| Classification | Regression | |
|---|---|---|
| Output | target variables are discrete | Continuous numerical value |
| Goal | Predict category | Predict exact numerical value |
| Evaluation | Precision, Recall, F1-Score | MSE, R2-Score, MAPE, RMSE |
| Decision boundary | Clearly defined between different classed | No distinct boundaries |
| Algorithm | Logistic regression, Decision trees, Support Vector Machines (SVM) | Linear Regression, Polynomial Regression, Decision Trees (with regression objective). |
|---|
MNIST는 일반적으로 classification 문제로 접근하지만, 본 실험에서는 regression 방식으로도 충분히 높은 성능을 낼 수 있는지 검증해보고자 하였음.
Regression은 label 간 거리 정보를 활용할 수 있어 더욱 정교한 loss 계산이 가능할 것
Image = matrix of numbers
ways to handle 2d structure:
Multilayer Perceptron: Fully Connected Layer로만 구성됨
기존 CNN의 Feature Extraction + Regression