Low-Rank Nonlinear Decoding of μ-ECoG from the Primary Auditory Cortex

Abstract

This paper considers the problem of neural decoding from parallel neural measurements systems such as micro-electrocorticography (µ-ECoG). In systems with large numbers of array elements at very high sampling rates, the dimension of the raw measurement data may be large. Learning neural decoders for this high-dimensional data can be challenging, particularly when the number of training samples is limited. To address this challenge, this work presents a novel neural network decoder with a low-rank structure in the first hidden layer. The lowrank constraints dramatically reduce the number of parameters in the decoder while still enabling a rich class of nonlinear decoder maps. The low-rank decoder is illustrated on µ-ECoG data from the primary auditory cortex (A1) of awake rats. This decoding problem is particularly challenging due to the complexity of neural responses in the auditory cortex and the presence of confounding signals in awake animals. It is shown that the proposed low-rank decoder significantly outperforms models using standard dimensionality reduction techniques such as principal component analysis (PCA).

PAC-Bayes Generalisation Bounds for Dynamical Systems including Stable RNNs

“…Our contribution is consistent with recent results for linear RNNs (Emami et al 2021;Cohen-Karlik et al 2023), on their ability to extrapolate to longer sequences by training on short sequences with stochastic gradient descent. Thus we provide a different perspective, while extending to a more general class of models.…”

Double-Descent Curves in Neural Networks: A New Perspective Using Gaussian Processes

“…Our work improves on this line of research by introducing the idea of width-dependent kernels, which is especially well-suited to the context of DNNs where double descent manifests as the network width tends to infinity. Recent studies of the double-descent phenomenon have focused on random features regressions in the case of shallow networks (Gerbelot, Abbara, and Krzakala 2020; Liao, Couillet, and Mahoney 2020;Emami et al 2020;Gerace et al 2020;Li, Zhou, and Gretton 2021;Adlam and Pennington 2020b;Belkin, Hsu, and Xu 2020;Chen and Schaeffer 2021;Bosch et al 2022;D'Ascoli et al 2020), or kernel regression with no dependence on the width (Liu, Liao, and Suykens 2020;Mallinar et al 2022).…”

Preparation of Glycan Arrays Using Glycopeptides Derived From Biomaterials

“…Statistical inference for generalized linear models (GLMs) with large has been extensively studied. There are four typical settings on : moderately high-dimensional with small / (Liang and Du, 2012;Hsu and Mazumdar, 2023;Kuchelmeister and van de Geer, 2024); sparse and high-dimensional (van de Geer, 2008;James and Radchenko, 2009;Levy and Abramovich, 2023); non-sparse and proportionally high-dimensional , i.e., / → ∈ (0, ∞) Salehi et al, 2019;Aubin et al, 2020;Emami et al, 2020;Sawaya et al, 2023); and non-sparse and possibly infinite-dimensional (Wu et al, 2023). Note that our study adopts the last setting.…”