Which statements about the self$-$attention mechanism are correct?

Verbleibende Zeit $0$:$50$:$23$

a$.$ Similar to the hidden state in an LSTM$,$ the attention matrix is updated in a recurrent manner for each token in an input sequence.

b$.$ The raw attention scores are normalized by applying a sigmoid activation function row$-$wise to the attention matrix.

c$.$ The vector at position $t$ in the output sequence of a self$-$attention layer is computed by summing the value vectors, each weighted by its corresponding attention score from the $t-$th row of the attention matrix.

d$.$ In the standard self$-$attention formulation, cosine similarity is used as a measure of alignment between keys and values.

e$.$ The Multi$-$head Attention mechanism consists of a total of four weight matrices.