Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Addressing Task A of the 1st DAFx Parameter Estimation Challenge, this work frames physical plate reverberation system identification as an amortized Simulation-Based Inference (SBI) problem to retrieve mechanical parameters directly from measured impulse responses. Rather than relying on slow numerical optimization or intractable likelihoods, a neural network is trained on synthetic impulse responses to directly estimate the parameter posterior in a single forward pass. To maximize estimation accuracy on target observations, the authors implement an active, multi-round scheme that simulates new data conditioned on the current posterior estimate to fine-tune an observation-specialized network.
Multi-round active simulation turns standard neural density estimators into high-precision, test-time-specialized solvers for notoriously ill-posed physical inverse problems.
We address Task A of the 1st DAFx Parameter Estimation Challenge, which aims to retrieve the physical parameters of a plate model from an impulse response. To do so, we use the Simulation-Based Inference (SBI) framework, in which we train a neural network to estimate a density over plate parameters given an impulse response, using a dataset generated by the simulator. Inference for a new impulse response then requires only a forward pass through the network, without involving the simulator. For each test observation, we fine-tune a specific network: additional simulation rounds are performed by sampling parameters from the current estimated distribution, simulating the corresponding impulse responses, and fine-tuning to produce the specialized network.