Gating, as used in gated recurrent units (GRUs) [4] and long short-term memory (LSTM) networks [12], has become common-place in recurrent architectures. Gates facilitate the learning of longer term temporal relationships [12]. Furthermore, in the presence of noise in the input signal, gates can protect the cell state from undesired updates, thereby improving overall stability and convergence.

This allowed FastGRNN to accurately recognize the "Hey Cortana" wakeword with a 1 KB model and to be

Recurrent neural networks (RNNs) are widely used for processing time series and sequential information.

We formulate and solve the fixed horizon linear quadratic covariance steering problem in continuous time with a terminal cost measured in Hilbert-Schmidt (i.e., Frobenius) norm error between the desired and the controlled terminal covariances. For this problem, the necessary conditions of optimality become a coupled matrix ODE two-point boundary value problem. To solve this system of equations, we design a matricial recursive algorithm and prove its convergence. The proposed algorithm and its analysis make use of the linear fractional transforms parameterized by the state transition matrix of the associated Hamiltonian matrix. To illustrate the results, we provide two numerical examples: one with a two dimensional and another with a six dimensional state space.

State-space models (SSMs) have recently gained attention in deep learning for their ability to efficiently model long-range dependencies, making them promising candidates for edge-AI applications. In this paper, we analyze the effects of quantization on small-scale SSMs with a focus on reducing memory and computational costs while maintaining task performance. Using the S4D architecture, we first investigate post-training quantization (PTQ) and show that the state matrix A and internal state x are particularly sensitive to quantization. Furthermore, we analyze the impact of different quantization techniques applied to the parameters and activations in the S4D architecture. To address the observed performance drop after Post-training Quantization (PTQ), we apply Quantization-aware Training (QAT), significantly improving performance from 40% (PTQ) to 96% on the sequential MNIST benchmark at 8-bit precision. We further demonstrate the potential of QAT in enabling sub-8-bit precisions and evaluate different parameterization schemes for QAT stability. Additionally, we propose a heterogeneous quantization strategy that assigns different precision levels to model components, reducing the overall memory footprint by a factor of 6x without sacrificing performance. Our results provide actionable insights for deploying quantized SSMs in resource-constrained environments.

Linear Recurrent Neural Networks (LRNNs) such as Mamba, RWKV, GLA, mLSTM, and DeltaNet have emerged as efficient alternatives to Transformers in large language modeling, offering linear scaling with sequence length and improved training efficiency. However, LRNNs struggle to perform state-tracking which may impair performance in tasks such as code evaluation or tracking a chess game. Even parity, the simplest state-tracking task, which non-linear RNNs like LSTM handle effectively, cannot be solved by current LRNNs. Recently, Sarrof et al. (2024) demonstrated that the failure of LRNNs like Mamba to solve parity stems from restricting the value range of their diagonal state-transition matrices to $[0, 1]$ and that incorporating negative values can resolve this issue. We extend this result to non-diagonal LRNNs, which have recently shown promise in models such as DeltaNet. We prove that finite precision LRNNs with state-transition matrices having only positive eigenvalues cannot solve parity, while complex eigenvalues are needed to count modulo $3$. Notably, we also prove that LRNNs can learn any regular language when their state-transition matrices are products of identity minus vector outer product matrices, each with eigenvalues in the range $[-1, 1]$. Our empirical results confirm that extending the eigenvalue range of models like Mamba and DeltaNet to include negative values not only enables them to solve parity but consistently improves their performance on state-tracking tasks. Furthermore, pre-training LRNNs with an extended eigenvalue range for language modeling achieves comparable performance and stability while showing promise on code and math data. Our work enhances the expressivity of modern LRNNs, broadening their applicability without changing the cost of training or inference.

We investigate whether transformers can learn to track a random process when given observations of a related process and parameters of the dynamical system that relates them as context. More specifically, we consider a finite-dimensional state-space model described by the state transition matrix $F$, measurement matrices $h_1, \dots, h_N$, and the process and measurement noise covariance matrices $Q$ and $R$, respectively; these parameters, randomly sampled, are provided to the transformer along with the observations $y_1,\dots,y_N$ generated by the corresponding linear dynamical system. We argue that in such settings transformers learn to approximate the celebrated Kalman filter, and empirically verify this both for the task of estimating hidden states $\hat{x}_{N|1,2,3,...,N}$ as well as for one-step prediction of the $(N+1)^{st}$ observation, $\hat{y}_{N+1|1,2,3,...,N}$. A further study of the transformer's robustness reveals that its performance is retained even if the model's parameters are partially withheld. In particular, we demonstrate that the transformer remains accurate at the considered task even in the absence of state transition and noise covariance matrices, effectively emulating operations of the Dual-Kalman filter.