Research Review
SCCN-Core: The Research Value of Adaptive Sparse Neural Systems
SCCN-Core starts from a different definition of machine intelligence: not only the ability to fit a task after
training, but the ability to keep reorganizing perception, state, and decision pathways while signals continue
to arrive. This turns neural computation from a static dense mapping into an adaptive connection process that
can learn during inference rather than waiting for a separate retraining stage.
Research Thesis
The central value of SCCN-Core is that it treats structure as an active part of intelligence. In conventional
ANN systems, the learned function is usually compressed into fixed parameters and then executed as a largely
stable inference graph. SCCN-Core studies a more dynamic form: a neural system whose useful pathways can become
more expressive through use, while irrelevant activity remains sparse and controllable.
The difference begins at the neuron model. A neuron in this system is not only a differentiable activation unit
inside a static matrix pipeline; it is a state-bearing computational element participating in local activity,
competition, reinforcement, and connection change. Learning is therefore not framed as global error propagation
through all layers. It is framed as local, online structural adaptation driven by signal, state, and feedback.
This research direction matters because many real tasks are not closed datasets. Vision, embodied interaction,
financial signals, scientific search, and autonomous agents all operate under changing distributions. A neural
core that can adapt during use is a step toward systems that learn continuously rather than repeatedly returning
to a full retraining cycle.
What Makes It Valuable
Neurons carry state
The neuron is treated as an active state unit, not just a pointwise activation function. Internal state,
firing behavior, and local context become part of how the system remembers, distinguishes, and adapts.
No backpropagation dependency
SCCN-Core does not depend on the traditional backpropagation loop as its learning foundation. It studies
online structural learning, where useful pathways are adjusted through local signal dynamics and feedback.
Sparsity becomes efficiency
Instead of paying the cost of broad dense activation for every signal, the system favors selective
computation. Fewer active pathways can mean lower marginal inference cost and more focused adaptation.
Inference becomes learning time
Inference is not only a readout phase. Each interaction can expose activity, update state, and increase
the future usefulness of pathways that repeatedly support correct perception or decision.
Beyond Traditional ANN Scaling
Dimension Traditional ANN SCCN-Core Research Value
Core assumption Intelligence is mostly captured in a trained parameter function. Intelligence also lives in adaptive connection organization and runtime state.
Neuron model Often abstracted as differentiable units in layered tensor operations. Stateful neural elements participate in firing, competition, and local structural change.
Learning signal Commonly depends on global loss optimization and backpropagation. Uses online signal dynamics, feedback, and local pathway adjustment as the learning substrate.
Computation Dense operations dominate many inference paths. Sparse active pathways make selective computation a first-class research object.
Learning cycle Capability often improves through larger offline training rounds. Capability can improve during inference through interaction, feedback, and structural reuse.
Sample efficiency Classic CNN training usually depends on repeated offline passes over large labeled datasets. In the current standard digit experiment, about 200 standard samples can produce strong accuracy and generalization behavior.
Interpretability Analysis often depends on external probes after training. Internal state and connection change are observable during operation.
Long-term path Progress is tied to more data, more parameters, and more compute. Progress can come from better organization, adaptive routing, and efficient memory of experience.
LeNet-5 as a Reference Point
LeNet-5 is a landmark in neural vision. It proved that convolutional structure, supervised optimization, and
backpropagation could build practical digit recognition systems. Its historical importance is clear: it helped
establish the path that later became modern CNN-based computer vision.
Design dimension LeNet-5 SCCN-Core
Parameter design Parameters are mainly organized as convolution kernels and fully connected weights learned through supervised optimization. The central object is not only a parameter tensor, but a stateful connection field where pathway activity, strength, and reuse matter.
Layer hierarchy Uses a fixed feed-forward hierarchy: convolution, subsampling, convolution, subsampling, and dense classification layers. Uses layered neural fields with observable state, sparse active pathways, and connection dynamics that can adapt during operation.
Neuron role Neurons are primarily activation units inside a differentiable network graph. Neurons are state-bearing elements participating in firing, competition, memory, and local structural change.
Learning mechanism Depends on offline supervised training and backpropagation through the fixed architecture. Emphasizes online structural adaptation through signal activity, state change, and feedback.
Adaptation after deployment Usually requires another training process to absorb new behavior. Can continue adapting while inference happens, making use itself part of learning.
SCCN-Core explores a different value curve. In the current standard digit recognition experiment, the system
can form strong recognition and generalization behavior from about 200 standard samples. The significance is not
merely a smaller training count; it is that learning happens through adaptive state and connection organization,
without relying on the classic offline backpropagation loop as the central mechanism.
This creates a practical research contrast. LeNet-5 demonstrates how a carefully designed ANN can learn visual
structure through batch training. SCCN-Core demonstrates how a sparse stateful neural system can keep learning
while it is being used, turning inference itself into an opportunity for adaptation.
The breakthrough is therefore architectural: the learnable object expands from fixed parameter layers to a
living organization of stateful neurons and sparse connections.
EVM-2 v0.25 Neural Architecture
EVM-2 is the current visual architecture built from this research direction. It keeps the useful inductive
bias of LeNet-style vision, but it does not reduce the model to a simple classifier memory. The forward path is
a real spiking visual hierarchy: a 32x32 retina, sparse S1 filter maps, C1/C2 pooled maps, descriptor-based
prototype memory, a composable assembly vector, and a batch multi-centroid readout. The verified v0.25
checkpoint reaches 92.3% on the 2000 / 300 validation run and 92.0% on the 8000 / 1000 run using MNIST only.
No offline alphabet or symbol dataset is used for these numbers.
The new matrix output experiment keeps that MNIST-only baseline explicit. Digits are the only offline training
distribution, while the live adapter exposes a 94-label ASCII output space for online updates. When a user draws
a letter or symbol and presses Train Drawn, EVM-2 adds that evidence after deployment instead of restarting a
batch training pipeline. This is the important research claim: the visual core can reuse digit-trained strokes,
parts, and detector memories to absorb new symbolic targets through incremental experience.
Module What it does Why it matters
Input / Retina Normalizes each digit into a centered 32x32 event image, giving the model 1024 sparse input neurons. Creates a stable visual field while preserving the event-like sparsity needed by the downstream spiking layers.
S1 filter maps Applies twelve 3x3 orientation and contrast kernels with sparse top-k activation on each map. Acts like a spiking version of early visual cortex: edges, strokes, and polarity-sensitive local signals are extracted before classification.
C1 pooled maps Uses 2x2 max pooling over S1 and keeps the strongest responses per map. Provides local translation tolerance and reduces the active search space without destroying stroke structure.
C2 pooled maps Applies another 2x2 pooling stage and sparse top-k selection. Builds a broader spatial abstraction similar to LeNet's visual hierarchy, but keeps computation sparse and inspectable.
Channel tokens Converts C1 and C2 activity into positioned channel tokens over 4x4 and 2x2 grids. Makes routing explicit: the model can reason about which visual channel fired and where it fired.
Stroke memory Stores prototype descriptors of local C1 stroke patterns. Gives the model reusable visual primitives instead of memorizing whole images or whole classes.
Part memory Composes neighboring stroke descriptors into larger local parts. Moves the representation from isolated edges toward reusable digit components such as corners, loops, and junctions.
Shared detector memory Learns class-agnostic detector prototypes from C1/C2 positioned patches. Creates a shared visual dictionary that can support multiple classes and reduces dependence on class-specific templates.
Composable part memory Builds detector-token pairs and pools their activations into a structured feature block. Turns memory into a compositional layer: evidence can be routed and recombined rather than stored as flat examples.
Assembly vector Concatenates retina summaries, C1/C2 grid statistics, co-activation grids, and prototype-memory activations into a normalized feature vector. Serves as the convergence layer where sparse visual evidence becomes a compact, inspectable representation.
Batch centroid readout Fits 32 centroids per class and scores input with top-2 blended cosine similarity. Provides the current strongest readout for v0.25 while keeping the decision surface transparent and reproducible.
Output logits Returns digit scores for MNIST validation and can expose an 8x12 ASCII matrix for interactive online learning. Separates visual representation from final decision, making it possible to test new symbols without retraining the whole visual stack.
The important design shift is that EVM-2 treats S1/C1/C2, stroke memory, part memory, and detector memory as
real visual layers. They are not labels attached to a nearest-neighbor classifier. They are separate stages that
transform raw visual activity into local descriptors, reusable prototypes, composable parts, and finally class
evidence. This is why the architecture can be studied as a general visual framework rather than only a digit
memory system. The online ASCII adapter is deliberately framed as a post-training generalization test: the base
model sees MNIST offline, then learns user-provided letters and symbols through live updates.
Why This Is a Revolutionary Direction
The revolution is not that SCCN-Core replaces every ANN technique. The deeper shift is that it changes where
intelligence is expected to emerge. Dense ANN scaling asks how far performance can go when parameters, data, and
compute grow. SCCN-Core asks how much intelligence can emerge when structure itself becomes adaptive, sparse,
inspectable, and shaped by use.
This matters because the next generation of intelligent systems will need to work under incomplete information,
changing environments, and long interaction histories. A system that can keep refining its active pathways
offers a different form of leverage: it can improve by organization, not only by accumulation. It also changes
the economics of intelligence: useful computation can concentrate on active structure instead of repeatedly
invoking the full cost profile of dense global updates.
Research and Engineering Value
As a research platform, SCCN-Core creates a bridge between neural computation, online learning, sparse systems,
and state inspection. It gives experiments a visible nervous system: input activity, inference activity, class
competition, energy, separability, active edges, and structural change can be watched together rather than
reduced to a final accuracy number.
As an engineering direction, the value is equally concrete. Adaptive sparse systems suggest smaller active
computation, faster incremental learning, better recovery from noisy samples, and architectures that can be
specialized for domains where data arrives continuously. Since learning can happen online and during inference,
the system can absorb new evidence without waiting for a heavy retraining pipeline. This is especially important
for intelligent systems that must operate close to the user, close to devices, or close to rapidly changing
real-world signals.
Outlook
SCCN-Core points toward a hybrid future. Large ANN models remain powerful representation engines, but adaptive
sparse neural cores can become the layer that learns from local experience, organizes signals over time, and
keeps the system responsive after deployment. The research value is therefore not incremental. It is a path
toward intelligent systems that are lighter, more adaptive, more inspectable, and closer to the dynamics of
living cognition.
研究命题
SCCN-Core 的核心价值在于把结构本身视为智能的一部分。在传统 ANN 系统中,学习结果通常被压缩进固定参数,
再以相对稳定的推理图执行。SCCN-Core 研究的是另一种更动态的形式:有用通路会随着使用变得更有表达力,
而无关活动保持稀疏和可控。
差异首先来自神经元模型。这里的神经元不只是静态矩阵管线中的可微激活单元,而是携带状态、参与局部活动、
竞争、强化和连接变化的计算元素。因此,学习不再被限定为跨层全局误差反向传播,而是由信号、状态和反馈驱动的
局部在线结构适应。
这个方向重要,是因为大量真实任务并不是封闭数据集。视觉、具身交互、金融信号、科学搜索和自主智能体都处在变化分布中。
一个能够在使用中调整内部组织的神经核心,意味着系统可以持续学习,而不是频繁回到完整再训练流程。
我们的研究价值
神经元携带状态
神经元不再只是点式激活函数,而是主动的状态单元。内部状态、触发行为和局部上下文共同参与记忆、区分和适应。
不依赖反向传播
SCCN-Core 不以传统反向传播循环作为学习基础,而是研究在线结构学习:有用通路通过局部信号动态和反馈持续调整。
稀疏成为效率
系统不需要为每个信号支付大范围稠密激活的成本,而是优先进行选择性计算。更少的活跃通路意味着更低的边际推理成本
和更聚焦的适应。
推理过程也是学习时间
推理不只是读出结果。每一次交互都可以暴露活动、更新状态,并提升那些反复支持正确感知或决策的通路在未来的价值。
超越传统 ANN 规模化
维度 传统 ANN SCCN-Core 研究价值
核心假设 智能主要被训练后的参数函数捕获。 智能也存在于自适应连接组织和运行状态中。
神经元模型 通常抽象为分层张量运算中的可微单元。 状态化神经元参与触发、竞争和局部结构变化。
学习信号 常依赖全局损失优化和反向传播。 以在线信号动态、反馈和局部通路调整作为学习底座。
计算方式 许多推理路径由稠密计算主导。 稀疏活跃通路让选择性计算成为核心研究对象。
学习周期 能力提升常依赖更大的离线训练轮次。 能力可以在推理中通过交互、反馈和结构复用持续提升。
样本效率 经典 CNN 训练通常依赖对大规模标注数据集进行多轮离线训练。 在当前标准数字实验中,约 200 个标准样本即可形成很好的准确率和泛化表现。
可解释性 分析往往依赖训练后的外部探针。 内部状态和连接变化可以在运行过程中被观察。
长期路径 进步高度依赖更多数据、更多参数和更多算力。 进步可以来自更好的组织方式、自适应路由和经验的高效记忆。
以 LeNet-5 作为参照
LeNet-5 是神经视觉历史上的重要里程碑。它证明了卷积结构、监督优化和反向传播可以构建实用的数字识别系统,
也奠定了后来 CNN 视觉路线的基础。
设计维度 LeNet-5 SCCN-Core
参数设计 参数主要组织为卷积核和全连接权重,并通过监督优化学习。 核心对象不只是参数张量,而是状态化连接场:通路活动、连接强度和复用价值都参与智能形成。
层次结构 采用固定前馈层级:卷积、下采样、卷积、下采样,再进入稠密分类层。 采用分层神经场:状态可观察、活跃通路稀疏,并且连接动态可在运行中适应。
神经元角色 神经元主要是可微网络图中的激活单元。 神经元是携带状态的计算元素,参与触发、竞争、记忆和局部结构变化。
学习机制 依赖固定架构上的离线监督训练和反向传播。 强调由信号活动、状态变化和反馈驱动的在线结构适应。
部署后适应 通常需要再次训练流程来吸收新行为。 可以在推理发生时继续适应,让使用过程本身成为学习的一部分。
SCCN-Core 探索的是另一条价值曲线。在当前标准数字识别实验中,系统可以通过约 200 个标准样本形成很好的识别和泛化表现。
其意义不只是训练样本数更少,而是学习来自自适应状态和连接组织,并不以经典离线反向传播循环作为中心机制。
这形成了一个清晰的研究对照:LeNet-5 展示了精心设计的 ANN 如何通过批训练学习视觉结构;SCCN-Core 展示了稀疏、
状态化神经系统如何在使用过程中继续学习,让推理本身成为适应机会。
因此,突破点是架构层面的:可学习对象从固定参数层,扩展为状态化神经元和稀疏连接共同组成的动态组织。
EVM-2 v0.25 神经网络架构
EVM-2 是当前基于这条研究方向构建的视觉架构。它保留了 LeNet 风格视觉层级中有价值的归纳偏置,
但不把模型简化成一个分类记忆器。真实前向路径由 32x32 retina、稀疏 S1 滤波图、C1/C2 池化图、
descriptor 原型记忆、可组合 assembly 向量以及 batch multi-centroid readout 组成。v0.25 checkpoint
在 2000 / 300 验证中达到 92.3%,在 8000 / 1000 验证中达到 92.0%。这些数字只来自 MNIST,
没有使用离线字母或符号数据集。
新的矩阵 output 实验刻意保留这个 MNIST-only 基线:离线训练分布只有数字,交互 adapter 则开放
94 个 ASCII 标签组成的输出空间。当用户画出字母或符号并点击 Train Drawn,EVM-2 会在部署后吸收这条证据,
而不是重启一次批量训练流程。这里的研究重点是:数字训练得到的笔画、部件和 detector memory 可以被复用,
通过在线增量经验吸收新的符号目标。
模块 作用 价值
Input / Retina 把数字归一化成居中的 32x32 事件图,对应 1024 个稀疏输入神经元。 建立稳定的视觉场,同时保留后续脉冲层需要的稀疏事件特征。
S1 滤波图 使用 12 个 3x3 方向和对比度核,并在每张图上保留 top-k 稀疏激活。 相当于脉冲版本的早期视觉皮层:先抽取边缘、笔画和极性敏感局部信号,再进入分类证据。
C1 池化图 对 S1 进行 2x2 max pooling,并在每张图上保留最强响应。 带来局部平移容忍度,降低活跃搜索空间,同时不破坏笔画结构。
C2 池化图 再次进行 2x2 pooling 和稀疏 top-k 选择。 形成更大感受野的空间抽象,类似 LeNet 的视觉层级,但计算仍然稀疏、可观察。
Channel tokens 把 C1 和 C2 活动转换成 4x4 与 2x2 网格上的位置化通道 token。 显式表达路由信息:模型可以知道哪个视觉通道在什么位置触发。
Stroke memory 存储 C1 局部笔画模式的原型 descriptor。 让模型复用视觉基元,而不是记忆整张图片或整类模板。
Part memory 把相邻 stroke descriptor 组合成更大的局部部件。 把表示从孤立边缘推进到可复用的数字组成部分,例如角、环、交叉和连接处。
Shared detector memory 从 C1/C2 位置化 patch 中学习类别无关 detector 原型。 形成共享视觉字典,可服务多个类别,降低对类别模板的依赖。
Composable part memory 构建 detector-token pair,并把激活池化成结构化特征块。 把 memory 升级成可组合视觉层:证据可以被路由和重组,而不是以平面样本形式保存。
Assembly vector 拼接 retina 摘要、C1/C2 网格统计、共激活网格和原型记忆激活,形成归一化特征向量。 作为收敛层,把稀疏视觉证据压缩成紧凑、可审计的内部表示。
Batch centroid readout 每类拟合 32 个 centroid,并用 top-2 blended cosine similarity 打分。 这是 v0.25 当前最强 readout,同时保留透明、可复现的决策边界。
Output logits 在 MNIST 验证中输出数字分数,在交互学习中可暴露 8x12 ASCII 矩阵输出。 把视觉表示和最终决策分开,使新符号测试不需要重训整个视觉栈。
关键设计变化在于:EVM-2 把 S1/C1/C2、stroke memory、part memory 和 detector memory 当作真正的视觉层,
而不是贴在 nearest-neighbor 分类器上的名称。它们分别把原始视觉活动转换为局部 descriptor、可复用原型、
可组合部件和类别证据。因此,这个架构可以作为通用视觉框架继续研究,而不仅是一个数字记忆系统。
在线 ASCII adapter 是一个部署后泛化测试:基础模型离线只看 MNIST,然后通过实时更新学习用户提供的字母和符号。
为什么这是革命性方向
SCCN-Core 的革命性不在于替代所有 ANN 技术,而在于改变了智能被期待产生的位置。稠密 ANN 规模化关注参数、
数据和算力增长后性能能走多远;SCCN-Core 关注当结构本身变得自适应、稀疏、可观察并由使用塑造时,
智能能形成怎样的新能力。
这很重要,因为下一代智能系统需要面对不完整信息、变化环境和长期交互历史。一个能够持续优化活跃通路的系统,
拥有另一种杠杆:它可以通过组织获得提升,而不只是通过累积获得提升。它也改变了智能的计算经济性:
有用计算可以集中在活跃结构上,而不是反复承担稠密全局更新的完整成本。
研究与工程价值
作为研究平台,SCCN-Core 在神经计算、在线学习、稀疏系统和状态观察之间建立了桥梁。它让实验拥有一个可见的神经系统:
输入活动、推理活动、类别竞争、能量、可分性、活跃连接和结构变化可以被同时观察,而不是只被压缩成最终准确率。
作为工程方向,它的价值同样具体。自适应稀疏系统意味着更小的活跃计算、更快的增量学习、更好的噪声样本恢复能力,
以及可面向连续数据流进行专门化的架构。由于学习可以在线发生、也可以在推理过程中发生,系统能够吸收新证据,
而不必等待沉重的再训练管线。这对贴近用户、贴近设备、贴近快速变化真实信号的智能系统尤其关键。
研究展望
SCCN-Core 指向一个混合未来。大型 ANN 仍然是强大的表征引擎,而自适应稀疏神经核心可以成为从本地经验中学习、
随时间组织信号、并在部署后保持响应性的层。因此,它的研究价值不是增量优化,而是一条通向更轻量、更自适应、
更可观察、也更接近生命认知动态的智能系统路线。