This is the code for constructing UNFs, from the paper Universal Neural Functionals. UNFs are architectures that can process the weights of other neural networks, while maintaining equivariance or invariance to the weight space permutation symmetries. In contrast to NFNs, UNFs can ingest weights from any architecture.
Equivalently, we can think of UNFs as equivariant architectures for processing any collection of tensors, where the action involves a shared set of permutations permuting the axes of the tensors in a given way.
The codebase requires JAX for core functionality and Flax for the example (though other Jax NN libraries are likely compatible as well). See usage in example.py.
The perm_spec is what tells our library the permutation symmetries it should be equivariant to. For example, suppose you have a collection of weight tensors corresponding to a simple MLP:
params = {
"params": {
"Dense_0": {
"kernel": Array[784, 512],
"bias": Array[512]
},
"Dense_1": {
"kernel": Array[512, 10],
"bias": Array[10]
}
}
}We can describe the permutation symmetry of this network as follows (assume the input and output neurons are also permutable).
- The weight tensors can be permuted by
$\sigma=(\sigma_0, \sigma_1, \sigma_2) \in S_{784} \times S_{512} \times S_{10}$ . -
$\sigma_0$ permutes the first dimension ofparams["params"]["Dense_0"]["kernel"]. -
$\sigma_1$ permutes the second dimension ofparams["params"]["Dense_0"]["kernel"], the vectorparams["params"]["Dense_0"]["bias"], and the first dimension ofparams["params"]["Dense_1"]["kernel"]. -
$\sigma_2$ permutes the second dimension ofparams["params"]["Dense_1"]["kernel"]and the vectorparams["params"]["Dense_1"]["bias"].
Then we number each permutation by integers:
perm_spec = {
"params": {
"Dense_0": {
"kernel": (0, 1),
"bias": (1,)
},
"Dense_1": {
"kernel": (1, 2),
"bias": (2,)
}
}
}Notice that nothing requires the input to be a collection of weight tensors. This library processes any collection of tensors if you give it a description of the permutation symmetries.