Stateful encoding of hypervectors

[cvigilv]

As discussed in person, an interesting idea for things to eventually add to the package is the idea of having some kind of way of maintaining a history on how hypervectors are constructed. The following snippet enables the creation of Memory structures, which have an internal structure that contains information on how a given hypervector was constructed.

I thinks this could be added as a new data structure once we finish the first pass through of the package. LMKWYT

abstract type AbstractMemory end
struct Memory{T <: AbstractHDV} <: AbstractMemory
   v::T
   values::Tuple{Vararg{Union{T,AbstractMemory}}}
   fun::Function
end

function Memory(fun::Function, hvs::Union{AbstractHDV,AbstractMemory}...)
   @assert length(unique(typeof.(hvs))) == 1
   Memory{typeof(collect(first(hvs)))}(fun(hvs...), hvs, fun)
end
Memory(hv::AbstractHDV) = Memory(hv, Tuple([hdv]), identity)

Base.show(io::IO, m::Memory) = print(io, "$(typeof(m.v))($(m.fun); $(length(m.values))x$(length(m.v)))")
Base.collect(m::Memory) = m.v

bundle(hvs::Union{AbstractMemory,AbstractHDV}...) = Memory(bundle(collect.(hvs)...), hvs, bundle)
bind(hvs::Union{AbstractHDV,AbstractMemory}...) = Memory(bind(collect.(hvs)...), hvs, bind)
shift(hv::Memory, i) = Memory(shift(collect.(hvs), i), hvs, x -> shift(x, i))

end

I think we can improve this a lot, but as a first prototype it gets the job done.

[cvigilv]

Relevant implementation on AbstractTrees.jl that we could use to make this better: link

[MichielStock]

Would first like to have the basics in order. This would be a super cool type for more advanced analysis but we should have some use cases to see which functionality to add (or not to add).

[cvigilv]

Sounds good. Maybe we can add this to the documentation as an example for "Advanced" use cases.

[cvigilv]

I'll add this to the tutorials as a "Extending HDC.jl" section, just to show more advanced usecases.