Fully backend-agnostic Process API

[mathisrichter]

User story

As a user, I want to replace the backend of any Lava Process without having to change its parameters. This is one of the core promises of Lava to enable productive work in heterogeneous hardware environments.

Conditions of satisfaction

• All parameters of Lava Processes are agnostic to the ProcessModel chosen for execution.
• All parameters are abstract enough to not be tied to a particular implementation.

Acceptance tests

• to be determined

Originally raised by @epaxon

Current behavior

• Setting the same parameter values/using the same processes has different behavior on Loihi 2 and CPU backends for both LIF and Dense.

# Loihi 2 sim (bit_accurate_loihi is the same):
inport_plug = LIF(shape=(4,), vth=10, du=0, dv=0, bias_mant=80)

dense_weights = Dense(weights=weights)

lif_layer = LIF(shape=(3,), vth=10, du=4095, dv=0, bias_mant=0)
inport_plug.s_out.connect(dense_weights.s_in)
dense_weights.a_out.connect(lif_layer.a_in)

# What it takes to get same result with default LIF:
inport_plug = LIF(shape=(4,), vth=10*2**6, du=0, dv=0, bias_mant=80)

dense_weights = Dense(weights=weights*2**6)

lif_layer = LIF(shape=(3,), vth=10*2**6, du=4096/2**12, dv=0, bias_mant=0)
inport_plug.s_out.connect(dense_weights.s_in)
dense_weights.a_out.connect(lif_layer.a_in)

[tim-shea]

I am pretty sure this is just not possible in theory. The best one could do in most cases is to give the user a bunch of convenience tools that pretend like two very different versions of a process model behave the same, but these kinds of tools have proven to be difficult research projects in themselves.

My counter-suggestion would be to strongly prefer, at a Lava community standards level, for Process developers to provide backend agnostic parameterizations, but also create a simple system for providing backend-specific parameterizations when that is what's needed. I think the HyperParameters object in Lava-Optimization could be a reasonable model for this.

[tim-shea]

Since there is no reply to the above, and this isn't blocking anything, and this issue doesn't correspond to a clearly defined set of changes needed in the code, I'm closing for now. I'm going to submit a new issue specific to LIF parameterization and I recommend we do the same for any other process models that have significant issues with hardware-specific parameters.