基于切削力的刀刃锋利度识别模型：损失下降但精度停滞问题

我正在构建一个神经网络模型，用于根据切口后特定距离的切削力识别刀刃锋利度。数据为CSV格式，采用含2个隐藏层的二分类模型，仅拥有45个输入数据点。运行模型时发现，损失持续下降，但精度在多个epoch后才发生变化。

#Initialising the neural network
Classifier = Sequential()

#Adding the input layer and the first hidden layer
Classifier.add(Dense(units=2, kernel_initializer= 'he_uniform', activation= 'relu',input_dim = 2))

Classifier.add(Dense(units=2, kernel_initializer= 'he_uniform', activation= 'relu',))


#Adding the output layer
Classifier.add(Dense(units =1, kernel_initializer='glorot_uniform', activation = 'sigmoid',))


Classifier.summary()

Epoch 177/2000
1/1 [==============================] - 0s 98ms/step - loss: 0.5921 - accuracy: 0.7222 - val_loss: 0.6642 - val_accuracy: 0.5000
Epoch 178/2000
1/1 [==============================] - 0s 72ms/step - loss: 0.5915 - accuracy: 0.7222 - val_loss: 0.6627 - val_accuracy: 0.5000
Epoch 179/2000
1/1 [==============================] - 0s 83ms/step - loss: 0.5908 - accuracy: 0.7222 - val_loss: 0.6612 - val_accuracy: 0.5000
Epoch 180/2000
1/1 [==============================] - 0s 82ms/step - loss: 0.5902 - accuracy: 0.7222 - val_loss: 0.6597 - val_accuracy: 0.5000
Epoch 181/2000
1/1 [==============================] - 0s 123ms/step - loss: 0.5896 - accuracy: 0.7222 - val_loss: 0.6581 - val_accuracy: 0.5000
Epoch 182/2000
1/1 [==============================] - 0s 77ms/step - loss: 0.5889 - accuracy: 0.7222 - val_loss: 0.6566 - val_accuracy: 0.5000
Epoch 183/2000
1/1 [==============================] - 0s 75ms/step - loss: 0.5883 - accuracy: 0.7500 - val_loss: 0.6550 - val_accuracy: 0.5000
Epoch 184/2000
1/1 [==============================] - 0s 73ms/step - loss: 0.5877 - accuracy: 0.8056 - val_loss: 0.6533 - val_accuracy: 0.5000
Epoch 185/2000
1/1 [==============================] - 0s 83ms/step - loss: 0.5870 - accuracy: 0.8056 - val_loss: 0.6517 - val_accuracy: 0.5000
Epoch 186/2000
1/1 [==============================] - 0s 103ms/step - loss: 0.5864 - accuracy: 0.8056 - val_loss: 0.6500 - val_accuracy: 0.5000
Epoch 187/2000
1/1 [==============================] - 0s 95ms/step - loss: 0.5857 - accuracy: 0.8056 - val_loss: 0.6484 - val_accuracy: 0.5000
Epoch 188/2000
1/1 [==============================] - 0s 69ms/step - loss: 0.5851 - accuracy: 0.8056 - val_loss: 0.6467 - val_accuracy: 0.5000
Epoch 189/2000
1/1 [==============================] - 0s 84ms/step - loss: 0.5845 - accuracy: 0.8056 - val_loss: 0.6450 - val_accuracy: 0.5000
Epoch 190/2000
1/1 [==============================] - 0s 94ms/step - loss: 0.5838 - accuracy: 0.8056 - val_loss: 0.6433 - val_accuracy: 0.5000
Epoch 191/2000
1/1 [==============================] - 0s 86ms/step - loss: 0.5832 - accuracy: 0.8056 - val_loss: 0.6416 - val_accuracy: 0.5000
Epoch 192/2000
1/1 [==============================] - 0s 80ms/step - loss: 0.5825 - accuracy: 0.8056 - val_loss: 0.6399 - val_accuracy: 0.5000
Epoch 193/2000
1/1 [==============================] - 0s 63ms/step - loss: 0.5818 - accuracy: 0.8056 - val_loss: 0.6381 - val_accuracy: 0.5000
Epoch 194/2000
1/1 [==============================] - 0s 79ms/step - loss: 0.5812 - accuracy: 0.8056 - val_loss: 0.6364 - val_accuracy: 0.5000
Epoch 195/2000
1/1 [==============================] - 0s 87ms/step - loss: 0.5805 - accuracy: 0.8056 - val_loss: 0.6347 - val_accuracy: 0.5000
Epoch 196/2000
1/1 [==============================] - 0s 90ms/step - loss: 0.5799 - accuracy: 0.8056 - val_loss: 0.6330 - val_accuracy: 0.5000
Epoch 197/2000
1/1 [==============================] - 0s 83ms/step - loss: 0.5792 - accuracy: 0.8056 - val_loss: 0.6313 - val_accuracy: 0.7500
Epoch 198/2000
1/1 [==============================] - 0s 191ms/step - loss: 0.5785 - accuracy: 0.8333 - val_loss: 0.6296 - val_accuracy: 1.0000
Epoch 199/2000
1/1 [==============================] - 0s 77ms/step - loss: 0.5779 - accuracy: 0.8333 - val_loss: 0.6278 - val_accuracy: 1.0000
Epoch 200/2000
1/1 [==============================] - 0s 122ms/step - loss: 0.5772 - accuracy: 0.8333 - val_loss: 0.6261 - val_accuracy: 1.0000
Epoch 201/2000
1/1 [==============================] - 0s 98ms/step - loss: 0.5765 - accuracy: 0.8333 - val_loss: 0.6244 - val_accuracy: 1.0000
Epoch 202/2000
1/1 [==============================] - 0s 85ms/step - loss: 0.5758 - accuracy: 0.8333 - val_loss: 0.6226 - val_accuracy: 1.0000
Epoch 203/2000
1/1 [==============================] - 0s 107ms/step - loss: 0.5752 - accuracy: 0.8333 - val_loss: 0.6209 - val_accuracy: 1.0000
Epoch 204/2000
1/1 [==============================] - 0s 54ms/step - loss: 0.5745 - accuracy: 0.8333 - val_loss: 0.6192 - val_accuracy: 1.0000
Epoch 205/2000
1/1 [==============================] - 0s 67ms/step - loss: 0.5738 - accuracy: 0.8333 - val_loss: 0.6175 - val_accuracy: 1.0000
Epoch 206/2000
1/1 [==============================] - 0s 125ms/step - loss: 0.5731 - accuracy: 0.8333 - val_loss: 0.6158 - val_accuracy: 1.0000
Epoch 207/2000
1/1 [==============================] - 0s 101ms/step - loss: 0.5725 - accuracy: 0.8333 - val_loss: 0.6140 - val_accuracy: 1.0000
Epoch 208/2000
1/1 [==============================] - 0s 146ms/step - loss: 0.5718 - accuracy: 0.8333 - val_loss: 0.6123 - val_accuracy: 1.0000
Epoch 209/2000
1/1 [==============================] - 0s 218ms/step - loss: 0.5711 - accuracy: 0.8333 - val_loss: 0.6106 - val_accuracy: 1.0000
Epoch 210/2000
1/1 [==============================] - 0s 174ms/step - loss: 0.5704 - accuracy: 0.8333 - val_loss: 0.6088 - val_accuracy: 1.0000

• 精度延迟更新的本质：损失是连续计算的数值，会随着权重微调持续下降；但精度是离散的分类指标，只有当模型预测概率跨过0.5的分类阈值时，对应样本的判断才会从错误转为正确，精度才会跳变。前期权重调整还没让足够多的样本越过阈值，所以精度保持不变。
• 小数据集的影响：仅45个数据点，模型容易卡在局部最优，需要更多迭代才能把权重调整到合适状态，让预测概率跨过阈值。
• 模型结构调整：
  • 当前隐藏层仅2个神经元，拟合能力不足，可增加到4-8个神经元，提升模型对数据特征的捕捉能力。
  • 添加Dropout(0.2)层，防止小数据集带来的过拟合问题。
• 训练策略优化：
  • 换用Adam自适应优化器，替代默认SGD，加快权重更新速度，让模型更快收敛。
  • 对输入数据做标准化/归一化处理，让梯度更新更稳定，加速模型调整。
  • 添加EarlyStopping回调，当验证精度连续多轮无提升时自动停止训练，避免无效迭代。

内容的提问来源于stack exchange，提问作者Jiten Prasad