What is the meaning of the word logits in TensorFlow?

Question

I can understand that we have to feed the activation of artificial neurons in the last layer of the TensorFlow function given below, but why we call it logits? I think it's a mathematical function:

loss_function = tf.nn.softmax_cross_entropy_with_logits(
     logits = last_layer,
     labels = target_output
)

Can anyone answer this?

Answer 1

Logits are values that are used as input to softmax. To understand this better click here this is official by tensorflow.

Hence, Logits are used to map probabilities [0,1] to R [-inf, inf]

L=ln(p/1-p)      p=1/1+e^-L

From limits we conclude that the probability approaching to 0.5 corresponds to a logits of 0. Therefore, +ive logits correspond to probability of greater than 0.5 and negative corresponds to a probability value of less than 0.5.

Sometimes they are also refer to inverse of sigmoid function. It's mainly used in Statistics.

Answer 2

Logits is a simple term which can mean many different things: In Mathematics, Logitis a function that is used to maps probabilities ( [0, 1] ) to R ( (-inf, inf) ) .

Ex-

 N=ln M - ln (1-M) , where M=1/(1+e^-N)

Ex-

 A logits of 0 corresponds to a probability of 0.5.

A negative logit corresponds to a probability of less than 0.5.

A positive logit corresponds to a probability greater than 0.5.

In ML, logits can be defined as a vector of raw predictions that a classification model generates and it is passed to a normalized function. In case the model is solving a classification problem having multiple class, logits will behave as an input to the softmax function. Then the softmax function generates a vector of probability having one value for each class. Logits sometime also refer to the element-wise inverse of the sigmoid function.

Hope this answer helps.