n<p>Key Highlights of my submission:</p>\n<p>Innovation: Used \"1<strong>D-Upsampling</strong>\" to maintain \"<strong>spatial-temporal alignment</strong>\".</p>\n<p>Robustness &amp; Stability: My model showed significant stability on unseen data. <strong>My rank improved by 33 places</strong>  (from <strong>389</strong> on the <strong>Public Leaderboard</strong> to <strong>356</strong> on the \"<strong>Private Leaderboard</strong>\") with a score of <strong>0.9</strong> . This improvement highlights the model's ability to generalize across different neural recording sessions.</p>\n<p>Efficiency: Optimized for stable gradient flow using \"<strong>Z-score normalization</strong>\" and \"<strong>AdamW</strong>\".\nI have made my notebook \"<strong>Public</strong>\" for the community and the organizers to review. I would love to hear feedback from the \"<strong>Blackrock Neurotech</strong>\" team and the organizers!</p>\n<p>Link to Notebook: [<a href=\"https://www.kaggle.com/code/skalahamasaedfatah/brain-to-text-skala\" target=\"_blank\">https://www.kaggle.com/code/skalahamasaedfatah/brain-to-text-skala</a>]</p>\n<p>Best regards, Skala Hamasaed Fatah</p>",
      "rawMarkdown": "Hi everyone,\n\nI wanted to share my approach for this competition, which focuses on achieving robust decoding using a \"**Transformer architecture**\" combined with \"**1D-Upsampling**\".\n\nIn my solution, I implemented a \"**4-layer Transformer Encoder**\" with \"**8-head Multi-Head Attention**\". To bridge the gap between high-dimensional neural signals and character sequences, I utilized a \"**ConvTranspose1d**\" layer for \"**temporal upsampling**\". This ensured that the output sequence length was optimal for \"**CTC Loss**\" decoding.\n\nKey Highlights of my submission:\n\nInnovation: Used \"1**D-Upsampling**\" to maintain \"**spatial-temporal alignment**\".\n\nRobustness & Stability: My model showed significant stability on unseen data. **My rank improved by 33 places**  (from **389** on the **Public Leaderboard** to **356** on the \"**Private Leaderboard**\") with a score of **0.9** . This improvement highlights the model's ability to generalize across different neural recording sessions.\n\nEfficiency: Optimized for stable gradient flow using \"**Z-score normalization**\" and \"**AdamW**\".\nI have made my notebook \"**Public**\" for the community and the organizers to review. I would love to hear feedback from the \"**Blackrock Neurotech**\" team and the organizers!\n\nLink to Notebook: [https://www.kaggle.com/code/skalahamasaedfatah/brain-to-text-skala]\n\nBest regards, Skala Hamasaed Fatah"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3384779": "Hi everyone,\n\nI wanted to share my approach for this competition, which focuses on achieving robust decoding using a \"**Transformer architecture**\" combined with \"**1D-Upsampling**\".\n\nIn my solution, I implemented a \"**4-layer Transformer Encoder**\" with \"**8-head Multi-Head Attention**\". To bridge the gap between high-dimensional neural signals and character sequences, I utilized a \"**ConvTranspose1d**\" layer for \"**temporal upsampling**\". This ensured that the output sequence length was optimal for \"**CTC Loss**\" decoding.\n\nKey Highlights of my submission:\n\nInnovation: Used \"1**D-Upsampling**\" to maintain \"**spatial-temporal alignment**\".\n\nRobustness & Stability: My model showed significant stability on unseen data. **My rank improved by 33 places**  (from **389** on the **Public Leaderboard** to **356** on the \"**Private Leaderboard**\") with a score of **0.9** . This improvement highlights the model's ability to generalize across different neural recording sessions.\n\nEfficiency: Optimized for stable gradient flow using \"**Z-score normalization**\" and \"**AdamW**\".\nI have made my notebook \"**Public**\" for the community and the organizers to review. I would love to hear feedback from the \"**Blackrock Neurotech**\" team and the organizers!\n\nLink to Notebook: [https://www.kaggle.com/code/skalahamasaedfatah/brain-to-text-skala]\n\nBest regards, Skala Hamasaed Fatah"
  }
}