Neural Adaptive Sequential Monte Carlo

Authors: Shixiang (Shane) Gu, Zoubin Ghahramani, Richard E. Turner

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments indicate that NASMC significantly improves inference in a non-linear state space model outperforming adaptive proposal methods including the Extended Kalman and Unscented Particle Filters. Experiments also indicate that improved inference translates into improved parameter learning when NASMC is used as a subroutine of Particle Marginal Metropolis Hastings. Finally we show that NASMC is able to train a latent variable recurrent neural network (LV-RNN) achieving results that compete with the state-of-the-art for polymorphic music modelling.
Researcher Affiliation Academia Shixiang Gu Zoubin Ghahramani Richard E. Turner University of Cambridge, Department of Engineering, Cambridge UK MPI for Intelligent Systems, T ubingen, Germany sg717@cam.ac.uk, zoubin@eng.cam.ac.uk, ret26@cam.ac.uk
Pseudocode Yes Algorithm 1 Stochastic Gradient Adaptive SMC (batch inference and learning variants)
Open Source Code No The paper thanks 'the authors of [5] for releasing the source code', indicating a third-party release, but does not explicitly state that the authors of *this* paper are releasing their own source code for the methodology described.
Open Datasets Yes In order to evaluate the effectiveness of our adaptive SMC method, we tested our method on a standard nonlinear state-space model often used to benchmark SMC algorithms [2, 3]. As a second and more physically meaningful system we considered a cart-pole system that consists of an inverted pendulum that rests on a movable base [16]. Finally, the new method is used to train a latent variable recurrent neural network (LV-RNN) for modelling four polymorphic music datasets of varying complexity [17].
Dataset Splits Yes The hyperparameters are tuned using the validation set [17].
Hardware Specification No No specific hardware details (such as GPU/CPU models, memory, or cloud instances) used for running the experiments are provided in the paper. Only general statements about 'differing levels of acceleration' for run times are mentioned.
Software Dependencies No The paper mentions 'Adam [20] is used as the optimizer' and thanks 'Theano developers for their toolkit', but it does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes For all experiments, the parameters in the non-linear state-space model were fixed to (σv, σw) = (√10, 1). A random walk proposal is used to sample θ = (σv, σw), q(θ |θ) = N(θ |θ, diag([0.15, 0.08])). The prior over θ is set as IG(0.01, 0.01). θ is initialized as (10, 10), and the PMMH is run for 500 iterations. Both the LSTM layers in the generative and proposal models are set as 1000 units and Adam [20] is used as the optimizer.