The dim keyword argument specify in which dimension the n-dimensional array is going to be reduced. dims=1 means that the mean is going to be taken reducing over rows. dims=2 means reducing over columns.
For a 3D array, you can see how it works with this example
julia> a = [3; 5;; 4; 5;;;
9; 8;; 4; 9;;;
9; 9;; 5; 6]
2×2×3 Array{Int64, 3}:
[:, :, 1] =
3 4
5 5
[:, :, 2] =
9 4
8 9
[:, :, 3] =
9 5
9 6
julia> size(a) # rows, columns, depth
(2, 2, 3)
julia> mean(a; dims=1) # reduced over rows
1×2×3 Array{Float64, 3}:
[:, :, 1] =
4.0 4.5
[:, :, 2] =
8.5 6.5
[:, :, 3] =
9.0 5.5
julia> mean(a; dims=2) # reduced over columns
2×1×3 Array{Float64, 3}:
[:, :, 1] =
3.5
5.0
[:, :, 2] =
6.5
8.5
[:, :, 3] =
7.0
7.5
julia> mean(a; dims=3) # reduced over depth
2×2×1 Array{Float64, 3}:
[:, :, 1] =
7.0 4.33333
7.33333 6.66667
Hope it makes it clear. 