Synapse.NN.Losses

Provides collection of functions that are used to as a reference of what needs to be minimised during training.

LossFn type alias represents those functions, and Synapse offers a variety of them.

LossFn type alias and Loss newtype

type LossFn a = SymbolMat a -> SymbolMat a -> SymbolMat a

LossFn type alias represents functions that are able to provide a reference of what relation between matrices needs to be minimised.

newtype Loss a

Loss newtype wraps LossFns - differentiable functions that are able to provide a reference of what relation between matrices needs to be minimised.

Every loss function must return symbol of singleton matrix.

Constructors

Loss

Fields

• unLoss :: LossFn a

  Unwraps Loss newtype.

Regression losses

mse :: (Symbolic a, Floating a) => LossFn a

Computes the mean of squares of errors.

msle :: (Symbolic a, Floating a) => LossFn a

Computes the mean squared logarithmic error.

mae :: (Symbolic a, Floating a) => LossFn a

Computes the mean of absolute error.

mape :: (Symbolic a, Floating a) => LossFn a

Computes the mean absolute percentage error.

logcosh :: (Symbolic a, Floating a) => LossFn a

Computes the logarithm of the hyperbolic cosine of the error.