20260518-morning:新增内容
This commit is contained in:
42
concepts/pre-activation-history.md
Normal file
42
concepts/pre-activation-history.md
Normal file
@@ -0,0 +1,42 @@
|
||||
---
|
||||
title: "Pre-Activation History"
|
||||
created: 2026-05-15
|
||||
updated: 2026-05-15
|
||||
type: concept
|
||||
tags: [neural-architecture, temporal-processing, memory]
|
||||
sources: [raw/papers/darlow-ctm-2025.md]
|
||||
---
|
||||
|
||||
# Pre-Activation History
|
||||
|
||||
**前激活历史** 是 CTM 中每个神经元维护的滚动缓冲区,存储最近 M 步的前激活值,供 [[neuron-level-models|NLM]] 处理。
|
||||
|
||||
## 定义
|
||||
|
||||
```
|
||||
A_t = [a_{t-M+1}, a_{t-M+2}, ..., a_t] ∈ R^{D×M}
|
||||
```
|
||||
|
||||
其中 a_t 是 [[synapse-model|Synapse Model]] 的输出(前激活)。A_t 以 FIFO 方式滚动更新。
|
||||
|
||||
对于第 d 个神经元:
|
||||
```
|
||||
A_t^d ∈ R^M → NLM g_{θ_d} → z_{t+1}^d
|
||||
```
|
||||
|
||||
## 为什么重要?
|
||||
|
||||
前激活历史是 NLMs 能够产生**复杂时序动态**的基础:
|
||||
- 没有历史 → NLM 退化为普通逐元素变换
|
||||
- M 较大 → 每个神经元可以检测 M 步的模式
|
||||
- 这类似于卷积的感受野,但在时间维度上且每个神经元独立
|
||||
|
||||
## 超参数 M
|
||||
|
||||
作者发现 M ≈ 10-100 在初始探索中有效:
|
||||
- 太小:缺乏足够的时序上下文
|
||||
- 太大:训练开销增加,可能稀释近期信号
|
||||
|
||||
## 来源
|
||||
|
||||
- [[darlow-ctm-2025|CTM 论文]]
|
||||
Reference in New Issue
Block a user