Recurrent Memory

Neural population activity often exhibits distinct dynamical features across time, which may correspond to distinct internal processes or behavior. Linear methods and variations thereof, such as Hidden Markov Model (HMM) and Switching Linear Dynamical System (SLDS), are often employed to identify discrete states with evolving neural dynamics. However, these techniques may not be able to capture the underlying nonlinear dynamics associated with neural propagation. Recurrent Neural Networks (RNNs) are commonly used to model neural dynamics thanks to their nonlinear characteristics. In our work, we develop Switching Recurrent Neural Networks (SRNN), RNNs with weights that switch across time, to reconstruct switching dynamics of neural time-series data. We apply these models to simulated data as well as cortical neural activity across mice and monkeys, which allows us to automatically detect discrete states that lead to the identification of varying neural dynamics. In a monkey reaching dataset with electrophysiology recordings, a mouse self-initiated lever pull dataset with widefield calcium recordings, and a mouse self-initiated decision making dataset with widefield calcium recording, SRNNs are able to automatically identify discrete states with distinct nonlinear neural dynamics. The inferred switches are aligned with the behavior, and the reconstructions show that the recovered neural dynamics are distinct across different stages of the behavior. We show that the neural dynamics have behaviorally-relevant switches across time and we are able to use SRNNs to successfully capture these switches and the corresponding dynamical features.

This paper advances temporal reasoning within dynamically changing high-dimensional noisy observations, focusing on a latent space that characterizes the nonlinear dynamics of objects in their environment. We introduce the *Gated Inference Network* (GIN), an efficient approximate Bayesian inference algorithm for state space models (SSMs) with nonlinear state transitions and emissions. GIN disentangles two latent representations: one representing the object derived from a nonlinear mapping model, and another representing the latent state describing its dynamics. This disentanglement enables direct state estimation and missing data imputation as the world evolves. To infer the latent state, we utilize a deep extended Kalman filter (EKF) approach that integrates a novel compact RNN structure to compute both the Kalman Gain (KG) and smoothing gain (SG), completing the data flow. This design results in a computational cost per step that is linearly faster than EKF but introduces issues such as the exploding gradient problem. To mitigate the exploding gradients caused by the compact RNN structure in our model, we propose a specialized learning method that ensures stable training and inference. The model is then trained end-to-end on videos depicting a diverse range of simulated and real-world physical systems, and outperforms its ounterparts —RNNs, autoregressive models, and variational approaches— in state estimation and missing data imputation tasks.

Natural behaviors, even stereotyped ones, exhibit variability. Despite its role in exploring and learning, the function and neural basis of this variability is still not well understood. Given the coupling between neural activity and behavior, we ask what type of neural variability does not compromise behavioral performance. While previous studies typically curtail variability to allow for high task performance in neural networks, our approach takes the reversed perspective. We investigate how to generate maximal neural variability while at the same time having high network performance. To do so, we extend to neural activity the maximum occupancy principle (MOP) developed for behavior, and refer to this new neural principle as NeuroMOP. NeuroMOP posits that the goal of the nervous system is to maximize future action-state entropy, a reward-free, intrinsic motivation that entails creating all possible activity patterns while avoiding terminal or dangerous ones. We show that this goal can be achieved through a neural network controller that injects currents (actions) into a recurrent neural network of fixed random weights to maximize future cumulative action-state entropy. High activity variability can be induced while adhering to an energy constraint or while avoiding terminal states defined by specific neurons' activities, also in a context-dependent manner. The network solves these tasks by flexibly switching between stochastic and deterministic modes as needed and projecting noise onto a null space. Based on future maximum entropy production, NeuroMOP contributes to a novel theory of neural variability that reconciles stochastic and deterministic behaviors within a single framework.

Transformers with linear attention (i.e., linear transformers) and state-space models have recently been suggested as a viable linear-time alternative to transformers with softmax attention. However, these models still underperform transformers especially on tasks that require in-context retrieval. While more expressive variants of linear transformers which replace the additive update in linear transformers with the delta rule (DeltaNet) have been found to be more effective at associative recall, existing algorithms for training such models do not parallelize over sequence length and are thus inefficient to train on modern hardware. This work describes a hardware-efficient algorithm for training linear transformers with the delta rule, which exploits a memory-efficient representation for computing products of Householder matrices. This algorithm allows us to scale up DeltaNet to standard language modeling settings. We train a 1.3B model for 100B tokens and find that it outperforms recent linear-time baselines such as Mamba and GLA in terms of perplexity and zero-shot performance on downstream tasks. We also experiment with two hybrid models which combine DeltaNet layers with (1) sliding-window attention layers every other layer or (2) two global attention layers, and find that these hybrids outperform strong transformer baselines.

How do brain circuits learn to generate behaviour? While significant strides have been made in understanding learning in artificial neural networks, applying this knowledge to biological networks remains challenging. For instance, while backpropagation is known to perform accurate credit assignment of error in artificial neural networks, how a similarly powerful process can be realized within the constraints of biological circuits remains largely unclear. One of the major challenges is that the brain's extensive recurrent connectivity requires the propagation of error through both space and time, a problem that is notoriously difficult to solve in vanilla recurrent neural networks. Moreover, the extensive feedback connections in the brain are known to influence forward network activity, but the interaction between feedback-driven activity changes and local, synaptic plasticity-based learning is not fully understood. Building on our previous work modelling motor learning, this work investigates the mechanistic properties of pre-trained networks with feedback control on a standard motor task. We show that feedback control of the ongoing recurrent network dynamics approximates the optimal first-order gradient with respect to the network activities, allowing for rapid, ongoing movement correction. Moreover, we show that trial-by-trial adaptation to a persistent perturbation using a local, biologically plausible learning rule that integrates recent activity and error feedback is both more accurate and more efficient with feedback control during learning, due to the decoupling of the recurrent network dynamics and the injection of an adaptive, second-order gradient into the network dynamics. Thus, our results suggest that feedback control may guide credit assignment in biological recurrent neural networks, enabling both rapid and efficient learning in the brain.

A wide array of sequence models are built on a framework modeled after Transformers, comprising alternating sequence mixer and channel mixer layers. This paper studies a unifying *matrix mixer* view of sequence mixers that can be conceptualized as a linear map on the input sequence. This framework encompasses a broad range of well-known sequence models, including the self-attention of Transformers as well as recent strong alternatives such as structured state space models (SSMs), and allows understanding downstream characteristics such as efficiency and expressivity through properties of their structured matrix class. We identify a key axis of matrix parameterizations termed *sequence alignment*, which increases the flexibility and performance of matrix mixers, providing insights into the strong performance of Transformers and recent SSMs such as Mamba. Furthermore, the matrix mixer framework offers a systematic approach to developing sequence mixers with desired properties, allowing us to develop several new sub-quadratic sequence models. In particular, we propose a natural bidirectional extension of the Mamba model (**Hydra**), parameterized as a *quasiseparable matrix mixer*, which demonstrates superior performance over other sequence models including Transformers on non-causal tasks. As a drop-in replacement for attention layers, \name outperforms BERT by 0.8 points on the GLUE benchmark and ViT by 2% Top-1 accuracy on ImageNet.

Memory models such as Recurrent Neural Networks (RNNs) and Transformers address Partially Observable Markov Decision Processes (POMDPs) by mapping trajectories to latent Markov states. Neither model scales particularly well to long sequences, especially compared to an emerging class of memory models called Linear Recurrent Models. We discover that the recurrent update of these models resembles a monoid, leading us to reformulate existing models using a novel monoid-based framework that we call memoroids. We revisit the traditional approach to batching in recurrent reinforcement learning, highlighting theoretical and empirical deficiencies. We leverage memoroids to propose a batching method that improves sample efficiency, increases the return, and simplifies the implementation of recurrent loss functions in reinforcement learning.

Exposure Correction (EC) aims to recover proper exposure conditions for images captured under over-exposure or under-exposure scenarios. While existing deep learning models have shown promising results, few have fully embedded Retinex theory into their architecture, highlighting a gap in current methodologies. Additionally, the balance between high performance and efficiency remains an under-explored problem for exposure correction task. Inspired by Mamba which demonstrates powerful and highly efficient sequence modeling, we introduce a novel framework based on \textbf{Mamba} for \textbf{E}xposure \textbf{C}orrection (\textbf{ECMamba}) with dual pathways, each dedicated to the restoration of reflectance and illumination map, respectively. Specifically, we firstly derive the Retinex theory and we train a Retinex estimator capable of mapping inputs into two intermediary spaces, each approximating the target reflectance and illumination map, respectively. This setup facilitates the refined restoration process of the subsequent \textbf{E}xposure \textbf{C}orrection \textbf{M}amba \textbf{M}odule (\textbf{ECMM}). Moreover, we develop a novel \textbf{2D S}elective \textbf{S}tate-space layer guided by \textbf{Retinex} information (\textbf{Retinex-SS2D}) as the core operator of \textbf{ECMM}. This architecture incorporates an innovative 2D scanning strategy based on deformable feature aggregation, thereby enhancing both efficiency and effectiveness. Extensive experiment results and comprehensive ablation studies demonstrate the outstanding performance and the importance of each component of our proposed ECMamba. Code is available at \url{https://github.com/LowlevelAI/ECMamba}.

Domain Generalization (DG) aims to enable models to generalize to unseen target domains by learning from multiple source domains. Existing DG methods primarily rely on convolutional neural networks (CNNs), which inherently learn texture biases due to their limited receptive fields, making them prone to overfitting source domains. While some works have introduced transformer-based methods (ViTs) for DG to leverage the global receptive field, these methods incur high computational costs due to the quadratic complexity of self-attention. Recently, advanced state space models (SSMs), represented by Mamba, have shown promising results in supervised learning tasks by achieving linear complexity in sequence length during training and fast RNN-like computation during inference. Inspired by this, we investigate the generalization ability of the Mamba model under domain shifts and find that input-dependent matrices within SSMs could accumulate and amplify domain-specific features, thus hindering model generalization. To address this issue, we propose a novel SSM-based architecture with saliency-based token-aware transformation (namely START), which achieves state-of-the-art (SOTA) performances and offers a competitive alternative to CNNs and ViTs. Our START can selectively perturb and suppress domain-specific features in salient tokens within the input-dependent matrices of SSMs, thus effectively reducing the discrepancy between different domains. Extensive experiments on five benchmarks demonstrate that START outperforms existing SOTA DG methods with efficient linear complexity. Our code is available at https://github.com/lingeringlight/START.

Sequential memory, the ability to form and accurately recall a sequence of events or stimuli in the correct order, is a fundamental prerequisite for biological and artificial intelligence as it underpins numerous cognitive functions (e.g., language comprehension, planning, episodic memory formation, etc.) However, existing methods of sequential memory suffer from catastrophic forgetting, limited capacity, slow iterative learning procedures, low-order Markov memory, and, most importantly, the inability to represent and generate multiple valid future possibilities stemming from the same context. Inspired by biologically plausible neuroscience theories of cognition, we propose Predictive Attractor Models (PAM), a novel sequence memory architecture with desirable generative properties. PAM is a streaming model that learns a sequence in an online, continuous manner by observing each input only once. Additionally, we find that PAM avoids catastrophic forgetting by uniquely representing past context through lateral inhibition in cortical minicolumns, which prevents new memories from overwriting previously learned knowledge. PAM generates future predictions by sampling from a union set of predicted possibilities; this generative ability is realized through an attractor model trained alongside the predictor. We show that PAM is trained with local computations through Hebbian plasticity rules in a biologically plausible framework. Other desirable traits (e.g., noise tolerance, CPU-based learning, capacity scaling) are discussed throughout the paper. Our findings suggest that PAM represents a significant step forward in the pursuit of biologically plausible and computationally efficient sequential memory models, with broad implications for cognitive science and artificial intelligence research. Illustration videos and code are available on our project page: https://ramymounir.com/publications/pam.

The vertebrate hippocampus is thought to use recurrent connectivity in area CA3 to support episodic memory recall from partial cues. This brain area also contains place cells, whose location-selective firing fields implement maps supporting spatial memory. Here we show that place cells emerge in networks trained to remember temporally continuous sensory episodes. We model CA3 as a recurrent autoencoder that recalls and reconstructs sensory experiences from noisy and partially occluded observations by agents traversing simulated arenas. The agents move in realistic trajectories modeled from rodents and environments are modeled as continuously varying, high-dimensional, sensory experience maps (spatially smoothed Gaussian random fields). Training our autoencoder to accurately pattern-complete and reconstruct sensory experiences with a constraint on total activity causes spatially localized firing fields, i.e., place cells, to emerge in the encoding layer. The emergent place fields reproduce key aspects of hippocampal phenomenology: a) remapping (maintenance of and reversion to distinct learned maps in different environments), implemented via repositioning of experience manifolds in the network’s hidden layer, b) orthogonality of spatial representations in different arenas, c) robust place field emergence in differently shaped rooms, with single units showing multiple place fields in large or complex spaces, and (d) slow representational drift of place fields. We argue that these results arise because continuous traversal of space makes sensory experience temporally continuous. We make testable predictions: a) rapidly changing sensory context will disrupt place fields, b) place fields will form even if recurrent connections are blocked, but reversion to previously learned representations upon remapping will be abolished, c) the dimension of temporally smooth experience sets the dimensionality of place fields, including during virtual navigation of abstract spaces.

Softmax attention is the principle backbone of foundation models for various artificial intelligence applications, yet its quadratic complexity in sequence length can limit its inference throughput in long-context settings. To address this challenge, alternative architectures such as linear attention, State Space Models (SSMs), and Recurrent Neural Networks (RNNs) have been considered as more efficient alternatives. While connections between these approaches exist, such models are commonly developed in isolation and there is a lack of theoretical understanding of the shared principles underpinning these architectures and their subtle differences, greatly influencing performance and scalability. In this paper, we introduce the Dynamical Systems Framework (DSF), which allows a principled investigation of all these architectures in a common representation. Our framework facilitates rigorous comparisons, providing new insights on the distinctive characteristics of each model class. For instance, we compare linear attention and selective SSMs, detailing their differences and conditions under which both are equivalent. We also provide principled comparisons between softmax attention and other model classes, discussing the theoretical conditions under which softmax attention can be approximated. Additionally, we substantiate these new insights with empirical validations and mathematical arguments. This shows the DSF's potential to guide the systematic development of future more efficient and scalable foundation models.

Structured state space models (SSMs), the core engine behind prominent neural networks such as S4 and Mamba, are linear dynamical systems adhering to a specified structure, most notably diagonal. In contrast to typical neural network modules, whose parameterizations are real, SSMs often use complex parameterizations. Theoretically explaining the benefits of complex parameterizations for SSMs is an open problem. The current paper takes a step towards its resolution, by establishing formal gaps between real and complex diagonal SSMs. Firstly, we prove that while a moderate dimension suffices in order for a complex SSM to express all mappings of a real SSM, a much higher dimension is needed for a real SSM to express mappings of a complex SSM. Secondly, we prove that even if the dimension of a real SSM is high enough to express a given mapping, typically, doing so requires the parameters of the real SSM to hold exponentially large values, which cannot be learned in practice. In contrast, a complex SSM can express any given mapping with moderate parameter values. Experiments corroborate our theory, and suggest a potential extension of the theory that accounts for selectivity, a new architectural feature yielding state of the art performance.

In long-term time series forecasting (LTSF) tasks, an increasing number of works have acknowledged that discrete time series originate from continuous dynamic systems and have attempted to model their underlying dynamics. Recognizing the chaotic nature of real-world data, our model, Attraos, incorporates chaos theory into LTSF, perceiving real-world time series as low-dimensional observations from unknown high-dimensional chaotic dynamical systems. Under the concept of attractor invariance, Attraos utilizes non-parametric Phase Space Reconstruction embedding along with a novel multi-resolution dynamic memory unit to memorize historical dynamical structures, and evolves by a frequency-enhanced local evolution strategy. Detailed theoretical analysis and abundant empirical evidence consistently show that Attraos outperforms various LTSF methods on mainstream LTSF datasets and chaotic datasets with only one-twelfth of the parameters compared to PatchTST.

Recently, recurrent models based on linear state space models (SSMs) have shown promising performance in language modeling (LM), competititve with transformers. However, there is little understanding of the in-principle abilities of such models, which could provide useful guidance to the search for better LM architectures. We present a comprehensive theoretical study of the capacity of such SSMs as it compares to that of transformers and traditional RNNs. We find that SSMs and transformers have overlapping but distinct strengths. In star-free state tracking, SSMs implement straightforward and exact solutions to problems that transformers struggle to represent exactly. They can also model bounded hierarchical structure with optimal memory even without simulating a stack. On the other hand, we identify a design choice in current SSMs that limits their expressive power. We discuss implications for SSM and LM research, and verify results empirically on a recent SSM, Mamba.

Success in collaborative and competitive environments, where agents must work with or against each other, requires individuals to encode the position and trajectory of themselves and others. Decades of neurophysiological experiments have shed light on how brain regions [e.g., medial entorhinal cortex (MEC), hippocampus] encode the self's position and trajectory. However, it has only recently been discovered that MEC and hippocampus are modulated by the positions and trajectories of others. To understand how encoding spatial information of multiple agents shapes neural representations, we train a recurrent neural network (RNN) model that captures properties of MEC to path integrate trajectories of two agents simultaneously navigating the same environment. We find significant differences between these RNNs and those trained to path integrate only a single agent. At the individual unit level, RNNs trained to path integrate more than one agent develop weaker grid responses, stronger border responses, and tuning for the relative position of the two agents. At the population level, they develop more distributed and robust representations, with changes in network dynamics and manifold topology. Our results provide testable predictions and open new directions with which to study the neural computations supporting spatial navigation.

Finetuning large language models (LLMs) in federated learning (FL) settings has become increasingly important as it allows resource-constrained devices to finetune a model using private data. However, finetuning LLMs using backpropagation requires excessive memory (especially from intermediate activations) for resource-constrained devices. While Forward-mode Auto-Differentiation (AD) can significantly reduce memory footprint from activations, we observe that directly applying it to LLM finetuning results in slow convergence and poor accuracy. In this paper, we introduce Spry, an FL algorithm that splits trainable weights of an LLM among participating clients, such that each client computes gradients using forward-mode AD that are closer estimations of the true gradients. Spry achieves a low memory footprint, high accuracy, and fast convergence. We formally prove that the global gradients in Spry are unbiased estimators of true global gradients for homogeneous data distributions across clients, while heterogeneity increases bias of the estimates. We also derive Spry's convergence rate, showing that the gradients decrease inversely proportional to the number of FL rounds, indicating the convergence up to the limits of heterogeneity. Empirically, Spry reduces the memory footprint during training by 1.4-7.1$\times$ in contrast to backpropagation, while reaching comparable accuracy, across a wide range of language tasks, models, and FL settings. Spry reduces the convergence time by 1.2-20.3$\times$ and achieves 5.2-13.5\% higher accuracy against state-of-the-art zero-order methods. When finetuning Llama2-7B with LoRA, compared to the peak memory consumption of 33.9GB of backpropagation, Spry only consumes 6.2GB of peak memory. For OPT13B, the reduction is from 76.5GB to 10.8GB. Spry makes feasible previously impossible FL deployments on commodity mobile and edge devices. Our source code is available for replication at https://github.com/Astuary/Spry.

Large language models (LLMs) have recently emerged as powerful tools for tackling many language-processing tasks. Despite their success, training and fine-tuning these models is still far too computationally and memory intensive. In this paper, we identify and characterise the important components needed for effective model convergence using gradient descent. In doing so we find that the intermediate activations used to implement backpropagation can be excessively compressed without incurring any degradation in performance. This result leads us to a cheap and memory-efficient algorithm for both fine-tuning and pre-training LLMs. The proposed algorithm simply divides the tokens up into smaller sub-tokens before projecting them onto a fixed 1-dimensional subspace during the forward pass. These features are then coarsely reconstructed during the backward pass to implement the update rules. We confirm the effectiveness of our algorithm as being complimentary to many state-of-the-art PEFT methods on the VTAB-1k fine-tuning benchmark. Furthermore, we outperform QLoRA for fine-tuning LLaMA and show competitive performance against other memory-efficient pre-training methods on the large-scale C4 dataset.

This paper studies the scaling behavior of state-space models (SSMs) and their structured variants, such as Mamba, that have recently arisen in popularity as alternatives to transformer-based neural network architectures. Specifically, we focus on the capability of SSMs to learn features as their network width approaches infinity. Our findings reveal that established scaling rules, such as the Maximal Update Parameterization, fail to support feature learning as these models cannot be represented in the form of Tensor Programs. Additionally, we demonstrate that spectral scaling conditions, shown to be effective for feature learning in a host of other architectures, do not hold the same implications for SSMs. Through a detailed signal propagation analysis in SSMs, both forward and backward, we identify the appropriate scaling necessary for non-trivial feature evolution in the infinite-width limit. Our proposed scaling shows behavior akin to the Maximal Update Parameterization, such as improved stability, better generalization, and transferability of optimal hyper-parameters from small to large scale SSMs.

The poor performance of transformers on arithmetic tasks seems to stem in large part from their inability to keep track of the exact position of each digit inside of a large span of digits. We mend this problem by adding an embedding to each digit that encodes its position relative to the start of the number. In addition to the boost these embeddings provide on their own, we show that this fix enables architectural modifications such as input injection and recurrent layers to improve performance even further. With positions resolved, we can study the logical extrapolation ability of transformers. Can they solve arithmetic problems that are larger and more complex than those in their training data? We find that training on only 20 digit numbers with a single GPU for one day, we can reach state-of-the-art performance, achieving up to 99% accuracy on 100 digit addition problems. Finally, we show that these gains in numeracy also unlock improvements on other multi-step reasoning tasks including sorting and multiplication.