Prior Network on Two Moons

A Prior Network parameterizes a Dirichlet distribution over class probabilities, with concentration parameters alpha = exp(logits). What distinguishes it from evidential classification is not so much the model as the training signal. Rather than learning from in-distribution data alone, a Prior Network is trained with explicit out-of-distribution (OOD) data: in-distribution inputs are pushed towards a sharp Dirichlet on the true class, and OOD inputs towards a flat one. The model is thereby encouraged to separate aleatoric uncertainty, which arises where the classes overlap, from epistemic uncertainty, which arises far from the training data.

from __future__ import annotations

from sklearn.datasets import make_moons
import torch
from torch import nn
from torch.distributions import Dirichlet, kl_divergence
from torch.nn import functional as F
from torch.utils.data import DataLoader, TensorDataset

from probly.losses.torch import make_ood_target_alpha
from probly.method.prior_network import prior_network
from probly.representer import representer

from examples.utils.model import MLPClassifier
from examples.utils.plotting import plot_example_uncertainty

Setup

X, y = make_moons(n_samples=500, noise=0.05, random_state=0)
X_tensor = torch.from_numpy(X).float()
y_tensor = torch.from_numpy(y).long()

dataset = TensorDataset(X_tensor, y_tensor)
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

OOD Data

Since the method requires OOD inputs during training, these are generated synthetically. Points are sampled uniformly from the plotting region, and only those at a distance of at least 0.5 from every training point are kept, so that no OOD target is placed on the moons.

ood_candidates = torch.rand(5000, 2) * 6 - 3
min_dist = torch.cdist(ood_candidates, X_tensor).min(dim=1).values
X_ood = ood_candidates[min_dist > 0.5]

Model

base_model = MLPClassifier()
pn_model = prior_network(base_model, predictor_type="logit_classifier")

Training

The loss mirrors pn_loss from probly.losses.torch, specialized to two classes. It sums the KL divergences between target and predicted Dirichlet distributions on in-distribution and OOD inputs, and adds a small negative log-likelihood term on the mean class probabilities, which stabilizes classification. In-distribution targets have concentration 10 on the true class and 1 on the other. The OOD target is the flat Dirichlet Dir(1, 1), as in the original paper. Its low total concentration yields a large mutual information, that is, a large epistemic uncertainty, away from the data.

def prior_network_loss(x_in: torch.Tensor, y_in: torch.Tensor, x_ood: torch.Tensor) -> torch.Tensor:
    alpha_in = pn_model(x_in)
    target_in = F.one_hot(y_in, 2).float() * 9 + 1
    kl_in = kl_divergence(Dirichlet(target_in), Dirichlet(alpha_in)).mean()

    probs_in = alpha_in / alpha_in.sum(dim=-1, keepdim=True)
    nll = F.nll_loss(torch.log(probs_in + 1e-8), y_in)

    alpha_ood = pn_model(x_ood)
    target_ood = make_ood_target_alpha(x_ood.size(0), num_classes=2, alpha0=2)
    kl_ood = kl_divergence(Dirichlet(target_ood), Dirichlet(alpha_ood)).mean()

    return kl_in + kl_ood + 0.1 * nll


opt = torch.optim.Adam(pn_model.parameters(), lr=1e-3)
grad_clip_norm = 0.5

pn_model.train()
for _epoch in range(200):
    for inputs, targets in dataloader:
        opt.zero_grad()
        ood_idx = torch.randint(0, X_ood.size(0), (inputs.size(0),))
        loss = prior_network_loss(inputs, targets, X_ood[ood_idx])
        loss.backward()
        nn.utils.clip_grad_norm_(pn_model.parameters(), grad_clip_norm)
        opt.step()

Evaluation

Away from the moons, where the model was trained to predict a flat Dirichlet, the epistemic uncertainty is high.

pn_model.eval()
rep = representer(pn_model)

plot = plot_example_uncertainty(X, y, rep, title="Prior Network Epistemic Uncertainty", notion="epistemic")
plot.show()
Prior Network Epistemic Uncertainty

Aleatoric uncertainty, in contrast, concentrates where the two classes are hard to tell apart, i.e., near the boundary between the moons.

plot = plot_example_uncertainty(X, y, rep, title="Prior Network Aleatoric Uncertainty", notion="aleatoric")
plot.show()
Prior Network Aleatoric Uncertainty

Total running time of the script: (0 minutes 10.584 seconds)

Gallery generated by Sphinx-Gallery