arXiv subjects

Jelmer P. Borst

Publications and source records attributed to Jelmer P. Borst.

4 recordsLinked to original sources

arXiv2026

Structured Secant Methods to Select Smoothing Parameters for General Smooth Models

General smooth models replace parameters of a regular likelihood with additive models. The models can include parametric terms, Gaussian random effects, and smooth functions of covariates. The latter are parameterized via a reduced-rank spline basis and regularized via weighted quadratic penalties placed on the basis coefficients. Estimates for these weights (i.e., smoothing parameters) can be obtained by optimizing the Laplace-approximate Bayesian marginal likelihood. Existing (second-order) methods require the Hessian of the log-likelihood to solve this optimization problem approximately - exact optimization requires up to fourth order derivatives - which can be difficult to derive and expensive to evaluate. To address these problems, we present a quasi-Newton variant of the second-order Extended Fellner-Schall (EFS) optimization method. Our qEFS method relies on structured limited-memory secant approximations to the Hessian of the log-likelihood and is principally first-order. However, the approximation can also be accumulated for a sub-block of the Hessian, with the remaining columns being constrained to match those of the actual Hessian. The exact columns then provide additional structure for the sub-block approximation, which becomes more accurate as a result. We show that the qEFS method converges to the EFS method under certain conditions and continues to provide good estimates beyond these circumstances, which we illustrate in simulation studies. Secondary tasks involving the Hessian (confidence interval coverage & model selection) require partial approximations to achieve close to nominal performance. We provide Hidden Markov and Tweedie model examples, for which the qEFS method is substantially easier to implement than alternative methods.

stat.ME↗

arXiv2025

The Mixed-Sparse-Smooth-Model Toolbox (MSSM): Efficient Estimation and Selection of Large Multi-Level Statistical Models

Additive smooth models, such as Generalized additive models (GAMs) of location, scale, and shape (GAMLSS), are a popular choice for modeling experimental data. However, software available to fit such models is usually not tailored specifically to the estimation of mixed models. As a result, estimation can slow down as the number of random effects increases. Additionally, users often have to provide a substantial amount of problem-specific information in case they are interested in more general non-standard smooth models, such as higher-order derivatives of the likelihood. Here we combined and extended recently proposed strategies to reduce memory requirements and matrix infill into a theoretical framework that supports efficient estimation of general mixed sparse smooth models, including GAMs & GAMLSS, based only on the Gradient and Hessian of the log-likelihood. To make non-standard smooth models more accessible, we developed an approximate estimation algorithm (the L-qEFS update) based on limited-memory quasi-Newton methods. This enables estimation of any general smooth model based only on the log-likelihood function. We also considered the problem of model selection for general mixed smooth models. To facilitate practical application we provide a Python implementation of the theoretical framework, algorithms, and model selection strategies presented here: the Mixed-Sparse-Smooth-Model (MSSM) toolbox. MSSM supports estimation and selection of massive additive multi-level models that are impossible to estimate with alternative software, for example of trial level EEG data. Additionally, when the L-qEFS update is used for estimation, implementing a new non-standard smooth model in MSSM is straightforward. Results from multiple simulation studies and real data examples are presented, showing that the framework implemented in MSSM is both efficient and robust to numerical instabilities.

stat.ME↗

arXiv2022

A Robust Learning Rule for Soft-Bounded Memristive Synapses Competitive with Supervised Learning in Standard Spiking Neural Networks

Memristive devices are a class of circuit elements that shows great promise as future building block for brain-inspired computing. One influential view in theoretical neuroscience sees the brain as a function-computing device: given input signals, the brain applies a function in order to generate new internal states and motor outputs. Therefore, being able to approximate functions is a fundamental axiom to build upon for future brain research and to derive more efficient computational machines. In this work we apply a novel supervised learning algorithm - based on controlling niobium-doped strontium titanate memristive synapses - to learning non-trivial multidimensional functions. By implementing our method into the spiking neural network simulator Nengo, we show that we are able to at least match the performance obtained when using ideal, linear synapses and - in doing so - that this kind of memristive device can be harnessed as computational substrate to move towards more efficient, brain-inspired computing.

cs.ET↗

arXiv2020

Learning to Approximate Functions Using Nb-doped SrTiO$_3$ Memristors

Memristors have attracted interest as neuromorphic computation elements because they show promise in enabling efficient hardware implementations of artificial neurons and synapses. We performed measurements on interface-type memristors to validate their use in neuromorphic hardware. Specifically, we utilised Nb-doped SrTiO$_3$ memristors as synapses in a simulated neural network by arranging them into differential synaptic pairs, with the weight of the connection given by the difference in normalised conductance values between the two paired memristors. This network learned to represent functions through a training process based on a novel supervised learning algorithm, during which discrete voltage pulses were applied to one of the two memristors in each pair. To simulate the fact that both the initial state of the physical memristive devices and the impact of each voltage pulse are unknown we injected noise at each timestep. Nevertheless, discrete updates based on local knowledge were shown to result in robust learning performance. Using this class of memristive devices as the synaptic weight element in a spiking neural network yields, to our knowledge, one of the first models of this kind, capable of learning to be a universal function approximator, and strongly suggests the suitability of these memristors for usage in future computing platforms.

cs.ET↗