SabaPivot's picture
Upgrade canonical logbook from stronger peer evidence with attribution
9b57190 verified
Raw
History Blame Contribute Delete
1.21 kB
[
"ViPll models the posterior P(Y|X,S) over true labels using a Dirichlet variational distribution whose parameters are output by an amortized inference network, factorizing the generative process as P(Y)P(X|Y)P(S|X,Y) instead of the standard discriminative formulation (Section 4.1).",
"ViPll incorporates a conditional variational autoencoder to model P(X|Y) (Section 4.2) and a maximum-entropy prior satisfying candidate-label-set constraints (Section 4.3).",
"Training proceeds in three phases: initialization, a 500-epoch CVAE warm-up, and a 1000-epoch main training loop using Monte Carlo sampling (Section 4.5).",
"On 5 real-world partial-label datasets, ViPll attains 76.15% (+/-1.56) accuracy on Bird-song and 78.52% (+/-2.27) on Lost, outperforming 9 baselines including PlKnn, Proden, PiCO, and CroSel on the majority of settings (Table 2).",
"On synthetic partial-label benchmarks, ViPll reaches 96.64% (+/-3.92) accuracy on CIFAR-10 and 77.76% (+/-0.51) on CIFAR-100 (Table 3).",
"An ablation removing the modeling of Markov equivalence in the generative structure causes significant accuracy degradation, validating the choice of generative modeling (Section 4, ablation study)."
]