{"id":"6d8brVnfZb","url":"https://pastebin.ca/6d8brVnfZb","raw_url":"https://raw.anybin.ca/6d8brVnfZb","visibility":"public","access":"public","created_at":1793951539324,"expires_at":1794556339324,"fetch_limit":null,"fetches_used":0,"reads_remaining":null,"size_bytes":796,"syntax_hint":null,"title":null,"filename":null,"change_note":null,"cipher":null,"cipher_meta":null,"parent_id":null,"root_id":"6d8brVnfZb","version":1,"owner_id":null,"recipient_id":null,"body":"如何识别过拟合现象？\n观察损失曲线：训练集损失持续下降，但是测试集（验证集）损失先下降，达到最低点之后开始上升；同时模型在训练集上准确率很高，在测试集上效果明显变差，就说明出现了过拟合。\n为了避免过拟合，训练应该在哪个Epoch停止？\n在测试（验证）损失达到最小值对应的Epoch停止训练，也就是测试损失即将由降转升的临界点。这就是早停法的核心依据。\n除了早停法，还有哪些避免过拟合的方法？\n正则化：L1、L2权重正则化\nDropout：训练时随机让一部分神经元失活\n增加训练数据、数据增强\n简化网络结构，减少网络参数\n加入噪声、批量归一化（BatchNorm）\n如果你想要更简短精简"}