Anand Subramoney

Anand Subramoney, Khaleelulla Khan Nazeer, Mark Schöne et al.

Recurrent neural networks (RNNs) are well suited for solving sequence tasks in resource-constrained systems due to their expressivity and low computational requirements. However, there is still a need to bridge the gap between what RNNs are capable of in terms of efficiency and performance and real-world application requirements. The memory and computational requirements arising from propagating the activations of all the neurons at every time step to every connected neuron, together with the sequential dependence of activations, contribute to the inefficiency of training and using RNNs. We propose a solution inspired by biological neuron dynamics that makes the communication between RNN units sparse and discrete. This makes the backward pass with backpropagation through time (BPTT) computationally sparse and efficient as well. We base our model on the gated recurrent unit (GRU), extending it with units that emit discrete events for communication triggered by a threshold so that no information is communicated to other units in the absence of events. We show theoretically that the communication between units, and hence the computation required for both the forward and backward passes, scales with the number of events in the network. Our model achieves efficiency without compromising task performance, demonstrating competitive performance compared to state-of-the-art recurrent network models in real-world tasks, including language modeling. The dynamic activity sparsity mechanism also makes our model well suited for novel energy-efficient neuromorphic hardware. Code is available at https://github.com/KhaleelKhan/EvNN/.

Anand Subramoney

Backpropagation through time (BPTT) is the standard algorithm for training recurrent neural networks (RNNs), which requires separate simulation phases for the forward and backward passes for inference and learning, respectively. Moreover, BPTT requires storing the complete history of network states between phases, with memory consumption growing proportional to the input sequence length. This makes BPTT unsuited for online learning and presents a challenge for implementation on low-resource real-time systems. Real-Time Recurrent Learning (RTRL) allows online learning, and the growth of required memory is independent of sequence length. However, RTRL suffers from exceptionally high computational costs that grow proportional to the fourth power of the state size, making RTRL computationally intractable for all but the smallest of networks. In this work, we show that recurrent networks exhibiting high activity sparsity can reduce the computational cost of RTRL. Moreover, combining activity and parameter sparsity can lead to significant enough savings in computational and memory costs to make RTRL practical. Unlike previous work, this improvement in the efficiency of RTRL can be achieved without using any approximations for the learning process.

Rishav Mukherji, Mark Schöne, Khaleelulla Khan Nazeer et al.

Artificial neural networks open up unprecedented machine learning capabilities at the cost of ever growing computational requirements. Sparsifying the parameters, often achieved through weight pruning, has been identified as a powerful technique to compress the number of model parameters and reduce the computational operations of neural networks. Yet, sparse activations, while omnipresent in both biological neural networks and deep learning systems, have not been fully utilized as a compression technique in deep learning. Moreover, the interaction between sparse activations and weight pruning is not fully understood. In this work, we demonstrate that activity sparsity can compose multiplicatively with parameter sparsity in a recurrent neural network model based on the GRU that is designed to be activity sparse. We achieve up to $20\times$ reduction of computation while maintaining perplexities below $60$ on the Penn Treebank language modeling task. This magnitude of reduction has not been achieved previously with solely sparsely connected LSTMs, and the language modeling performance of our model has not been achieved previously with any sparsely activated recurrent neural networks or spiking neural networks. Neuromorphic computing devices are especially good at taking advantage of the dynamic activity sparsity, and our results provide strong evidence that making deep learning models activity sparse and porting them to neuromorphic devices can be a viable strategy that does not compromise on task performance. Our results also drive further convergence of methods from deep learning and neuromorphic computing for efficient machine learning.

Reza Sedghi, Robin Schiewer, Anand Subramoney et al.

At typical context lengths, the feed-forward MLP block accounts for a large share of a transformer's compute budget, motivating sparse alternatives to dense MLP blocks. We study sparse, tree-structured feed-forward layers as drop-in replacements for MLP blocks in deep transformer architectures, enabling conditional computation via hard hierarchical routing without a separate router network. We demonstrate for the first time that this form of tree-structured conditional sparsity can be applied for autoregressive language modeling and downstream question answering, including zero- and few-shot settings, and its scalability beyond 1B parameters. Despite activating fewer than 5% of the feed-forward block's units per token, our models match dense baselines under controlled training and fine-tuning protocols. We further analyze training dynamics and identify an emergent auto-pruning effect: the interaction of hard routing with asymmetric nonlinearities progressively deactivates unused paths, yielding partial conversion of dynamic routing into static structural sparsity. We show that simple architectural choices can modulate this behavior and recover balanced trees without auxiliary losses. Overall, our work demonstrates that tree-structured feed-forward layers provide a scalable and controllable mechanism for sparsifying large transformer models.

Lukas König, Manuel Kuhn, David Kappel et al.

Existing frameworks for gradient-based training of spiking neural networks face a trade-off: discrete-time methods using surrogate gradients support arbitrary neuron models but introduce gradient bias and constrain spike-time resolution, while continuous-time methods that compute exact gradients require analytical expressions for spike times and state evolution, restricting them to simple neuron types such as Leaky Integrate and Fire (LIF). We introduce the Eventax framework, which resolves this trade-off by combining differentiable numerical ODE solvers with event-based spike handling. Built in JAX, our frame-work uses Diffrax ODE-solvers to compute gradients that are exact with respect to the forward simulation for any neuron model defined by ODEs . It also provides a simple API where users can specify just the neuron dynamics, spike conditions, and reset rules. Eventax prioritises modelling flexibility, supporting a wide range of neuron models, loss functions, and network architectures, which can be easily extended. We demonstrate Eventax on multiple benchmarks, including Yin-Yang and MNIST, using diverse neuron models such as Leaky Integrate-and-fire (LIF), Quadratic Integrate-and-fire (QIF), Exponential integrate-and-fire (EIF), Izhikevich and Event-based Gated Recurrent Unit (EGRU) with both time-to-first-spike and state-based loss functions, demonstrating its utility for prototyping and testing event-based architectures trained with exact gradients. We also demonstrate the application of this framework for more complex neuron types by implementing a multi-compartment neuron that uses a model of dendritic spikes in human layer 2/3 cortical Pyramidal neurons for computation. Code available at https://github.com/efficient-scalable-machine-learning/eventax.

Jacques Kaiser, Michael Hoff, Andreas Konle et al.

The endeavor to understand the brain involves multiple collaborating research fields. Classically, synaptic plasticity rules derived by theoretical neuroscientists are evaluated in isolation on pattern classification tasks. This contrasts with the biological brain which purpose is to control a body in closed-loop. This paper contributes to bringing the fields of computational neuroscience and robotics closer together by integrating open-source software components from these two fields. The resulting framework allows to evaluate the validity of biologically-plausibe plasticity models in closed-loop robotics environments. We demonstrate this framework to evaluate Synaptic Plasticity with Online REinforcement learning (SPORE), a reward-learning rule based on synaptic sampling, on two visuomotor tasks: reaching and lane following. We show that SPORE is capable of learning to perform policies within the course of simulated hours for both tasks. Provisional parameter explorations indicate that the learning rate and the temperature driving the stochastic processes that govern synaptic learning dynamics need to be regulated for performance improvements to be retained. We conclude by discussing the recent deep reinforcement learning techniques which would be beneficial to increase the functionality of SPORE on visuomotor tasks.

Mark Schöne, Neeraj Mohan Sushma, Jingyue Zhuge et al.

Event-based sensors are well suited for real-time processing due to their fast response times and encoding of the sensory data as successive temporal differences. These and other valuable properties, such as a high dynamic range, are suppressed when the data is converted to a frame-based format. However, most current methods either collapse events into frames or cannot scale up when processing the event data directly event-by-event. In this work, we address the key challenges of scaling up event-by-event modeling of the long event streams emitted by such sensors, which is a particularly relevant problem for neuromorphic computing. While prior methods can process up to a few thousand time steps, our model, based on modern recurrent deep state-space models, scales to event streams of millions of events for both training and inference. We leverage their stable parameterization for learning long-range dependencies, parallelizability along the sequence dimension, and their ability to integrate asynchronous events effectively to scale them up to long event streams. We further augment these with novel event-centric techniques enabling our model to match or beat the state-of-the-art performance on several event stream benchmarks. In the Spiking Speech Commands task, we improve state-of-the-art by a large margin of 7.7% to 88.4%. On the DVS128-Gestures dataset, we achieve competitive results without using frames or convolutional neural networks. Our work demonstrates, for the first time, that it is possible to use fully event-based processing with purely recurrent networks to achieve state-of-the-art task performance in several event-based benchmarks.

Khaleelulla Khan Nazeer, Mark Schöne, Rishav Mukherji et al.

As large language models continue to scale in size rapidly, so too does the computational power required to run them. Event-based networks on neuromorphic devices offer a potential way to reduce energy consumption for inference significantly. However, to date, most event-based networks that can run on neuromorphic hardware, including spiking neural networks (SNNs), have not achieved task performance even on par with LSTM models for language modeling. As a result, language modeling on neuromorphic devices has seemed a distant prospect. In this work, we demonstrate the first-ever implementation of a language model on a neuromorphic device - specifically the SpiNNaker 2 chip - based on a recently published event-based architecture called the EGRU. SpiNNaker 2 is a many-core neuromorphic chip designed for large-scale asynchronous processing, while the EGRU is architected to leverage such hardware efficiently while maintaining competitive task performance. This implementation marks the first time a neuromorphic language model matches LSTMs, setting the stage for taking task performance to the level of large language models. We also demonstrate results on a gesture recognition task based on inputs from a DVS camera. Overall, our results showcase the feasibility of this neuro-inspired neural network in hardware, highlighting significant gains versus conventional hardware in energy efficiency for the common use case of single batch inference.

Neeraj Mohan Sushma, Yudou Tian, Harshvardhan Mestha et al.

Deep state-space models (Deep SSMs) are becoming popular as effective approaches to model sequence data. They have also been shown to be capable of in-context learning, much like transformers. However, a complete picture of how SSMs might be able to do in-context learning has been missing. In this study, we provide a direct and explicit construction to show that state-space models can perform gradient-based learning and use it for in-context learning in much the same way as transformers. Specifically, we prove that a single structured state-space model layer, augmented with multiplicative input and output gating, can reproduce the outputs of an implicit linear model with least squares loss after one step of gradient descent. We then show a straightforward extension to multi-step linear and non-linear regression tasks. We validate our construction by training randomly initialized augmented SSMs on linear and non-linear regression tasks. The empirically obtained parameters through optimization match the ones predicted analytically by the theoretical construction. Overall, we elucidate the role of input- and output-gating in recurrent architectures as the key inductive biases for enabling the expressive power typical of foundation models. We also provide novel insights into the relationship between state-space models and linear self-attention, and their ability to learn in-context.

Rishav Mukherji, Mark Schöne, Khaleelulla Khan Nazeer et al.

Activity and parameter sparsity are two standard methods of making neural networks computationally more efficient. Event-based architectures such as spiking neural networks (SNNs) naturally exhibit activity sparsity, and many methods exist to sparsify their connectivity by pruning weights. While the effect of weight pruning on feed-forward SNNs has been previously studied for computer vision tasks, the effects of pruning for complex sequence tasks like language modeling are less well studied since SNNs have traditionally struggled to achieve meaningful performance on these tasks. Using a recently published SNN-like architecture that works well on small-scale language modeling, we study the effects of weight pruning when combined with activity sparsity. Specifically, we study the trade-off between the multiplicative efficiency gains the combination affords and its effect on task performance for language modeling. To dissect the effects of the two sparsities, we conduct a comparative analysis between densely activated models and sparsely activated event-based models across varying degrees of connectivity sparsity. We demonstrate that sparse activity and sparse connectivity complement each other without a proportional drop in task performance for an event-based neural network trained on the Penn Treebank and WikiText-2 language modeling datasets. Our results suggest sparsely connected event-based neural networks are promising candidates for effective and efficient sequence modeling.

Mark Schöne, Yash Bhisikar, Karan Bania et al. · cmu

Handling sparse and unstructured geometric data, such as point clouds or event-based vision, is a pressing challenge in the field of machine vision. Recently, sequence models such as Transformers and state-space models entered the domain of geometric data. These methods require specialized preprocessing to create a sequential view of a set of points. Furthermore, prior works involving sequence models iterate geometric data with either uniform or learned step sizes, implicitly relying on the model to infer the underlying geometric structure. In this work, we propose to encode geometric structure explicitly into the parameterization of a state-space model. State-space models are based on linear dynamics governed by a one-dimensional variable such as time or a spatial coordinate. We exploit this dynamic variable to inject relative differences of coordinates into the step size of the state-space model. The resulting geometric operation computes interactions between all pairs of N points in O(N) steps. Our model deploys the Mamba selective state-space model with a modified CUDA kernel to efficiently map sparse geometric data to modern hardware. The resulting sequence model, which we call STREAM, achieves competitive results on a range of benchmarks from point-cloud classification to event-based vision and audio classification. STREAM demonstrates a powerful inductive bias for sparse geometric data by improving the PointMamba baseline when trained from scratch on the ModelNet40 and ScanObjectNN point cloud analysis datasets. It further achieves, for the first time, 100% test accuracy on all 11 classes of the DVS128 Gestures dataset.

Jan Rathjens, Robin Schiewer, Laurenz Wiskott et al.

Mamba has attracted widespread interest as a general-purpose sequence model due to its low computational complexity and competitive performance relative to transformers. However, its performance can degrade when inference sequence lengths exceed those seen during training. We study this phenomenon using a controlled vision task in which Mamba reconstructs images from sequences of image patches. By analyzing reconstructions at different stages of sequence processing, we reveal that Mamba qualitatively adapts its behavior to the distribution of sequence lengths encountered during training, resulting in strategies that fail to generalize beyond this range. To support our analysis, we introduce a length-adaptive variant of Mamba that improves performance across training sequence lengths. Our results provide an intuitive perspective on length generalization in Mamba and suggest directions for improving the architecture.

Reza Sedghi, Anand Subramoney, David Kappel

Efficient inference with transformer-based models remains a challenge, especially in vision tasks like object detection. We analyze the inherent sparsity in the MLP layers of DETR and introduce two methods to exploit it without retraining. First, we propose Static Indicator-Based Sparsification (SIBS), a heuristic method that predicts neuron inactivity based on fixed activation patterns. While simple, SIBS offers limited gains due to the input-dependent nature of sparsity. To address this, we introduce Micro-Gated Sparsification (MGS), a lightweight gating mechanism trained on top of a pretrained DETR. MGS predicts dynamic sparsity using a small linear layer and achieves up to 85 to 95% activation sparsity. Experiments on the COCO dataset show that MGS maintains or even improves performance while significantly reducing computation. Our method offers a practical, input-adaptive approach to sparsification, enabling efficient deployment of pretrained vision transformers without full model retraining.

Robin Schiewer, Anand Subramoney, Laurenz Wiskott

Hierarchical model-based reinforcement learning (HMBRL) aims to combine the benefits of better sample efficiency of model based reinforcement learning (MBRL) with the abstraction capability of hierarchical reinforcement learning (HRL) to solve complex tasks efficiently. While HMBRL has great potential, it still lacks wide adoption. In this work we describe a novel HMBRL framework and evaluate it thoroughly. To complement the multi-layered decision making idiom characteristic for HRL, we construct hierarchical world models that simulate environment dynamics at various levels of temporal abstraction. These models are used to train a stack of agents that communicate in a top-down manner by proposing goals to their subordinate agents. A significant focus of this study is the exploration of a static and environment agnostic temporal abstraction, which allows concurrent training of models and agents throughout the hierarchy. Unlike most goal-conditioned H(MB)RL approaches, it also leads to comparatively low dimensional abstract actions. Although our HMBRL approach did not outperform traditional methods in terms of final episode returns, it successfully facilitated decision making across two levels of abstraction using compact, low dimensional abstract actions. A central challenge in enhancing our method's performance, as uncovered through comprehensive experimentation, is model exploitation on the abstract level of our world model stack. We provide an in depth examination of this issue, discussing its implications for the field and suggesting directions for future research to overcome this challenge. By sharing these findings, we aim to contribute to the broader discourse on refining HMBRL methodologies and to assist in the development of more effective autonomous learning systems for complex decision-making environments.

David Kappel, Khaleelulla Khan Nazeer, Cabrel Teguemne Fokam et al.

The ubiquitous backpropagation algorithm requires sequential updates through the network introducing a locking problem. In addition, back-propagation relies on the transpose of forward weight matrices to compute updates, introducing a weight transport problem across the network. Locking and weight transport are problems because they prevent efficient parallelization and horizontal scaling of the training process. We propose a new method to address both these problems and scale up the training of large models. Our method works by dividing a deep neural network into blocks and introduces a feedback network that propagates the information from the targets backwards to provide auxiliary local losses. Forward and backward propagation can operate in parallel and with different sets of weights, addressing the problems of locking and weight transport. Our approach derives from a statistical interpretation of training that treats output activations of network blocks as parameters of probability distributions. The resulting learning framework uses these parameters to evaluate the agreement between forward and backward information. Error backpropagation is then performed locally within each block, leading to "block-local" learning. Several previously proposed alternatives to error backpropagation emerge as special cases of our model. We present results on a variety of tasks and architectures, demonstrating state-of-the-art performance using block-local learning. These results provide a new principled framework for training networks in a distributed setting.

Alper Yegenoglu, Anand Subramoney, Thorsten Hater et al.

Neuroscience models commonly have a high number of degrees of freedom and only specific regions within the parameter space are able to produce dynamics of interest. This makes the development of tools and strategies to efficiently find these regions of high importance to advance brain research. Exploring the high dimensional parameter space using numerical simulations has been a frequently used technique in the last years in many areas of computational neuroscience. High performance computing (HPC) can provide today a powerful infrastructure to speed up explorations and increase our general understanding of the model's behavior in reasonable times.

Anand Subramoney, Franz Scherr, Wolfgang Maass

We consider reservoirs in the form of liquid state machines, i.e., recurrently connected networks of spiking neurons with randomly chosen weights. So far only the weights of a linear readout were adapted for a specific task. We wondered whether the performance of liquid state machines can be improved if the recurrent weights are chosen with a purpose, rather than randomly. After all, weights of recurrent connections in the brain are also not assumed to be randomly chosen. Rather, these weights were probably optimized during evolution, development, and prior learning experiences for specific task domains. In order to examine the benefits of choosing recurrent weights within a liquid with a purpose, we applied the Learning-to-Learn (L2L) paradigm to our model: We optimized the weights of the recurrent connections -- and hence the dynamics of the liquid state machine -- for a large family of potential learning tasks, which the network might have to learn later through modification of the weights of readout neurons. We found that this two-tiered process substantially improves the learning speed of liquid state machines for specific tasks. In fact, this learning speed increases further if one does not train the weights of linear readouts at all, and relies instead on the internal dynamics and fading memory of the network for remembering salient information that it could extract from preceding examples for the current learning task. This second type of learning has recently been proposed to underlie fast learning in the prefrontal cortex and motor cortex, and hence it is of interest to explore its performance also in models. Since liquid state machines share many properties with other types of reservoirs, our results raise the question whether L2L conveys similar benefits also to these other reservoirs.

Guillaume Bellec, Darjan Salaj, Anand Subramoney et al.

Recurrent networks of spiking neurons (RSNNs) underlie the astounding computing and learning capabilities of the brain. But computing and learning capabilities of RSNN models have remained poor, at least in comparison with artificial neural networks (ANNs). We address two possible reasons for that. One is that RSNNs in the brain are not randomly connected or designed according to simple rules, and they do not start learning as a tabula rasa network. Rather, RSNNs in the brain were optimized for their tasks through evolution, development, and prior experience. Details of these optimization processes are largely unknown. But their functional contribution can be approximated through powerful optimization methods, such as backpropagation through time (BPTT). A second major mismatch between RSNNs in the brain and models is that the latter only show a small fraction of the dynamics of neurons and synapses in the brain. We include neurons in our RSNN model that reproduce one prominent dynamical process of biological neurons that takes place at the behaviourally relevant time scale of seconds: neuronal adaptation. We denote these networks as LSNNs because of their Long short-term memory. The inclusion of adapting neurons drastically increases the computing and learning capability of RSNNs if they are trained and configured by deep learning (BPTT combined with a rewiring algorithm that optimizes the network architecture). In fact, the computational performance of these RSNNs approaches for the first time that of LSTM networks. In addition RSNNs with adapting neurons can acquire abstract knowledge from prior learning in a Learning-to-Learn (L2L) scheme, and transfer that knowledge in order to learn new but related tasks from very few examples. We demonstrate this for supervised learning and reinforcement learning.

Mihai A. Petrovici, Sebastian Schmitt, Johann Klähn et al.

Despite being originally inspired by the central nervous system, artificial neural networks have diverged from their biological archetypes as they have been remodeled to fit particular tasks. In this paper, we review several possibilites to reverse map these architectures to biologically more realistic spiking networks with the aim of emulating them on fast, low-power neuromorphic hardware. Since many of these devices employ analog components, which cannot be perfectly controlled, finding ways to compensate for the resulting effects represents a key challenge. Here, we discuss three different strategies to address this problem: the addition of auxiliary network components for stabilizing activity, the utilization of inherently robust architectures and a training method for hardware-emulated networks that functions without perfect knowledge of the system's dynamics and parameters. For all three scenarios, we corroborate our theoretical considerations with experimental results on accelerated analog neuromorphic platforms.