Limited availability of experimental data from different testing configurations poses major challenges for the identification of constitutive model parameters in complex non-linear, anisotropic and time-dependent material behavior. Soft biological tissues, characterized by anisotropic and non-linear viscoelastic responses, are a prominent example. When experimental data are insufficient to reliably calibrate a constitutive model for generic loading conditions, determining plausible ranges of model responses rather than a single deterministic solution becomes pivotal. This motivates uncertainty analysis strategies [1]. However, Monte-Carlo approaches based on repeated forward evaluations remain computationally demanding for advanced constitutive models with internal state variables due to implicit time integration and high-dimensional parameter spaces. Even relatively simple methods such as Generalized Likelihood Uncertainty Estimation (GLUE) [2] therefore become prohibitively expensive when combined with traditional implicit solvers. In this contribution, we compare two neural-network surrogate modeling strategies for efficient likelihood-based analysis. The approach is demonstrated using GLUE applied to a non-linear, anisotropic, compressible and viscoelastic constitutive model for soft tissue membranes [3,4], calibrated using only uniaxial tension data at different strain rates. As a reference, GLUE is first applied to implicitly integrated solutions, highlighting the high computational cost associated with likelihood-weight parameter sampling and forward prediction. Physics-informed neural networks are then introduced as surrogate models enabling fast, one-shot forward evaluations via GPU acceleration, but are shown to be trainable only over relatively narrow parameter ranges. Alternatively, feature-based artificial neural networks can be proposed as a surrogate strategy without the limit of narrow parameter ranges. By representing stress–strain responses using low-dimensional B-spline features, Feature-NNs enable efficient forward evaluation across broader parameter spaces and thus to efficiently perform GLUE-based analyses. Overall, the proposed framework provides an efficient and practical approach for uncertainty-aware forward prediction in complex nonlinear viscoplastic constitutive models in view of limited experimental data unseen, achieving substantially reduced computational cost.