Artificial Neural Networks in Neuroscience

• Artificial neural networks are computational models that simulate neural circuits using layered architectures and non-linear activations, widely applied to study brain functions.
• They serve as testbeds for experimental design by replicating neural dynamics and behavioral outcomes through lesion studies and performance benchmarks.
• Interpretability techniques, such as saliency maps and representational similarity analysis, bridge model components with empirical neural data to refine understanding.

Artificial neural networks (ANNs) constitute a foundational modeling and analysis toolset for neuroscientists, both as computational analogues of biological networks and as testbeds for developing and validating experimental and analytic methods. ANNs operationalize central questions in neuroscience, such as how neural circuits instantiate behavior, how synaptic or circuit changes cause observable performance shifts, and how distributed population codes emerge and can be decoded. This article surveys the core architecture, theoretical grounding, biological mappings, experimental paradigms, interpretability challenges, and future directions that define ANN use in contemporary neuroscience research.

1. Fundamental Principles and Architectural Foundations

ANNs are composed of interconnected units (“neurons”) organized into layers, where each unit computes a weighted sum of its inputs followed by a non-linear activation. For a single layer, the operation is $y = \sigma(W x + b)$, where $W$ is a weight matrix, $b$ a bias vector, and $\sigma(\cdot)$ a non-linearity (e.g. ReLU, sigmoid, tanh). Deep neural networks (DNNs) stack these layers, yielding highly expressive input–output maps $f(x; \Theta)$ trained to minimize a loss function such as cross-entropy or mean-squared error using back-propagation and stochastic gradient descent (Yang et al., 2020, Storrs et al., 2019).

Key ANN architectures relevant to neuroscience include convolutional neural networks (CNNs), which implement local receptive fields and weight sharing as in cortical visual areas; recurrent neural networks (RNNs), including rate-based and spiking variants, which capture temporal dynamics and working memory; and reinforcement learning (RL) agents, which learn sequential policies via policy gradients or Q-learning updates (Lindsay, 2023). In all cases, models are explicitly optimized for quantifiable behavioral objectives—accuracy, reaction time, reward, etc.—making them suited for linking microscopic neural dynamics to macroscopic behavior.

2. Mapping Biological Computation and Circuit Motifs to ANNs

Neuroscientific modeling with ANNs exploits analogies between artificial and biological systems at multiple levels (He et al., 2024, Storrs et al., 2019, Saxe et al., 2020). CNNs embody retinotopic and hierarchical organization of the ventral visual pathway (V1/V2/V4/IT), with learned filters paralleling cortical receptive fields. RNNs and long short-term memory units encode recurrent cortical loops and emergent dynamical motifs such as attractors and context-dependent routing observed in prefrontal and parietal cortex (Parga et al., 2022, Yang et al., 2020). Dale’s law and synaptic constraint can be imposed to restrict connectivity to realistic excitatory/inhibitory motifs (Yang et al., 2020).

Distinct families of artificial neuron models map to biological diversity: beyond standard sigmoidal or rectified-linear units, quadratic (high-order), dendritic, and spiking neuron models encode sub-compartmental integration, multiplicative interactions, burst firing, and event-driven dynamics. These augmentations confer gains in efficiency, memory, and interpretability and are being explored as a route to NeuroAI—a convergence between neural computation and engineered learning systems (Fan et al., 2023).

3. Experimental Design: Linking Behavior, Representation, and Mechanism

ANNs enable rigorous, causal tests of hypotheses regarding the mapping from neural activity to behavior (Lindsay, 2023, Saxe et al., 2020, Parga et al., 2022). A “behavior-first” approach grounds neuroscience in observable task contrasts—before/after learning, with/without attention, etc.—and tasks ANNs with reproducing both performance and underlying neural changes. Model perturbations (e.g., silencing units, weight lesions, gain modifications) are systematically imposed, with impacts quantified via the same performance metrics used for biological systems.

Case studies show that, when engineered to replicate psychophysical and neurobiological experimental paradigms, ANNs can recapitulate both qualitative and quantitative in vivo findings. For example, silencing RL agent neurons encoding value and policy can mirror the effects of optogenetic silencing in mice, including drop in performance and effect sizes (Lindsay, 2023). In composition learning, ablation of Q-value–encoding units doubles the episodes required for convergence, paralleling behavioral learning deficits in animals. CNN gain modulation in spatial attention tasks quantitatively tracks human improvements in detection d′.

4. Interpretability: Connecting ANN Components to Neural Data and Function

Interpretability in ANNs spans two distinct but overlapping objectives: (1) AI-centric interpretability, focused on attributing output changes to internal components for engineered systems, and (2) neuroscience-centric interpretability, which demands explicit mapping between model components and empirically measurable brain constructs, such as brain areas, receptive fields, or task variables (Kar et al., 2022). Saliency maps, gradient-based attribution, integrated gradients, deep dream/feature visualization, class activation mapping, and layer-wise relevance propagation constitute the core set of tools for component-wise analysis.

For neuroscientific interpretability, these techniques are employed alongside explicit alignment procedures—layer/unit-to-area mapping, receptive field fitting, and population-level representational similarity analysis (RSA) correlating inter-stimulus geometry in ANNs and biological recordings. Tuning curves, spatial/temporal receptive field structure, and recurrent/feedback dynamics in ANNs are compared to neurophysiological and functional imaging observations. Rigorous validation against withheld neural or behavioral data is recommended to prevent overinterpretation, and explanations are to be benchmarked using metrics such as Brain-Score or neural RSA (Kar et al., 2022, He et al., 2024).

5. Analysis Toolbox: Neuroscientific Methods Applied to Artificial Models

ANNs serve as tractable platforms for testing and refining analytical methods from systems and cognitive neuroscience, owing to their full observability and manipulability (Lindsay, 2022). A suite of tools—including principal component analysis (PCA), demixed PCA, latent factor modeling (Gaussian process or probabilistic latent dynamical system), network/graph-theoretic analysis, encoding/decoding models, and information-theoretic measures—are systematically deployed on unit activations to uncover dimensionality, manifold geometry, representational structure, and functional connectivity.

Procedures involve applying these methods to population codes, validating insights through in silico perturbations (e.g., lesioning subspaces or network nodes), and scoring success based on experimental predictive power. For example, demixed PCA applied to a network trained on noisy digit classification allowed specific isolation and removal of noise dimensions, resulting in improved task performance—an experimentally validated outcome (Lindsay, 2022).

6. Case Studies, Quantitative Benchmarks, and Implementation Examples

Empirical studies systematically compare model–brain alignment at the representational and behavioral levels, using metrics such as RSA correlation values, encoding/decoding accuracy, variance explained, and effect sizes from targeted interventions. Examples include:

• RSA between CNN layers and inferotemporal cortex population responses (correlation ≈0.8 for late layers) (He et al., 2024).
• Lesion studies: ablation of functional units in LLMs sharply increases or decreases behavioral task scores, mapping to in vivo lesion and inactivation results (Lindsay, 2023).
• Mixed selectivity and superposed codes in transformers directly mirror population coding principles found in prefrontal cortex recordings (He et al., 2024, Parga et al., 2022).

ANNs trained to model complex tasks—object recognition, working memory, context-dependent decision-making—exhibit emergent computational motifs (line attractors, context gating, bump attractors, trajectory manifolds) directly comparable to those observed in single-unit and population-level cortical recordings (Parga et al., 2022). Implemented using standard frameworks (e.g., PyTorch), with code structures and parameterizations tailored for biological realism (sparse/Dale-constrained connectivity, spiking dynamics), these models provide a computational microscope for dissecting neural algorithms (Yang et al., 2020).

7. Challenges, Recommendations, and Future Directions

As scaling continues, analysis at all three of Marr’s levels—computation/behavior, algorithm/representation, and implementation—remains essential (He et al., 2024). Practical guidelines emphasize matching the analysis level to the neuroscientific question, combining behavioral benchmarks, representational analyses, and circuit interventions. Model architectures are increasingly heterogeneous, incorporating diverse artificial neuron types, with ongoing efforts to build neuroinformatics infrastructure for catalogue and comparison (Fan et al., 2023).

Open problems include scaling interpretability and alignment techniques to billion-parameter models, integrating temporally resolved neural data, bridging developmental phases of representation formation, and disambiguating multiplexed codes (He et al., 2024, Kar et al., 2022). Systematic, pre-registered testing of neuroscience tools on ANNs is advocated to empirically calibrate mechanistic understanding and drive next-generation analytic methodology (Lindsay, 2022).

Artificial neural networks thus offer neuroscientists a laboratory for mechanistic, scalable, and testable hypotheses—enabling a deepened convergence between biological and artificial intelligence through shared mathematics, architectures, and analysis paradigms.