{
  "id": 579020,
  "title": "Why would aux loss hurt the performance?",
  "url": "/competitions/birdclef-2025/discussion/579020",
  "author_name": "",
  "post_date": "2025-05-14T17:15:42.345691600Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Anyone tried aux loss for the groups (Amphibians, Insects, Aves, Mammals)? <br>\nFor me it hurts the performance a lot but what would that mean?</p>\n<p>My assumption is, without aux loss I might be overfitting to the LB because the features learned by my model are not discriminative compared to what it should, but when enforced using aux loss it starts learning what it should, but the performance is not good enough.</p>\n<p>Any thoughts on this?</p>",
  "messages": [
    {
      "id": "3201984",
      "postDate": "05/14/2025 17:15:42",
      "content": "<p>Anyone tried aux loss for the groups (Amphibians, Insects, Aves, Mammals)? <br>\nFor me it hurts the performance a lot but what would that mean?</p>\n<p>My assumption is, without aux loss I might be overfitting to the LB because the features learned by my model are not discriminative compared to what it should, but when enforced using aux loss it starts learning what it should, but the performance is not good enough.</p>\n<p>Any thoughts on this?</p>",
      "rawMarkdown": "Anyone tried aux loss for the groups (Amphibians, Insects, Aves, Mammals)? \nFor me it hurts the performance a lot but what would that mean?\n\nMy assumption is, without aux loss I might be overfitting to the LB because the features learned by my model are not discriminative compared to what it should, but when enforced using aux loss it starts learning what it should, but the performance is not good enough.\n\n Any thoughts on this?",
      "votes": null
    },
    {
      "id": "3202022",
      "postDate": "05/14/2025 18:11:50",
      "content": "<p>I tried PairLogit loss to directly optimise ROC AUC in different setups both for pure classes and for class groups<br>\nIt dropped LB score for me</p>",
      "rawMarkdown": "I tried PairLogit loss to directly optimise ROC AUC in different setups both for pure classes and for class groups\nIt dropped LB score for me",
      "votes": null
    },
    {
      "id": "3202121",
      "postDate": "05/14/2025 21:14:35",
      "content": "<p>I think that a severe imbalance of samples between groups, different recording domains and the fact that some groups are not easily distinguishable from each other cause the model to overfit to irrelevant features when using the auxiliary loss.</p>",
      "rawMarkdown": "I think that a severe imbalance of samples between groups, different recording domains and the fact that some groups are not easily distinguishable from each other cause the model to overfit to irrelevant features when using the auxiliary loss.",
      "votes": null
    },
    {
      "id": "3202347",
      "postDate": "05/15/2025 08:53:25",
      "content": "<p>Thank you. Yes, I just listened to the recordings, In most of the insect group recordings you can sometimes hear birds in the background. </p>",
      "rawMarkdown": "Thank you. Yes, I just listened to the recordings, In most of the insect group recordings you can sometimes hear birds in the background.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3202022,
      "author_name": "veshkinartem",
      "author_url": "",
      "post_date": "05/14/2025 18:11:50",
      "content": "<p>I tried PairLogit loss to directly optimise ROC AUC in different setups both for pure classes and for class groups<br>\nIt dropped LB score for me</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3202121,
      "author_name": "nikitababich",
      "author_url": "",
      "post_date": "05/14/2025 21:14:35",
      "content": "<p>I think that a severe imbalance of samples between groups, different recording domains and the fact that some groups are not easily distinguishable from each other cause the model to overfit to irrelevant features when using the auxiliary loss.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3202347,
          "author_name": "salmanahmedtamu",
          "author_url": "",
          "post_date": "05/15/2025 08:53:25",
          "content": "<p>Thank you. Yes, I just listened to the recordings, In most of the insect group recordings you can sometimes hear birds in the background. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3201984": "Anyone tried aux loss for the groups (Amphibians, Insects, Aves, Mammals)? \nFor me it hurts the performance a lot but what would that mean?\n\nMy assumption is, without aux loss I might be overfitting to the LB because the features learned by my model are not discriminative compared to what it should, but when enforced using aux loss it starts learning what it should, but the performance is not good enough.\n\n Any thoughts on this?",
    "3202022": "I tried PairLogit loss to directly optimise ROC AUC in different setups both for pure classes and for class groups\nIt dropped LB score for me",
    "3202121": "I think that a severe imbalance of samples between groups, different recording domains and the fact that some groups are not easily distinguishable from each other cause the model to overfit to irrelevant features when using the auxiliary loss.",
    "3202347": "Thank you. Yes, I just listened to the recordings, In most of the insect group recordings you can sometimes hear birds in the background."
  },
  "source": "meta"
}