Neocognitron

The Neocognitron is a self-organizing, hierarchical multilayered artificial neural network model proposed by Kunihiko Fukushima in 1980 for performing visual pattern recognition that is invariant to shifts in position, small deformations, and variations in size.^[1] Designed to emulate the human visual system's ability to perceive patterns based on their essential geometrical features (Gestalt organization), it operates without supervised training, relying instead on unsupervised reinforcement of synaptic connections through repeated exposure to stimuli.^[1] The architecture draws inspiration from the hierarchical structure of the mammalian visual cortex, particularly the findings of Hubel and Wiesel on simple and complex cells.^[2] It consists of an input layer followed by successive modules, each comprising two types of cell layers: S-cells (simple cells) that detect specific features like lines or edges at precise locations using modifiable excitatory synapses, and C-cells (complex cells) that achieve positional tolerance by pooling inputs from multiple S-cells via fixed connections, allowing responses to features regardless of exact position.^[2] Receptive fields expand progressively across layers, enabling the extraction of increasingly abstract features, from basic orientations in early stages to complete pattern representations in later ones.^[2] Learning occurs through a self-organization process where S-cell synapses are selectively strengthened proportional to the output activity elicited by input patterns, without requiring error signals or a teacher.^[2] The Neocognitron laid foundational groundwork for subsequent neural network advancements, particularly influencing the development of convolutional neural networks (CNNs) by introducing key concepts like shared weights, feature mapping, and hierarchical feature extraction.^[3] Early applications demonstrated its effectiveness in tasks such as handwritten character recognition, and later refinements extended its capabilities to handle more complex deformations and incremental learning.^[4] Despite limitations in scalability compared to modern backpropagation-trained models, its biologically motivated design remains a benchmark for understanding shift-invariant pattern recognition in artificial systems.^[4]

History

Invention and Early Development

The Neocognitron was proposed by Kunihiko Fukushima in 1980 as a hierarchical, self-organizing neural network designed to achieve robust pattern recognition invariant to positional shifts.^[1] This model addressed limitations in earlier artificial neural networks by incorporating mechanisms for feature extraction and tolerance to deformations, drawing initial inspiration from the hierarchical organization observed in the visual cortex.^[1] The Neocognitron evolved directly from Fukushima's Cognitron, a multilayered neural network introduced in 1975 that excelled at extracting specific features like line segments but failed to generalize recognition across shifted positions due to its rigid connectivity.^[5] By modifying the Cognitron's architecture to include inhibitory connections and broader receptive fields in alternate layers, the Neocognitron gained the desired shift invariance, marking a pivotal advancement in unsupervised learning for visual tasks.^[1] Early validation came through computer simulations detailed in the 1980 publication, where the network successfully recognized five simple patterns shaped as the digits '0' to '4,' even when presented in various locations within the input field.^[1] These experiments highlighted the model's practical utility for basic optical character recognition without requiring supervised training data. Fukushima's work on the Neocognitron continued with refinements to enhance its efficiency and applicability, including a 1988 reformulation that simplified the network's structure for better visual pattern recognition while preserving self-organization.^[6] Further advancements appeared in 2003, where an improved version demonstrated high accuracy on larger datasets of handwritten digits, achieving recognition rates above 98% on test sets after training on thousands of samples.^[7]

Inspirations from Neuroscience

The Neocognitron's design draws significant inspiration from the neurophysiological studies of David Hubel and Torsten Wiesel, who conducted pioneering experiments on the visual cortex of cats in the late 1950s and early 1960s. Using microelectrodes to record from single neurons, they identified distinct cell types responsive to specific visual stimuli, such as oriented edges and lines, revealing how the cortex processes visual information through specialized receptive fields. Their work demonstrated that simple cells exhibit excitatory and inhibitory regions aligned to precise stimulus orientations and positions, while complex cells maintain responsiveness to the same features across a broader range of positions, providing tolerance to shifts in input location.^[8] This hierarchical organization of the visual system, as elucidated by Hubel and Wiesel, forms a foundational influence on the Neocognitron. Visual processing begins with photoreceptors in the retina, which detect light and transmit signals via ganglion cells through the optic nerve to the lateral geniculate nucleus (LGN) of the thalamus. The LGN relays this information to the primary visual cortex (V1), where layered processing occurs, with successive stages building increasingly abstract representations through interconnected simple and complex cells.^[9] In the Neocognitron, this progression is mirrored by alternating layers that progressively expand receptive fields and enhance feature selectivity, emulating the brain's multi-stage filtering from basic edges to complex patterns.^[1] The model's cell types directly map to these neurobiological findings: S-cells correspond to simple cells, possessing fixed, feature-specific receptive fields that detect precise local patterns, whereas C-cells align with complex cells, offering positional invariance by integrating inputs from multiple S-cells to tolerate variations in stimulus location.^[2] Kunihiko Fukushima explicitly drew on Hubel and Wiesel's model of cortical hierarchy—from LGN inputs to simple cells, then complex cells, and onward to higher-order processing—to construct the Neocognitron's architecture, aiming to replicate the visual system's robustness.^[1] A core objective of the Neocognitron is to emulate the brain's self-organizing capabilities, enabling robust pattern recognition without supervised teaching signals. This unsupervised approach is inspired by developmental processes in neural systems, such as the self-organization observed in visual pathway formation, allowing the network to adaptively tune connections based on input statistics alone, much like cortical plasticity in vertebrates.^[2]

Architecture

Overall Structure

The Neocognitron features a hierarchical, multi-layered architecture that processes visual inputs through a series of alternating layers, enabling robust pattern recognition with tolerance to positional shifts. The network begins with an input layer, denoted as U_0, which simulates an array of retinal photoreceptors and directly receives the stimulus pattern as a two-dimensional excitation map. This is followed by a cascade of simple cell layers (S-layers, U_{S_i}) and complex cell layers (C-layers, U_{C_i}), forming the core of the model's modular organization.^[1] The modular structure consists of multiple processing stages, each comprising an S-layer followed by a corresponding C-layer. In a standard simulation, the Neocognitron is implemented with seven layers: the input layer U_0, stage 1 (U_{S1} and U_{C1}), stage 2 (U_{S2} and U_{C2}), and stage 3 (U_{S3} and U_{C3}). This staged arrangement allows information to flow sequentially from low-level to high-level processing, with each module building upon the outputs of the previous one to refine feature representations. S-cells extract specific features, while C-cells aggregate these to achieve positional invariance.^[1] Progressing through the hierarchy, the Neocognitron achieves increasing levels of abstraction, where initial layers near the input detect rudimentary features like oriented edges or lines, and subsequent layers integrate these into more sophisticated patterns such as contours or complete shapes. This progression is accompanied by enhanced tolerance to translations and slight deformations in the input pattern's position, mimicking aspects of biological visual processing.^[1] Cell density diminishes with network depth, while receptive field sizes expand correspondingly, allowing higher layers to capture broader contextual information with greater efficiency. For instance, in the deepest C-layer, each feature-specific cell plane typically contains only a single cell whose receptive field spans the entire input area, culminating in shift-invariant recognition of complex patterns.^[1]

Cell Types and Layers

The Neocognitron employs two primary types of cells, S-cells and C-cells, arranged in alternating layers to facilitate hierarchical feature processing. S-cells, also known as simple cells, are responsible for feature extraction by detecting specific local patterns such as line segments or edges. S-cells receive modifiable excitatory inputs from cells in the preceding C-layer and modifiable inhibitory inputs from accompanying V-cells. The V-cells receive fixed excitatory inputs from the same group of C-cells and provide shunting inhibition to normalize responses and suppress background activity. This allows adaptation during learning to sharpen selectivity for particular features.^[2]^[10]^[11] In contrast, C-cells, or complex cells, contribute to positional invariance by integrating signals from multiple S-cells within a defined local region of the previous S-layer. Their connections are fixed and unmodifiable, typically involving a summation or averaging mechanism that activates the C-cell if sufficient presynaptic S-cells respond, thereby tolerating minor shifts or deformations in the input pattern without altering the feature representation.^[2]^[10] The architecture alternates between S-layers and C-layers across multiple modules, where each S-layer draws inputs exclusively from the prior C-layer, and each C-layer aggregates outputs from the subsequent S-layer. This sequential arrangement enables progressive abstraction, with early layers handling basic features and later ones combining them into more complex representations.^[2] Within each layer, cells are organized into multiple cell-planes, where each plane is tuned to a distinct feature type, such as oriented lines or endpoints, and consists of cells with identical receptive field properties but positioned at different locations across the input field. For instance, simulations of the model often incorporate around 24 cell-planes per layer to cover a variety of feature orientations and types.^[2]^[10]

Receptive Fields and Cell-Planes

In the Neocognitron, receptive fields define the spatial regions over which individual cells integrate inputs from preceding layers, characterized by fixed positions within their respective cell-planes and the pattern of synaptic connections from the preceding layer, allowing cells to detect specific local features such as lines or edges. The excitatory connections are modifiable, while inhibition via V-cells helps suppress responses to irrelevant activity, enhancing contrast for the detected features.^[2] This organization, inspired by visual cortex mechanisms, allows cells to detect local features while suppressing irrelevant background activity. As layers progress deeper in the network, receptive field sizes systematically increase, enabling the aggregation of increasingly complex patterns from broader input areas—starting small in early layers for basic edge detection and expanding to encompass nearly the entire visual field in the final layers.^[2] Cell-planes consist of groups of cells arranged in a two-dimensional array, where each plane comprises cells sharing identical receptive field shapes but with progressively shifted positions to ensure comprehensive coverage of the input space without gaps or overlaps.^[11] Within a single cell-plane, the fixed positional offsets between cells allow the plane as a whole to map the entire input layer, with each cell's receptive field overlapping slightly with its neighbors to maintain continuity.^[2] S-cells and C-cells, which play roles in feature extraction and integration respectively, are organized into these planes, facilitating a modular representation of the visual field.^[11] Feature selectivity in the Neocognitron arises from the specialized tuning of receptive fields within each cell-plane, where planes are dedicated to detecting specific orientations or patterns, such as vertical or horizontal lines in early layers and more complex motifs like corners or curves in later ones.^[2] This selectivity builds hierarchically: simpler features extracted in shallow layers, like oriented edges, serve as building blocks for higher-layer planes that combine them into compound patterns, such as character strokes, through the excitatory-inhibitory dynamics of the receptive fields.^[11] Each cell-plane thus functions as a feature map, with multiple planes per layer capturing a repertoire of distinct features to support robust pattern discrimination. The arrangement of receptive fields across cell-planes preserves translational symmetry, enabling the network to recognize patterns invariant to small positional shifts in the input without requiring explicit invariance training.^[2] By virtue of the shifted receptive fields within planes and the pooling-like integration in deeper layers, a feature detected in one location can activate corresponding cells elsewhere, allowing consistent responses to translated inputs while tolerating minor deformations.^[11] This structural property underpins the Neocognitron's position-tolerant recognition, distinguishing it from rigid template matching approaches.

Mathematical Formulation

Output Functions of Cells

In the Neocognitron model, the output functions of S-cells and C-cells form the basis for feature extraction and invariance to shifts, respectively, through weighted summations over specific receptive field regions combined with normalization mechanisms. S-cells, which detect local features similar to simple cells in visual cortex, compute their outputs by integrating excitatory inputs from preceding C-cells while incorporating inhibitory modulation for selectivity. C-cells, akin to complex cells, further process S-cell outputs to achieve positional tolerance by pooling over broader regions with a non-linear response characteristic.^[2] The output of an S-cell in the l-th layer, denoted U_{sl}(k_l, n), where k_l indexes the cell-plane and n the position, is given by the following equation:

U_{sl}(k_l, n) = \frac{\sum_{k_{l-1}=1}^{K_{l-1}} \sum_{v \in S_l} a_l(k_{l-1}, v, k_l) \cdot U_{cl-1}(k_{l-1}, n + v)}{1 + r_l \cdot b_l(k_l) \cdot V_{cl-1}(n)},

where the summation over v \in S_l aggregates excitatory contributions across the receptive field S_l of the S-cell, a_l(k_{l-1}, v, k_l) are the modifiable excitatory weights connecting C-cells from the previous layer's K_{l-1} planes, U_{cl-1} represents the inputs from those C-cells, r_l scales the inhibitory influence, b_l(k_l) are inhibitory coefficients specific to the cell-plane, and V_{cl-1}(n) captures the local inhibitory signal from the preceding layer.^[2] This divisive normalization in the denominator enhances the S-cell's response to specific feature alignments by suppressing activity when surrounding inhibition is high, with the size of S_l increasing across layers to capture progressively larger patterns.^[2] Similarly, the C-cell output U_{cl}(k_l, n) normalizes and non-linearly transforms pooled S-cell activations to produce shift-invariant representations:

U_{cl}(k_l, n) = \phi\left[\frac{\sum_{v \in D_l} d_l(v) \cdot U_{sl}(k_l, n + v)}{1 + V_{sl}(n)}\right],

where \phi = \frac{x}{\alpha + x} provides a saturating non-linearity (with \alpha > 0 controlling the saturation level), the sum over v \in D_l weights inputs from the same S-cell plane across the broader receptive field D_l using fixed coefficients d_l(v) that typically decrease with distance from the center, and V_{sl}(n) denotes the inhibitory term from the S-layer.^[2] The role of D_l is to average feature detections over slight positional variations, ensuring the C-cell fires robustly as long as the feature appears within its field, thus building hierarchical invariance.^[2] These computations integrate excitatory summation with normalization to mimic biological selectivity while enabling the network's tolerance to input distortions.^[2]

Inhibitory Mechanisms

In the Neocognitron, inhibitory mechanisms are primarily implemented through V-cells, which provide surround inhibition to modulate the responses of S-cells and C-cells, thereby promoting selective feature detection. A key component is the V-cell associated with the previous C-layer, which computes an inhibitory signal based on the root-mean-square activity across multiple cell-planes to suppress extraneous responses in the subsequent S-layer. This is formulated as V_{cl-1}(n) = \sqrt{ \frac{1}{K_{l-1}} \sum_{k_{l-1}=1}^{K_{l-1}} \sum_{v \in S_l} c_{l-1}(v) \cdot U_{cl-1}(k_{l-1}, n + v)^2 }, where c_{l-1}(v) are fixed coefficients that decrease with distance |v| from the center, defining a surround receptive field for suppression.^[2] Another V-cell type targets the S-layer outputs to normalize inputs to the C-layer, averaging excitatory activities over cell-planes and spatial offsets. Its output is given by V_{sl}(n) = \frac{1}{K_l} \sum_{k_l=1}^{K_l} \sum_{v \in D_l} d_l'(v) \cdot U_{sl}(k_l, n + v), with d_l'(v) as fixed weighting coefficients that emphasize central regions while attenuating peripheral ones. These coefficients remain unchanged during learning, ensuring stable inhibitory dynamics.^[2] The primary role of these V-cells is to prevent over-response to noise or intense local stimuli by implementing lateral and surround inhibition, which sharpens feature selectivity and enhances response stability across positional variations in input patterns. By integrating these inhibitory signals into the denominator of S-cell and C-cell output functions, the network achieves robust normalization that mitigates the effects of stimulus distortions.^[2]

Learning Process

Unsupervised Learning Rule

The Neocognitron employs an unsupervised learning paradigm that operates without external labels or teacher supervision, allowing the network to self-organize by repeatedly exposing it to stimulus patterns at the input layer. This teacherless approach enables the network to reinforce synaptic connections based on the responses elicited by input stimuli, gradually building an internal representation capable of recognizing patterns based on shape similarity while remaining unaffected by positional shifts.^[2] Central to this learning rule is the winner-take-all mechanism used to select representative S-cells. First, candidate S-cells are identified as those producing sufficiently large outputs. Then, within each S-plane, if multiple candidates exist, only the one with the highest output is selected as the representative for that plane, ensuring at most one representative per S-plane to avoid redundancy and focus on prominent feature detectors. An S-column refers to the group of S-cells from all S-planes at a given position.^[2] The learning process unfolds in iterative phases: first, the network is exposed to a series of training patterns; second, representative cells are identified via the winner-take-all selection from those S-cells producing the strongest activations; and third, connections are reinforced for these selected representatives across multiple iterations, progressively constructing hierarchical feature representations from simple to complex patterns.^[2] Synapse modifications, which detail the precise strengthening of these connections, follow this selection but are guided by the overall self-organizing rule. This unsupervised framework preserves translational invariance through the averaging function of C-layers, where C-cells integrate inputs from neighboring S-cells using fixed, unmodifiable synapses, thereby smoothing positional variations without relying on backpropagation or supervised error signals.^[2]

Synapse Modification

In the Neocognitron, synapse modification occurs during the unsupervised learning phase in the S-layers, where excitatory and inhibitory connections are selectively reinforced to enhance feature detection capabilities. This process targets only the representative S-cells within each S-plane, which are those exhibiting the strongest responses to input patterns via a winner-take-all mechanism.^[1] The reinforcement of excitatory synapses, denoted as a_l(k_{l-1}, v, k_l), connects presynaptic C-cells in the previous layer to postsynaptic S-cells in layer l. The update rule is given by

\Delta a_l(k_{l-1}, v, k_l) = q_l \cdot c_{l-1}(v) \cdot U_{C,l-1}(k_{l-1}, n + v),

where q_l is a layer-specific learning rate (e.g., q_1 = 1.0, q_2 = q_3 = 16.0), c_{l-1}(v) represents the fixed inhibitory synapse weights from V-cells, and U_{C,l-1}(k_{l-1}, n + v) is the output of the presynaptic C-cell. This adjustment strengthens connections from active presynaptic cells to the selected representative S-cell, promoting specialization for specific features.^[1] In parallel, the inhibitory synapse strength b_l(k_l) for each S-cell plane in layer l is modified to balance excitation. The update is

\Delta b_l(k_l) = \frac{q_l}{2} \cdot V_{C,l-1}(n),

where V_{C,l-1}(n) denotes the output of the inhibitory V-cells in the previous C-layer. This halves the learning rate relative to excitation, ensuring gradual adjustment of the overall inhibitory intensity across the cell-plane without altering individual inhibitory connections c_{l-1}.^[1] Notably, connections in the C-layers, including weights d_l from S-cells to C-cells and the inhibitory synapses c_{l-1}, remain fixed after initial setup to preserve positional invariance. Through repeated exposure to training patterns, these iterative updates progressively reinforce relevant synaptic pathways, allowing the network to develop hierarchical feature representations over multiple learning cycles.^[1]

Applications and Influence

Pattern Recognition Capabilities

The Neocognitron demonstrates robust pattern recognition by achieving invariance to positional shifts, size variations, and minor shape distortions through a hierarchical structure where tolerance builds progressively across layers. In simulations, simple cells (S-cells) in earlier layers detect local features with limited positional tolerance, while complex cells (C-cells) in subsequent layers integrate these to form more invariant representations, enabling the network to handle deformations without retraining.^[1] Experimental tests conducted in 1980 on handwritten-like digit patterns ("0" through "4") showcased this capability, with the network correctly discriminating all five classes even under random positional shifts and deformations such as added noise or partial occlusions. The final layer's C-cells produced distinct, non-overlapping response patterns for each digit variant, yielding 100% accuracy in recognition. This performance highlights the model's ability to generalize from learned prototypes to deformed inputs, as verified through computer simulations.^[1] Despite these strengths, early Neocognitron implementations had limitations, including the need for manual initialization of parameters like cell sensitivities and the number of cell-planes, which could degrade performance if not optimized. The model was also sensitive to extreme deformations, such as large rotations or severe distortions, where recognition accuracy dropped significantly without additional stages or adjustments.^[1] In engineering contexts, the Neocognitron's distortion-tolerant recognition made it suitable for applications like optical character recognition (OCR), where input patterns often vary in position, size, or slight form due to scanning imperfections. It offered a foundational approach for visual pattern matching tasks requiring robustness to real-world variability.^[1]

Relation to Convolutional Neural Networks

The Neocognitron, introduced by Kunihiko Fukushima in 1980, served as a foundational precursor to modern convolutional neural networks (CNNs) by pioneering key architectural principles such as shared weights, local connectivity, and hierarchical feature extraction. In the Neocognitron, all simple cells (S-cells) within a given S-layer share the same synaptic weights and spatial arrangement when connecting to the preceding layer, enabling efficient feature detection across the input without position-specific tuning. Local connectivity restricts inputs to small receptive fields, mimicking localized processing in the visual cortex, while the multi-layered structure progressively builds more abstract representations through alternating S-layers for feature extraction and C-layers for invariance. These elements were later formalized and extended in Yann LeCun's convolutional architecture, notably in his 1989 work on handwritten zip code recognition, where shared-weight layers and subsampling modules directly echoed the Neocognitron's design to reduce parameters and enhance generalization on image tasks.^[2]^[12] Key similarities between the Neocognitron and CNNs lie in their layered processing and mechanisms for translation invariance. Both architectures alternate between convolutional-like operations—S-layers in the Neocognitron detect oriented features via weighted sums over local regions, akin to convolution kernels in CNNs—and pooling operations, where C-layers in the Neocognitron perform fixed averaging over shifted positions to achieve positional tolerance, comparable to subsampling or pooling layers in CNNs that downsample and provide shift invariance. This hierarchical alternation allows both models to extract invariant features, starting from edge detectors in early layers to complex patterns in deeper ones, fundamentally influencing CNN designs like LeNet for tasks requiring robustness to translations and minor distortions.^[2]^[13]^[12] Despite these parallels, notable differences distinguish the Neocognitron from contemporary CNNs, particularly in learning paradigms and pooling flexibility. The Neocognitron relies on unsupervised learning rules, where synaptic weights in S-layers self-organize through competitive Hebbian-like reinforcement based on cell activations during repeated stimulus exposure, without labeled supervision. In contrast, CNNs, as developed by LeCun, employ end-to-end supervised training via backpropagation, allowing gradients to optimize all layers jointly for task-specific performance. Additionally, Neocognitron pooling in C-layers uses fixed, non-trainable connections that broadly inhibit or average inputs for invariance, whereas CNN pooling layers are often trainable or selectable (e.g., max or average pooling) and integrated into the differentiable optimization process.^[2]^[14]^[12]^[13] Fukushima's 1980 Neocognitron directly inspired LeCun's convolutional architecture for handwriting recognition, bridging early neural models to practical deep learning systems. LeCun explicitly referenced the Neocognitron in his 1989 implementation, adapting its modular structure—feature maps followed by subsampling—for zip code digit classification, which achieved high accuracy on real-world data and laid the groundwork for subsequent CNN advancements. This influence is further acknowledged in LeCun's later reviews, positioning the Neocognitron as a conceptual ancestor that motivated the shift toward trainable, supervised convolutional networks.^[2]^[12]^[13]

Modern Developments

In 2003, Kunihiko Fukushima introduced an improved version of the neocognitron designed to handle complex patterns such as handwritten digits through enhanced feature extraction and learning mechanisms.^[15] Key modifications included an inhibitory surround in connections from S-cells to C-cells for sharper feature detection, a contrast-extracting preprocessing layer ahead of edge extraction, and self-organization of line-extracting cells via unsupervised learning, followed by supervised competitive learning in the final stage.^[15] This version achieved a 98.6% recognition rate on a blind test set of 3000 patterns from the ETL1 database of Japanese handwritten digits, demonstrating robustness to variations in writing styles.^[15] Fukushima further advanced the model in 2021 with the Deep CNN Neocognitron, incorporating deeper layers to build a more hierarchical feature extraction process inspired by the mammalian visual cortex. Unlike standard convolutional neural networks (CNNs), which rely heavily on supervised backpropagation and mathematical optimization, this version emphasizes biological fidelity through unsupervised self-organization and the use of S-cells for selective feature detection and C-cells for position-tolerant responses. These elements enable superior handling of distortions and noise. Techniques such as interpolating vectors, developed earlier, have demonstrated reductions in error rates, for example from 1.52% to 1.02% on a 5000-digit blind test set.^[16]^[17] Subsequent enhancements have integrated deeper architectures with subtractive inhibition from V-cells to S-cells, providing adaptive suppression of irrelevant signals and improving overall tolerance to input variations.^[18] Applications to larger datasets, such as the ETL1 collection, have validated these features by scaling the network to process thousands of diverse samples while maintaining high accuracy in feature hierarchy formation.^[15] In the 2020s, the neocognitron has found use in bio-inspired vision systems for robust object recognition, leveraging its distortion-resistant properties to simulate human-like perceptual invariance. Hybrid models combining the neocognitron with deep learning techniques, such as evolutionary algorithms, have emerged for specialized tasks; for instance, the 2024 Evolutionary Gravity Neocognitron Neural Network (EGraNNN) integrates gravitational search optimization to classify rodent behaviors from video data, achieving over 95% accuracy across nine action categories like eating and jumping.^[19] Fukushima's ongoing research, conducted as a senior research scientist at the Fuzzy Logic Systems Institute, continues to refine unsupervised learning rules for enhanced robustness in pattern recognition, with applications extending to biologically plausible vision models that influence broader neural network developments. In 2021, he received the Bower Award for his pioneering work on the neocognitron.^[20]^[21]