英文:
Can a model learn from validation set?
问题
模型可以从验证集和测试集中学习。
英文:
I have just joined in the AI concept and started to learn machine learning, I saw that we split the dataset into train, valid and testing set. I have read that valid set can be used to avoid overfitting and underfitting but I don't actually know why so maybe our model can learn from valid set but can't learn from testing set isn't it?
The concise answer about model can learn from valid and testing set.
答案1
得分: 0
以下是您要翻译的内容：
模型既不从验证集中学习，也不从测试集中学习。
- 您使用验证集来验证与您选择的算法和超参数相关的训练结果。模型不会以任何方式对实际训练数据进行训练或以其进行训练。
- 您在完成调优和训练模型后使用测试集来获得对模型质量没有影响的观点。同样，模型不以任何方式对测试集进行训练。
为什么我们不只是使用测试集来调优和最终评估模型？
作为数据科学家，我们要调整超参数。因此，验证数据间接影响我们的模型，因为它直接影响我们的设计决策。我们通过微调模型使其与验证数据良好匹配，这可能会引入倾斜。只有通过使用数据的第三部分 - 测试集 - 我们才能获得对模型最终质量的完全中立观点。
英文:
The model does learn from neither the validation set nor the test set.
- You use the validation set to validate results from training in regard to the algorithm and hyperparameters you choose. The model does not «see» the data for actual training or is in any way trained on it.
- You use the test set after you finish tuning and training the model to get an unaffected view of you models quality. Again the model is not trained in any way on the test set.
Why don't we just use the test set for tuning and final assesment of the model?
We as data scientists tune hyperparameters. Therefore the validation data indirectly influences our model because it directly influences our design decisions. We nudge our model to work well with the validation data and that can possibly bring in a tilt. Only by using a third part of the data – the test set - we get this entirely neutral view on the final quality of the model.
通过集体智慧和协作来改善编程学习和解决问题的方式。致力于成为全球开发者共同参与的知识库，让每个人都能够通过互相帮助和分享经验来进步。
评论