➩ More data can be created for training using generative models or unsupervised learning techniques.

➩ A validation set can be used (similar to pruning for decision trees) to train the network until the loss (or accuracy) on the validation set begins to increase.

➩ Dropout can be used: randomly omitting units (random pruning of weights) during training so the rest of the units will have a better performance: Wikipedia.

➩ \(L_{1}\) regularization adds \(L_{1}\) norm of the weights and biases to the loss, or \(C = C_{1} + C_{2} + ... + C_{n} + \lambda \left\|\begin{bmatrix} w \\ b \end{bmatrix}\right\|_{1}\), for example, if there are no hidden layers, \(\left\|\begin{bmatrix} w \\ b \end{bmatrix}\right\|_{1} = \left| w_{1} \right| + \left| w_{2} \right| + ... + \left| w_{m} \right| + \left| b \right|\). Linear regression with \(L_{1}\) regularization is also called LASSO (Least Absolute Shrinkage and Selector Operator): Wikipedia.

➩ \(L_{2}\) regularization adds \(L_{2}\) norm of the weights and biases to the loss, or \(C = C_{1} + C_{2} + ... + C_{n} + \lambda \left\|\begin{bmatrix} w \\ b \end{bmatrix}\right\|^{2}_{2}\), for example, if there are no hidden layers, \(\left\|\begin{bmatrix} w \\ b \end{bmatrix}\right\|^{2}_{2} = \left(w_{1}\right)^{2} + \left(w_{2}\right)^{2} + ... + \left(w_{m}\right)^{2} + \left(b\right)^{2}\). Linear regression with \(L_{2}\) regularization is also called ridge regression: Wikipedia.

➩ \(L_{1}\) regularization often leads to more weights that are exactly \(0\), which is useful for feature selection.

➩ \(L_{2}\) regularization is easier for gradient descent since it is differentiable.

➩ Try \(L_{1}\) vs \(L_{2}\) regularization here: Link.