Brain-Like Neural Networks Overview

• BLNNs are biologically inspired neural systems that emulate key brain features such as local learning, sparse connectivity, and modular organization.
• They integrate diverse methods including spiking dynamics, Hebbian updates, and probabilistic learning to approximate neural activity and cognitive functions.
• BLNN research supports practical applications through scalable hardware implementations and validation against neural data.

Searching arXiv for papers on brain-like neural networks and related brain-inspired learning to support the article. Brain-Like Neural Networks (BLNNs) are a heterogeneous class of artificial neural systems that draw explicit inspiration from biological brains at one or more levels of analysis, including network architecture, learning rules, neural coding, dynamical regimes, hardware substrate, or empirical correspondence to brain activity and behavior. The term does not denote a single canonical model family. In contemporary usage, it ranges from biologically motivated training principles for conventional artificial neural networks, through spiking and neuromorphic systems, to graph-structured and dynamical models that use neural-network machinery to estimate functional brain organization, and to deep models evaluated by their emergence of brain-like perceptual or neural properties rather than by benchmark accuracy alone. Across this literature, BLNN research is unified less by one architecture than by a recurring set of objectives: local or biologically plausible credit assignment, sparse and modular computation, structured interactions among subsystems, continual adaptation, and alignment with neural or cognitive phenomena (Xiao et al., 2019, Maass, 2023, Rajesh et al., 2024).

1. Conceptual scope and definitional boundaries

BLNNs are defined in multiple, non-equivalent ways across the literature. One broad conceptual framing treats artificial neural networks as computational abstractions of biological neural systems, emphasizing neurons as processing units, synapses as adjustable couplings, thresholded or nonlinear activation, distributed parallelism, and learning through weight change (Nwadiugwu, 2020). In this view, even conventional ANNs are brain-inspired at a coarse level, but stronger forms of brain-likeness require importing additional biological principles such as local learning, feedback, dendritic or temporal computation, sparse activity, or continual adaptation (Etter, 6 Nov 2025).

A narrower and more technical usage reserves BLNNs for models that are constrained by cortical or neuromorphic principles more tightly than mainstream deep networks. In this sense, BLNNs are associated with sparse connectivity, modular cortical-style organization, local Hebbian or Bayesian plasticity, divisive normalization, unsupervised or semi-supervised learning, and deployment on neuromorphic or edge-efficient substrates. The Bayesian Confidence Propagation Neural Network (BCPNN) is treated as a canonical example because it combines hypercolumns and minicolumns, sparse distributed coding, local probability traces, and probabilistic weight semantics (Hafiz et al., 23 Jun 2025, Podobas et al., 2021, Ravichandran et al., 2024).

A third usage is correspondence-based rather than mechanistic. Here a network is considered brain-like because its internal representations or emergent properties align with neural recordings, perceptual effects, or brain organization. This includes visual CNNs benchmarked against Brain-Score and then adapted to model cortical prosthetic phosphenes (Granley et al., 2022), deep networks evaluated on perceptual and neural emergent properties such as mirror confusion, scene incongruence, and global advantage (Rajesh et al., 2024), and vision-LLMs whose neuron-level temporal features predict fMRI responses across multiple brain systems (Ren et al., 19 Oct 2025).

The literature also marks an important boundary: not every use of a neural network in neuroscience is itself a brain-like architecture. For example, ANN-based estimation of functional connectivity graphs from fMRI can be highly relevant to BLNN research on cognition and brain organization while still using a simple linear two-layer estimator rather than a biologically realistic neural design (Kivilcim et al., 2018). This suggests a useful distinction between BLNNs as brain-inspired computational systems and neural-network tools for modeling brain data.

2. Major design principles

Several recurring principles organize the BLNN literature, even though different papers instantiate different subsets.

One major principle is local learning. Standard backpropagation is repeatedly criticized as biologically implausible because it depends on exact loss derivatives, nonlocal information, and weight-transport-like assumptions. Alternatives include generalized likelihood ratio learning from scalar loss values with injected neuronal noise (Xiao et al., 2019), online unsupervised similarity-matching algorithms with local Hebbian and anti-Hebbian updates (Pehlevan et al., 2019), and neuromorphic learning rules derived as sparse approximations to Real-Time Recurrent Learning in spiking recurrent systems (Zenke et al., 2020). A consistent theme is that a synapse should update from variables locally available at that synapse, possibly modulated by a global or targeted evaluative signal.

A second principle is modularity and structured competition. Cortical microcircuit-inspired accounts emphasize repeated local modules, neuron-type diversity, selective routing, and segregation plus integration rather than homogeneous dense networks (Maass, 2023). In BCPNN, this appears as hypercolumns containing competing minicolumns with soft winner-take-all normalization (Hafiz et al., 23 Jun 2025, Ravichandran et al., 2024). In feedforward unsupervised BCPNN, hidden representations are sparse at the level of individual inputs while unit usage remains broadly distributed over the dataset, producing low-entropy within-hypercolumn activity alongside relatively uniform marginal utilization (Ravichandran et al., 2024).

A third principle is sparse and patchy connectivity. The idea that brain-like systems should not be fully dense appears in several forms: sparse local learning rules in BCPNN (Hafiz et al., 23 Jun 2025), structural plasticity and patchy feedforward connectivity in unsupervised feedforward BCPNN (Ravichandran et al., 2024), sparse temporal event coding in neuromorphic systems (Maass, 2023), and the broader claim that edge-efficient BLNNs benefit from modular sparse organization rather than overparameterized dense architectures (Hafiz et al., 23 Jun 2025).

A fourth principle is temporal and dynamical computation. Spiking neural networks are framed as central to brain-like AI because they process information through sparse event-driven temporal signals rather than static activations (Sima et al., 2023). More generally, BLNNs often model cognition as unfolding dynamics rather than static mappings. BrainDyn, for instance, combines node-local LSTM state, sheaf-based message passing, and a neural ODE to generate continuous-time brain-like activity on structured graphs (Viswanath et al., 19 May 2026). The same emphasis appears in neuromorphic substrates, where temporal data streams and online learning are treated as first-class requirements rather than secondary capabilities (Zenke et al., 2020).

A fifth principle is embodied or substrate-level adaptation. BLNN work is not limited to software models. In iontronic networks, “weights” become physical variables such as channel geometry, pressure, and concentration; the forward pass is physical ionic relaxation; and learning combines physical evolution with numerical steepest descent (Conte et al., 6 Nov 2025). This suggests a broader interpretation of brain-likeness in which memory and adaptation are stored in material state rather than abstract tensors.

3. Learning rules and plasticity mechanisms

The strongest technical divide in BLNN research concerns learning. Mainstream deep learning relies on backpropagation, whereas BLNN work explores biologically motivated alternatives.

One family of methods retains gradient-based optimization while making credit assignment more local or physically plausible. In neuromorphic substrates, a direct connection is drawn between RTRL and biologically plausible spiking-network learning. The recurrent Jacobian is decomposed into implicit local recurrence from membrane, synaptic, or compartmental dynamics and explicit recurrence through network couplings; dropping the explicit nonlocal term yields online local eligibility traces with modulatory learning signals (Zenke et al., 2020). This produces rules of the form

$$\frac{\partial \mathcal{L}^t}{\partial W} = \frac{\partial \mathcal{L}^t}{\partial S^t}\,\sigma'(U^t)\,Q_{\mathrm{in}}^t,$$

which are naturally interpretable as three-factor rules involving a learning signal, postsynaptic sensitivity, and presynaptic trace (Zenke et al., 2020).

A second family abandons direct differentiation of activations or losses. The generalized likelihood ratio method injects neuron-level noise and derives an unbiased estimator for the gradient of expected loss that depends on the scalar loss value and the score function of the noise density (Xiao et al., 2019). For threshold networks, the core identity is

$$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$

allowing training with discontinuous threshold activations and discontinuous $0$–$1$ losses that standard backpropagation cannot handle directly (Xiao et al., 2019). The resulting updates use a scalar evaluative signal and local perturbation/activity terms, which the authors interpret as more brain-like.

A third family uses normative objective functions specifically chosen so that online local learning emerges naturally. In similarity-based unsupervised learning, the objective is to match output similarities to input similarities: $$\min_{\mathbf{y}_1,\ldots,\mathbf{y}_T} \frac{1}{T^2} \sum_{t=1}^T\sum_{t'=1}^T \left( \mathbf{x}_t^\top \mathbf{x}_{t'} - \mathbf{y}_t^\top \mathbf{y}_{t'} \right)^2.$$ A variable-substitution trick converts this into a minimax problem whose dynamics are implemented by recurrent neural activity and local Hebbian/anti-Hebbian plasticity (Pehlevan et al., 2019). The resulting updates,

$$W_{ij} \leftarrow W_{ij} + \eta \left(y_i x_j - W_{ij}\right),\qquad M_{ij} \leftarrow M_{ij} + \frac{\eta}{2}\left(y_i y_j - M_{ij}\right),$$

yield biologically plausible networks for principal-subspace learning, whitening, sparse and nonnegative feature extraction, clustering, and manifold learning (Pehlevan et al., 2019).

A fourth family is explicitly Bayesian-Hebbian. In BCPNN, synaptic weights and biases encode marginal and joint probabilities: $$b_j = \log p_j,\qquad w_{ij} = \log\left(\frac{p_{ij}}{p_i \cdot p_j}\right),$$ so a synapse measures a pointwise-mutual-information-like dependency rather than storing an arbitrary learned coefficient (Hafiz et al., 23 Jun 2025). In feedforward unsupervised BCPNN, online traces $P_{im}$, $P_{jk}$, and $P_{imjk}$ are updated by exponential moving averages, and weights are recomputed as log-ratios of coactivation versus independence (Ravichandran et al., 2024). This gives the model probabilistic interpretability together with strictly local synaptic learning.

Finally, some BLNN proposals split learning across functional subsystems. A “brain version of backpropagation” called Layer-Wise Backpropagation replaces end-to-end gradient propagation with local module-wise supervision, using shortcut pathways and module-local losses; paired with an Engram Autoencoder, it is presented as a cortex-like learning plus hippocampus-like sparse memory framework (Mao, 2020). This suggests a broader trend: BLNN research often decomposes learning into local adaptation, sparse indexing, and system-level specialization rather than seeking a single universal algorithm.

4. Architectural families

BLNN architectures span several distinct families whose similarities are principled rather than formal.

Spiking and neuromorphic systems

Spiking neural networks are repeatedly presented as one of the cores of brain-like artificial intelligence because they use pulse-based communication, event-driven computation, and natural temporal processing (Sima et al., 2023). Surveyed neuron models include Hodgkin–Huxley, IF, LIF, SRM, and Izhikevich, with the LIF model described as the most common compromise between realism and simplicity (Sima et al., 2023). Topologies include feedforward, recurrent, recursive, evolutionary, and hybrid SNNs, though practical systems still rely heavily on feedforward networks (Sima et al., 2023).

Neuromorphic-hardware-oriented work argues that merely replacing artificial neurons by spiking neurons is insufficient. Genuine brain-like functionality, on this view, requires richer cortical microcircuit principles: many neuron types, sparse temporal coding, segregation and integration, and diverse local plasticity rules (Maass, 2023). This critique shifts BLNN emphasis from the spiking signal format alone to microcircuit organization and learning.

BCPNN and cortical modular models

BCPNN constitutes a major BLNN lineage. It organizes hidden layers into hypercolumns and minicolumns, uses soft winner-take-all normalization, and learns through local probabilities and joint statistics rather than error gradients (Hafiz et al., 23 Jun 2025, Podobas et al., 2021, Ravichandran et al., 2024). Feedforward BCPNN has been extended to unsupervised representation learning by adding structural plasticity and sparse patchy connectivity (Ravichandran et al., 2024). It has also been made practical on heterogeneous hardware through StreamBrain, which maps the model’s dominant batched outer-product and matrix operations to CPUs, GPUs, FPGAs, and clusters (Podobas et al., 2021), and to embedded FPGA edge deployment with online-learning and inference-only kernels (Hafiz et al., 23 Jun 2025).

Structured dynamical graph models

A different BLNN family models neural systems as structured dynamical graphs. BrainDyn is representative: node-local recent histories are encoded by LSTMs into stalks, projected through edge-specific restriction maps into shared spaces, compared through a sheaf Laplacian, and evolved by a neural ODE (Viswanath et al., 19 May 2026). The model is explicitly intended as a “virtual brain” that can forecast fMRI, EEG, and simulated spiking dynamics and support in silico perturbation prediction (Viswanath et al., 19 May 2026). Its brain-likeness lies in region-aligned latent states, heterogeneous edge-specific communication, and continuous-time evolution rather than in local synaptic plasticity.

ANN-based brain-network estimators

Some BLNN-relevant work uses neural networks primarily as devices for inferring brain organization. In task-fMRI decoding, directed and undirected Artificial Brain Networks are estimated by training a linear two-layer ANN to reconstruct each anatomical region’s BOLD activity from all others over a temporal window: $$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$0 The learned weights define directed or symmetric undirected connectivity graphs whose edge weights then serve as features for SVM classification (Kivilcim et al., 2018). This does not instantiate a brain-like architecture in the mechanistic sense, but it contributes to BLNN research on how cognition can be represented through learned network organization.

Physical and iontronic networks

Iontronic BLNNs push the architectural idea into soft matter. Nodes are electrolyte reservoirs, edges are conical microfluidic channels, and the synapse-like quantity is a history-dependent conductance $$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$1 obeying

$$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$2

Learning can modify pressures, concentrations, lengths, or radii rather than digital weights (Conte et al., 6 Nov 2025). These systems are brain-like mainly in their distributed analog adaptation, multimodal responsiveness, and embodied memory.

5. Brain-likeness as empirical correspondence

A substantial strand of BLNN research defines brain-likeness by what a model predicts or reproduces rather than by how it is trained.

In visual prosthesis modeling, a Brain-Score-validated EfficientNet-B1 is adapted into a retinotopic V1/V2-like population by computing maximally activating stimuli for individual units, mapping receptive-field centers into cortical coordinates with the Wedge-Dipole model, and decoding electrically stimulated unit populations into phosphenes using weighted sums of preferred stimuli with local normalization (Granley et al., 2022). The resulting phosphenes are qualitatively round, yellow-white, and sometimes patchy or shimmering, and V2 stimulation yields percepts on average 41% larger than V1 at the same visual-field location (Granley et al., 2022). The model is brain-like mainly because it uses a brain-aligned representation and a population-decoding readout under causal intervention.

A more general emergent-property approach evaluates whether deep networks display qualitative perceptual and neural effects characteristic of biological vision. Across 32 models and 15 properties, architecture is reported to have the strongest effect on brain-like emergent properties compared with dataset and training regime (Rajesh et al., 2024). Several properties, such as object normalization, scene incongruence, mirror confusion, and some correlated-sparseness effects, emerge broadly; others, such as relative size encoding, surface invariance, and 3D processing, are generally weak or absent (Rajesh et al., 2024). No single model dominates across all properties. The top Brain Property Match scores are diverse, with VOneCORnet-S, Inception v3, a Places365-trained ViT-B, a DINO-trained ViT-B, and DenseNet161 all ranking highly (Rajesh et al., 2024). This suggests that brain-likeness is multidimensional and architecture-sensitive rather than reducible to one benchmark number.

At the multimodal level, neuron-level analysis of vision-LLMs shows that artificial neurons derived from attention substructures can predict fMRI responses across language, vision, attention, and default-mode systems during movie viewing (Ren et al., 19 Oct 2025). The framework uses sparse dictionary learning on artificial-neuron time series and voxel-wise encoding to identify shared temporal motifs. Reported findings include functional redundancy, polarity patterns in artificial and biological units, and architecture-dependent specialization versus integration: CLIP’s separate branches align with modality-specific specialization, whereas METER’s cross-modal design yields more unified cross-modal activation (Ren et al., 19 Oct 2025). This broadens BLNN evaluation from univariate encoding to internal organizational principles.

6. Dynamics, information flow, and evolutionary hypotheses

BLNN research also includes work on what kinds of dynamical regimes or growth rules might generate brain-like networks.

A notable hypothesis is that brain networks evolve by maximizing internal information flow capacity, operationalized as an upper bound on mutual information rate given by the difference of the two largest Lyapunov exponents: $$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$3 In networks of Hindmarsh–Rose neurons grown by adding inter-cluster chemical connections only when they increase $$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$4, the resulting dynamics and graph spectra resemble those of Caenorhabditis elegans and human connectomes more closely than alternatives (Antonopoulos et al., 2015). A key conclusion is that the most brain-like regime is not maximal synchronization but moderately low global synchronization combined with high information flow capacity (Antonopoulos et al., 2015). This suggests a BLNN design principle: useful brain-like computation may require integrated yet non-collapsed dynamics, with local specialization and only partial global coherence.

A related message appears in neuromorphic-hardware critiques, where soft rank-order coding, modular segregation and integration, and diverse local plasticity rules are proposed as principles more likely to support brain-like function than simply spiking versions of ANN-style architectures (Maass, 2023). Together, these works position BLNN design near a tradeoff region between coherence and heterogeneity, not at either extreme.

7. Hardware, systems, and deployment

The practical realization of BLNNs is a major theme. Several papers argue that local learning, sparse activity, and modular architecture are not only biologically motivated but also advantageous for low-power or scalable systems.

StreamBrain shows that BCPNN can be expressed as a Keras-like domain-specific framework and mapped to CPUs, GPUs, FPGAs, and distributed-memory systems. By identifying batched outer products and matrix multiplications as the dominant kernels, it trains MNIST in about 10 seconds on an NVIDIA A100, achieves above 95% accuracy across implementations, and explores custom reduced-precision formats on FPGAs, finding that BF20 and above preserve accuracy while BF16 causes only modest degradation (Podobas et al., 2021). This matters because it demonstrates that a local-probabilistic BLNN need not remain a neuroscience toy model.

Embedded FPGA acceleration pushes this toward edge deployment. On a Zynq UltraScale+ SoC, BCPNN inference-only kernels achieve up to $$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$5 speedup and up to $$\frac{\partial }{\partial \theta_{a,b}^{(l)} } \mathbb{E}\!\left[ L(\vec{Z}(n),\vec{O}(n))\right] = \mathbb{E}\!\left[ - L(\vec{Z}(n),\vec{O}(n))\, z_b^{(l)}(n) \frac{\partial \log f_{a,l}(r_a^{(l)}(n))}{\partial r_a^{(l)}(n)} \right],$$6 board-energy savings over ARM baselines on MNIST, Pneumonia, and Breast Cancer tasks while preserving FP32/FP16 accuracy (Hafiz et al., 23 Jun 2025). The hardware supports online learning in smaller models and scalable inference in larger ones, showing a concrete path from biologically inspired local learning to deployable embedded systems (Hafiz et al., 23 Jun 2025).

Neuromorphic surveys complement these systems papers by reviewing hardware platforms such as Neurogrid, Braindrop, ROLLS, BrainScaleS, DYNAPs, TrueNorth, SpiNNaker, Loihi, Darwin, and Tianjic/Tianji, emphasizing tradeoffs among on-chip plasticity, routing, configurability, scale, and energy efficiency (Sima et al., 2023). A recurrent conclusion is that BLNN deployment requires hardware-software co-design, not merely porting ANN assumptions onto different substrates.

The iontronic-network work extends this systems perspective further by suggesting that BLNN hardware need not even remain electronic. In that setting, forward passes are physical ionic relaxation, and learning is adaptation of geometry or environmental variables rather than stored numerical weights (Conte et al., 6 Nov 2025). This opens a distinct branch of BLNN research centered on adaptive materials.

8. Strengths, limitations, and recurrent controversies

Several strengths recur across BLNN research. Local plasticity rules provide mechanistic interpretability and compatibility with neuromorphic hardware (Xiao et al., 2019, Pehlevan et al., 2019, Zenke et al., 2020, Ravichandran et al., 2024). Modular sparse architectures often yield structured, low-entropy representations and reduce energy or communication costs (Hafiz et al., 23 Jun 2025, Ravichandran et al., 2024). Correspondence-based evaluations can reveal qualitative similarities missed by raw benchmark scores (Granley et al., 2022, Rajesh et al., 2024, Ren et al., 19 Oct 2025). Structure-aware dynamical models support perturbation analysis and interpretable regional organization (Viswanath et al., 19 May 2026).

The limitations are equally clear. Many biologically plausible learning rules have high estimator variance, slower convergence, or modest benchmark scale relative to backpropagation (Xiao et al., 2019, Zenke et al., 2020). Spiking alone does not guarantee brain-like function if network organization and training remain ANN-like (Maass, 2023). Some models are brain-inspired in representation but not in learning, as in prosthesis modeling with a pretrained EfficientNet (Granley et al., 2022). Others are useful to neuroscience while only weakly connected to biologically inspired ANN design, as in ANN-based fMRI connectivity estimation (Kivilcim et al., 2018).

A frequent controversy concerns what counts as “brain-like.” One line of work emphasizes mechanistic similarity: spikes, local rules, interneuron diversity, dendritic gating, and neuromorphic substrates (Maass, 2023, Zenke et al., 2020, Sima et al., 2023). Another emphasizes functional correspondence: predict neural data, reproduce perceptual signatures, or model causal interventions in cortex (Granley et al., 2022, Rajesh et al., 2024, Ren et al., 19 Oct 2025). A third emphasizes structural estimation of brain organization, such as learning connectivity graphs or virtual-brain dynamics (Kivilcim et al., 2018, Viswanath et al., 19 May 2026). These agendas overlap but are not interchangeable.

Another recurring caution is that some notions of directionality are weaker than they may sound. In ANN-estimated directed brain networks from fMRI, “directed” means asymmetric predictive coefficients, not causal or effective connectivity in the strong neuroscientific sense (Kivilcim et al., 2018). Similarly, alignment of vision-language-model neurons with fMRI voxels is representational and correlational, not mechanistic identity (Ren et al., 19 Oct 2025). BLNN literature therefore often requires careful separation of structural analogy, computational utility, and biological claim strength.

9. Research directions

Several directions emerge repeatedly across the surveyed work. One is richer biologically plausible credit assignment: integrating local eligibility traces, modulatory signals, interneuron gating, and dendritic computation into scalable algorithms that compete more directly with backpropagation (Zenke et al., 2020, Etter, 6 Nov 2025, Maass, 2023). Another is broadening the scope of evaluation beyond neural-predictivity benchmarks to include controlled emergent properties, causal intervention settings, and multimodal correspondence (Granley et al., 2022, Rajesh et al., 2024, Ren et al., 19 Oct 2025).

A third direction is integrating modular sparse local-learning architectures with modern hardware. BCPNN already has evidence for CPU/GPU/FPGA scalability and embedded energy efficiency (Podobas et al., 2021, Hafiz et al., 23 Jun 2025), while neuromorphic critiques point toward richer microcircuit motifs, soft rank-order coding, and context-gated plasticity (Maass, 2023). A fourth is advancing virtual-brain and structured-dynamics models that retain anatomical alignment and support perturbation prediction (Viswanath et al., 19 May 2026). A fifth is exploring non-electronic substrates and morphological learning, where computation and memory are inseparable from material state (Conte et al., 6 Nov 2025).

A plausible implication is that BLNN research will continue to diversify rather than converge on one dominant architecture. The field’s central question is not whether a single model can be maximally brain-like in every sense, but which biological principles are computationally essential, scalable, and useful. Current evidence suggests that architecture matters strongly for emergent brain-like properties (Rajesh et al., 2024), that local and probabilistic learning can be made practical in specialized regimes (Ravichandran et al., 2024, Hafiz et al., 23 Jun 2025), and that correspondence to brain organization can be probed at finer levels than final-layer similarity (Ren et al., 19 Oct 2025). In that sense, BLNNs remain best understood as a research program linking neuroscience, machine learning, dynamical systems, and adaptive hardware rather than as one settled model class.