Neuroscience and Artificial Intelligence (AI) have made impressive progress in recent years but remain only loosely interconnected. Based on a workshop convened by the National Science Foundation in August 2025, we identify three fundamental capability gaps in current AI: the inability to interact with the physical world, inadequate learning that produces brittle systems, and unsustainable energy and data inefficiency. We describe the neuroscience principles that address each: co-design of body and controller, prediction through interaction, multi-scale learning with neuromodulatory control, hierarchical distributed architectures, and sparse event-driven computation. We present a research roadmap organized around these principles at near, mid, and long-term horizons. We argue that realizing this program requires a new generation of researchers trained across the boundary between neuroscience and engineering, and describe the institutional conditions: interdisciplinary training, hardware access, community standards, and ethics, needed to support them. We conclude that NeuroAI, neuroscience-informed artificial intelligence, has the potential to overcome limitations of current AI while deepening our understanding of biological neural computation.

This work introduces a novel generative continual learning framework based on self-organizing maps (SOMs), a brain-inspired natural computing model, extended with learned distributional statistics and encoder–decoder models for class incremental continual learning. These extended SOMs enable exemplar-free replay with fixed-capacity statistical memory, eliminating the need to store raw data samples. For high-dimensional input spaces, the SOM operates over the latent space of the encoder–decoder, while for lower-dimensional inputs, the SOM operates in a standalone fashion. Our method stores a running mean, variance, and covariance for each SOM unit, from which synthetic samples are then generated during future learning iterations. For the encoder–decoder method, generated samples are fed through the decoder to be used in subsequent replay. Experimental results on standard class-incremental benchmarks show that our approach performs competitively with state-of-the-art memory-based methods and outperforms memory-free methods, notably improving over the best state-of-the-art single-class incremental performance without pretrained encoders on CIFAR-10 and CIFAR-100 by nearly 10

In this article, we review a class of neuro-mimetic computational models that we place under the label of spiking predictive coding. Specifically, we review the general framework of predictive processing in the context of neurons that emit discrete action potentials, i.e., spikes. Theoretically, we structure our survey around how prediction errors are represented. This results in an organization into three broad classes of approaches: prediction errors in explicit groups of error neurons, in membrane potentials, and implicit prediction error encoding. Furthermore, we examine some applications of spiking predictive coding that utilize more energy-efficient, edge-computing hardware platforms. Finally, we highlight important future directions and challenges in this emerging line of inquiry in brain-inspired computing. Building on the prior results of work in computational cognitive neuroscience, machine intelligence, and neuromorphic engineering, we hope that this review of neuromorphic formulations and implementations of predictive coding will encourage and guide future research and development in this emerging research area.

Self-supervised learning has become an increasingly important paradigm in the domain of machine intelligence. Furthermore, evidence for self-supervised adaptation, such as contrastive formulations, has emerged in recent computational neuroscience and brain-inspired research. Nevertheless, current work on self-supervised learning relies on biologically implausible credit assignment -- in the form of backpropagation of errors -- and feedforward inference, typically a forward-locked pass. Predictive coding, in its mechanistic form, offers a biologically plausible means to sidestep these backprop-specific limitations. However, unsupervised predictive coding rests on learning a generative model of raw pixel input (akin to ``generative AI'' approaches), which entails predicting a potentially high dimensional input; on the other hand, supervised predictive coding, which learns a mapping between inputs to target labels, requires human annotation, and thus incurs the drawbacks of supervised learning. In this work, we present a scheme for self-supervised learning within a neurobiologically plausible framework that appeals to the free energy principle, constructing a new form of predictive coding that we call meta-representational predictive coding (MPC). MPC sidesteps the need for learning a generative model of sensory input (e.g., pixel-level features) by learning to predict representations of sensory input across parallel streams, resulting in an encoder-only learning and inference scheme. This formulation rests on active inference (in the form of sensory glimpsing) to drive the learning of representations, i.e., the representational dynamics are driven by sequences of decisions made by the model to sample informative portions of its sensorium.

Although research has produced promising results demonstrating the utility of active inference (AIF) in Markov decision processes (MDPs), there is relatively less work that builds AIF models in the context of environments and problems that take the form of partially observable Markov decision processes (POMDPs). In POMDP scenarios, the agent must infer the unobserved environmental state from raw sensory observations, e.g., pixels in an image. Additionally, less work exists in examining the most difficult form of POMDP-centered control: continuous action space POMDPs under sparse reward signals. In this work, we address issues facing the AIF modeling paradigm by introducing novel prior preference learning techniques and self-revision schedules to help the agent excel in sparse-reward, continuous action, goal-based robotic control POMDP environments. Empirically, we show that our agents offer improved performance over state-of-the-art models in terms of cumulative rewards, relative stability, and success rate.

Parameter prediction is essential for many applications, facilitating insightful interpretation and decision-making. However, in many real life domains, such as power systems, medicine, and engineering, it can be very expensive to acquire ground truth labels for certain datasets as they may require extensive and expensive laboratory testing. In this work, we introduce a semi-supervised learning approach based on topological projections in self-organizing maps (SOMs), which significantly reduces the required number of labeled data points to perform parameter prediction, effectively exploiting information contained in large unlabeled datasets. While few-shot learning has seen significant advances in recent years, the majority of existing approaches focus on classification tasks, making our regression-based method particularly novel for continuous parameter estimation problems. Our proposed method first trains SOMs on unlabeled data followed by only using a minimal number of available labeled data points to assign targets to key best matching units (BMU). The values estimated for newly-encountered data points are computed utilizing the average of the N closest labeled data points in the SOM's U-matrix in tandem with a topological shortest path distance calculation scheme. The effectiveness of our approach has been validated through practical application in power engineering, specifically for estimating critical coal property values in coal power plants, where traditional laboratory testing is both time-consuming and costly. Our results indicate that the proposed minimally supervised model significantly outperforms traditional regression techniques, including linear and polynomial regression, Gaussian process regression, K-nearest neighbors, as well as deep neural network models and related clustering schemes.

Synaptic delays play a crucial role in biological neuronal networks, where their modulation has been observed in mammalian learning processes. In the realm of neuromorphic computing, although spiking neural networks (SNNs) aim to emulate biology more closely than traditional artificial neural networks do, synaptic delays are rarely incorporated into their simulation. We introduce a novel learning rule for simultaneously learning synaptic connection strengths and delays, by extending spike-timing dependent plasticity (STDP), a Hebbian method commonly used for learning synaptic weights. We validate our approach by extending a widely-used SNN model for classification trained with unsupervised learning. Then we demonstrate the effectiveness of our new method by comparing it against another existing methods for co-learning synaptic weights and delays as well as against STDP without synaptic delays. Results demonstrate that our proposed method consistently achieves superior performance across a variety of test scenarios. Furthermore, our experimental results yield insight into the interplay between synaptic efficacy and delay.

An intelligent system capable of continual learning is one that can process and extract knowledge from potentially infinitely long streams of pattern vectors. The major challenge that makes crafting such a system difficult is known as catastrophic forgetting - an agent, such as one based on artificial neural networks (ANNs), struggles to retain previously acquired knowledge when learning from new samples. Furthermore, ensuring that knowledge is preserved for previous tasks becomes more challenging when input is not supplemented with task boundary information. Although forgetting in the context of ANNs has been studied extensively, there still exists far less work investigating it in terms of unsupervised architectures such as the venerable self-organizing map (SOM), a neural model often used in clustering and dimensionality reduction. While the internal mechanisms of SOMs could, in principle, yield sparse representations that improve memory retention, we observe that, when a fixed-size SOM processes continuous data streams, it experiences concept drift. In light of this, we propose a generalization of the SOM, the continual SOM (CSOM), which is capable of online unsupervised learning under a low memory budget. Our results, on benchmarks including MNIST, Kuzushiji-MNIST, and Fashion-MNIST, show almost a two times increase in accuracy, and CIFAR-10 demonstrates a state-of-the-art result when tested on (online) unsupervised class incremental learning setting.

This work introduces a novel, nature-inspired neural architecture search (NAS) algorithm based on ant colony optimization, Continuous Ant-based Neural Topology Search (CANTS), which utilizes synthetic ants that move over a continuous search space based on the density and distribution of pheromones, strongly inspired by how ants move in the real world. The paths taken by the ant agents through the search space are utilized to construct artificial neural networks (ANNs). This continuous search space allows CANTS to automate the design of ANNs of any size, removing a key limitation inherent to many current NAS algorithms that must operate within structures of a size predetermined by the user. CANTS employs a distributed asynchronous strategy which allows it to scale to large-scale high performance computing resources, works with a variety of recurrent memory cell structures, and uses of a communal weight sharing strategy to reduce training time. The proposed procedure is evaluated on three real-world, time series prediction problems in the field of power systems and compared to two state-of-the-art algorithms. Results show that CANTS is able to provide improved or competitive results on all of these problems while also being easier to use, requiring half the number of user-specified hyper-parameters.

In this work, we propose time-integrated spike-timing-dependent plasticity (TI-STDP), a mathematical model of synaptic plasticity that allows spiking neural networks to continuously adapt to sensory input streams in an unsupervised fashion. Notably, we theoretically establish and formally prove key properties related to the synaptic adjustment mechanics that underwrite TI-STDP. Empirically, we demonstrate the efficacy of TI-STDP in simulations of jointly learning deeper spiking neural networks that process input digit pixel patterns, at both full image and patch-levels, comparing to two powerful historical instantations of STDP; trace-based STDP (TR-STDP) and event-based post-synaptic STDP (EV-STDP). Usefully, we demonstrate that not only are all forms of STDP capable of meaningfully adapting the synaptic efficacies of a multi-layer biophysical architecture, but that TI-STDP is notably able to do so without requiring the tracking of a large window of pre- and post-synaptic spike timings, the maintenance of additional parameterized traces, or the restriction of synaptic plasticity changes to occur within very narrow windows of time. This means that our findings show that TI-STDP can efficiently embody the benefits of models such as canonical STDP, TR-STDP, and EV-STDP without their costs or drawbacks. Usefully, our results further demonstrate the promise of using a spike-correlation scheme such as TI-STDP in conducting credit assignment in discrete pulse-based neuromorphic models, particularly those than acquire a lower-level distributed representation jointly with an upper-level, more abstract representation that self-organizes to cluster based on inherent cross-pattern similarities. We further demonstrate TI-STDP's effectiveness in adapting a simple neuronal circuit that learns a simple bi-level, part-whole hierarchy from sensory input patterns.

Energy-based learning algorithms, such as predictive coding (PC), have garnered significant attention in the machine learning community due to their theoretical properties, such as local operations and biologically plausible mechanisms for error correction. In this work, we rigorously analyze the stability, robustness, and convergence of PC through the lens of dynamical systems theory. We show that, first, PC is Lyapunov stable under mild assumptions on its loss and residual energy functions, which implies intrinsic robustness to small random perturbations due to its well-defined energy-minimizing dynamics. Second, we formally establish that the PC updates approximate quasi-Newton methods by incorporating higher-order curvature information, which makes them more stable and able to converge with fewer iterations compared to models trained via backpropagation (BP). Furthermore, using this dynamical framework, we provide new theoretical bounds on the similarity between PC and other algorithms, i.e., BP and target propagation (TP), by precisely characterizing the role of higher-order derivatives. These bounds, derived through detailed analysis of the Hessian structures, show that PC is significantly closer to quasi-Newton updates than TP, providing a deeper understanding of the stability and efficiency of PC compared to conventional learning methods.

Eye image segmentation is a critical step in eye tracking that has great influence over the final gaze estimate. Segmentation models trained using supervised machine learning can excel at this task, their effectiveness is determined by the degree of overlap between the narrow distributions of image properties defined by the target dataset and highly specific training datasets, of which there are few. Attempts to broaden the distribution of existing eye image datasets through the inclusion of synthetic eye images have found that a model trained on synthetic images will often fail to generalize back to real-world eye images. In remedy, we use dimensionality-reduction techniques to measure the overlap between the target eye images and synthetic training data, and to prune the training dataset in a manner that maximizes distribution overlap. We demonstrate that our methods result in robust, improved performance when tackling the discrepancy between simulation and real-world data samples.

While current deep learning algorithms have been successful for a wide variety of artificial intelligence (AI) tasks, including those involving structured image data, they present deep neurophysiological conceptual issues due to their reliance on the gradients that are computed by backpropagation of errors (backprop). Gradients are required to obtain synaptic weight adjustments but require knowledge of feed-forward activities in order to conduct backward propagation, a biologically implausible process. This is known as the "weight transport problem". Therefore, in this work, we present a more biologically plausible approach towards solving the weight transport problem for image data. This approach, which we name the error kernel driven activation alignment (EKDAA) algorithm, accomplishes through the introduction of locally derived error transmission kernels and error maps. Like standard deep learning networks, EKDAA performs the standard forward process via weights and activation functions; however, its backward error computation involves adaptive error kernels that propagate local error signals through the network. The efficacy of EKDAA is demonstrated by performing visual-recognition tasks on the Fashion MNIST, CIFAR-10 and SVHN benchmarks, along with demonstrating its ability to extract visual features from natural color images. Furthermore, in order to demonstrate its non-reliance on gradient computations, results are presented for an EKDAA trained CNN that employs a non-differentiable activation function.

Brain-inspired machine intelligence research seeks to develop computational models that emulate the information processing and adaptability that distinguishes biological systems of neurons. This has led to the development of spiking neural networks, a class of models that promisingly addresses the biological implausibility and the lack of energy efficiency inherent to modern-day deep neural networks. In this work, we address the challenge of designing neurobiologically motivated schemes for adjusting the synapses of spiking networks and propose contrastive signal-dependent plasticity, a process which generalizes ideas behind self-supervised learning to facilitate local adaptation in architectures of event-based neuronal layers that operate in parallel. Our experimental simulations demonstrate a consistent advantage over other biologically plausible approaches when training recurrent spiking networks, crucially side-stepping the need for extra structure such as feedback synapses.

Spiking neural networks have become an important family of neuron-based models that sidestep many of the key limitations facing modern-day backpropagation-trained deep networks, including their high energy inefficiency and long-criticized biological implausibility. In this work, we design and investigate a proof-of-concept instantiation of contrastive-signal-dependent plasticity (CSDP), a neuromorphic form of forward-forward-based, backpropagation-free learning. Our experimental simulations demonstrate that a hardware implementation of CSDP is capable of learning simple logic functions without the need to resort to complex gradient calculations.

Multi-armed bandits and contextual multi-armed bandits have demonstrated their proficiency in a variety of application areas. However, these models are highly susceptible to volatility and often exhibit knowledge gaps due to a limited understanding of future states. In this paper, we propose a new bandit framework for what we refer to as informed contextual multi armed bandits (iCMABs) to mitigate these gaps, facilitating "informed" decisions based on predicted future contexts. The performance of an iCMAB is thus highly dependent on the accuracy of the forecast it uses. We examine the use of recurrent neural networks (RNNs) evolved through the EXAMM neuroevolution algorithm as compared to other time series forecasting (TSF) methods and evaluate our iCMAB framework's ability to make stock market trading decisions for the Dow-Jones Index (DJI) in comparison to other decision making strategies using these forecasts. Our results demonstrate that an iCMAB, driven by evolved RNN architectures, performs better than statistical TSF methods, fixed architecture RNNs for TSF, and other CMAB methods. Using evolved RNNs, iCMAB is able to achieve the highest return of over 21%, a similar to 7% improvement over not incorporating forecasted values, and a similar to 5% improvement over DJI's return for that time period.

One major criticism of deep learning centers around the biological implausibility of the credit assignment schema used for learning -- backpropagation of errors. This implausibility translates into practical limitations, spanning scientific fields, including incompatibility with hardware and non-differentiable implementations, thus leading to expensive energy requirements. In contrast, biologically plausible credit assignment is compatible with practically any learning condition and is energy-efficient. As a result, it accommodates hardware and scientific modeling, e.g. learning with physical systems and non-differentiable behavior. Furthermore, it can lead to the development of real-time, adaptive neuromorphic processing systems. In addressing this problem, an interdisciplinary branch of artificial intelligence research that lies at the intersection of neuroscience, cognitive science, and machine learning has emerged. In this paper, we survey several vital algorithms that model bio-plausible rules of credit assignment in artificial neural networks, discussing the solutions they provide for different scientific fields as well as their advantages on CPUs, GPUs, and novel implementations of neuromorphic hardware. We conclude by discussing the future challenges that will need to be addressed in order to make such algorithms more useful in practical applications.

We present the COGnitive Neural GENerative system (CogNGen), a cognitive architecture that combines two neurobiologically-plausible, computational models: predictive processing and hyperdimensional/vector-symbolic models. We draw inspiration from architectures such as ACT-R and Spaun/Nengo. CogNGen is in broad agreement with these, providing a level of detail between ACT-R's high-level symbolic description of human cognition and Spaun's low-level neurobiological description, furthermore creating the groundwork for designing agents that learn continually from diverse tasks and model human performance at larger scales than what is possible with current systems. We test CogNGen on four maze-learning tasks, including those that test memory and planning, and find that CogNGen matches performance of deep reinforcement learning models and exceeds on a task designed to test memory.

Training deep neural networks on large-scale datasets requires significant hardware resources whose costs (even on cloud platforms) put them out of reach of smaller organizations, groups, and individuals. Backpropagation (backprop), the workhorse for training these networks, is an inherently sequential process that is difficult to parallelize. Furthermore, researchers must continually develop various specialized techniques, such as particular weight initializations and enhanced activation functions, to ensure stable parameter optimization. Our goal is to seek an effective, neuro-biologically plausible alternative to backprop that can be used to train deep networks. In this paper, we propose a backprop-free procedure, recursive local representation alignment, for training large-scale architectures. Experiments with residual networks on CIFAR-10 and the large benchmark, ImageNet, show that our algorithm generalizes as well as backprop while converging sooner due to weight updates that are parallelizable and computationally less demanding. This is empirical evidence that a backprop-free algorithm can scale up to larger datasets.

In this work, we develop convolutional neural generative coding (Conv-NGC), a generalization of predictive coding to the case of convolution/deconvolution-based computation. Specifically, we concretely implement a flexible neurobiologically-motivated algorithm that progressively refines latent state feature maps in order to dynamically form a more accurate internal representation/reconstruction model of natural images. The performance of the resulting sensory processing system is evaluated on complex datasets such as Color-MNIST, CIFAR-10, and Street House View Numbers (SVHN). We study the effectiveness of our brain-inspired model on the tasks of reconstruction and image denoising and find that it is competitive with convolutional auto-encoding systems trained by backpropagation of errors and outperforms them with respect to out-of-distribution reconstruction (including the full 90k CINIC-10 test set).