{
  "id": 667060,
  "title": "Loss discussion",
  "url": "/competitions/csiro-biomass/discussion/667060",
  "author_name": "Ali Ozan Memetoglu",
  "post_date": "2026-01-10T22:20:07.216000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello, I’d appreciate your advice on which loss function to use for a model that predicts three targets clover, green, and dead using separate regression heads.</p>\n<p>At the moment, I’m seeing a consistent trade-off across epochs: when the green and dead predictions improve, clover tends to get worse, and when clover improves, the other two often degrade.</p>\n<p>Right now I’m using Smooth L1 (Huber) loss with different beta values per target: 2 for green, 5 for dead, and 7 for clover. This came out best in a quick ablation study and seems reasonable given the target distributions (clover has the longest tail, followed by dead, then green). Even so, training remains quite unstable and difficult.</p>\n<p>I’d be grateful for any suggestions or ideas you might have.</p>",
  "messages": [
    {
      "id": 3389313,
      "postDate": "2026-01-10T22:20:07.217Z",
      "content": "<p>Hello, I’d appreciate your advice on which loss function to use for a model that predicts three targets clover, green, and dead using separate regression heads.</p>\n<p>At the moment, I’m seeing a consistent trade-off across epochs: when the green and dead predictions improve, clover tends to get worse, and when clover improves, the other two often degrade.</p>\n<p>Right now I’m using Smooth L1 (Huber) loss with different beta values per target: 2 for green, 5 for dead, and 7 for clover. This came out best in a quick ablation study and seems reasonable given the target distributions (clover has the longest tail, followed by dead, then green). Even so, training remains quite unstable and difficult.</p>\n<p>I’d be grateful for any suggestions or ideas you might have.</p>",
      "rawMarkdown": "Hello, I’d appreciate your advice on which loss function to use for a model that predicts three targets clover, green, and dead using separate regression heads.\n\nAt the moment, I’m seeing a consistent trade-off across epochs: when the green and dead predictions improve, clover tends to get worse, and when clover improves, the other two often degrade.\n\nRight now I’m using Smooth L1 (Huber) loss with different beta values per target: 2 for green, 5 for dead, and 7 for clover. This came out best in a quick ablation study and seems reasonable given the target distributions (clover has the longest tail, followed by dead, then green). Even so, training remains quite unstable and difficult.\n\nI’d be grateful for any suggestions or ideas you might have.",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3389313": "Hello, I’d appreciate your advice on which loss function to use for a model that predicts three targets clover, green, and dead using separate regression heads.\n\nAt the moment, I’m seeing a consistent trade-off across epochs: when the green and dead predictions improve, clover tends to get worse, and when clover improves, the other two often degrade.\n\nRight now I’m using Smooth L1 (Huber) loss with different beta values per target: 2 for green, 5 for dead, and 7 for clover. This came out best in a quick ablation study and seems reasonable given the target distributions (clover has the longest tail, followed by dead, then green). Even so, training remains quite unstable and difficult.\n\nI’d be grateful for any suggestions or ideas you might have."
  }
}