Huge differences between loss functions [closed]

I have a model on TensorFlow that has 2 inputs: some text data and a set of 3 numerical values. It outputs only one single float value. Here is its structure.

num_input = keras.Input(shape= (30,3), name="nums")
nmlstm = layers.LSTM(128, return_sequences=True)(num_input)
nmdense1 = layers.Dense(128, activation='relu')(nmlstm)
nmdense2 = layers.Dense(128, activation='relu')(nmdense1)
nmdense3 = layers.Dense(128, activation='relu')(nmdense2)

text_input = keras.Input(shape= (30, max_sequence_length), name="text")
text_vec = layers.Embedding(vocab_size, mask_zero=True, output_dim=embedding_dim)(text_input)
txtds1 = keras.layers.TimeDistributed(layers.LSTM(64, return_sequences=True))(text_vec)
txtds2 = keras.layers.TimeDistributed(layers.LSTM(64, return_sequences=True))(txtds1)
txtds3 = keras.layers.TimeDistributed(layers.LSTM(64, return_sequences=True))(txtds2)
txrspd = layers.Reshape((30, 128))(txtds3)
txdense = layers.Dense(128, activation='relu')(txrspd)

concat = layers.concatenate([nmdense3, txdense])
prcs1 = layers.Dense(128, activation='relu')(concat)
prcs2 = layers.Dense(128, activation='relu')(prcs1)
pooling = layers.GlobalAveragePooling1D()(prcs2)
last = layers.Dense(1, name='prediction')(pooling)

I tried different loss functions for the model and got different results, which vary largely.

So, these are the results:

MeanSquaredError

loss: 0.038  

                           

MeanAbsoluteError

loss: 0.108 

                                        

Huber

loss: 0.0185

Does using Huber over Mean Squared Error loss function mean that the model will perform different? What will change in the training process if switch the loss function? (I assume something is because of the huge differences) Also, if "loss" is the difference between input and output values, what does loss even mean when there are strings and multiple floats as input and a single float output?

source https://stackoverflow.com/questions/77397729/huge-differences-between-loss-functions