Which statements about attention mechanism are true?

a$.$ A spatial attention mechanism causes the network to treat each pixel in an image as equally important.

b$.$ The scaled dot$-$product attention mechanism generally has the form softmax$(\frac{1}{\sqrt[\phantom{2}]{d}}Q{K}^{TT})V,$ where $K,Q,Vin{R}^{Ld}$ are keys, queries, and values, respectively, and $L$ is the sequence length and $d$ is the hidden dimension.

c$.$ The attention score $e(i)$ is obtained from the attention vector $a(i)$ via the softmax function.

d$.$ An LSTM with self$-$attention has access to all previous hidden states and cell states via an softmax$-$based attention mechanism.

e$.$ The LSTM gates can be interpreted as a temporal attention mechanism.