🎛 权重与偏置：两个最忙的配角🎛 Weights & Biases: the Two Busiest Bit-Parts
上一片里，门卫把分数 z 打理得服服帖帖。可 z 本身是谁算出来的？片场里站着两位最忙的配角——每条线、每个神经元都离不开它们。Last leaf, the gatekeeper tidied up the score z. But who actually computes z? Two busiest bit-parts stand behind every wire and every neuron.
第 1 步：两位配角登场Step 1: Meet the two bit-parts
先认识它们——一个管音量，一个管基础心情。Say hello — one turns the volume, one sets the starting mood.
糖卡和盐卡各自过一道音量旋钮：旋钮把信号拧大或拧小，这就是权重 w。两路信号汇进小机器人，再加上一点点「保底分」——偏置 b（它的基础心情），最后合成得分 z。写成一个式子：z = w₁·糖 + w₂·盐 + b。The sugar card and the salt card each pass through a volume knob that dials the signal up or down — that's a weight w. The signals merge into the little robot, and a pinch of "starting mood" — the bias b — is added on top. Together they make the score z: z = w₁·sugar + w₂·salt + b.
那你可能要问：拧这两个旋钮，到底会改变什么？You might ask: what do these two knobs actually change?
第 2 步：一个管倾斜，一个管高低Step 2: One tilts, one shifts
把数据画成散点，像一座往上爬的小山坡。权重 w 决定直线的倾斜：w 变大，线变陡；偏置 b 决定整条线的高低：b 变大，线整体上移。只调这两个数，让线穿过点群中间——这一步就叫「拟合」。Plot the data as dots and it looks like a little hill climbing up. The weight w sets the tilt: bigger w, steeper line. The bias b sets the height: bigger b lifts the whole line. Tune just these two numbers until the line threads through the cloud — that step is called "fitting".
会调一条线只是热身——真实网络里，这样的配角有多少个？One line is just a warm-up — how many of these bit-parts live in a real network?
第 3 步：真实网络里，它们无处不在Step 3: Everywhere in a real network
每条连线都有自己的权重（就像线上的小旋钮），每个神经元都带自己的偏置（脚下的小基座）。一个几百万参数的网络，主角就是它们；训练做的唯一一件事，就是一点点微调这几百万到几十亿个数字，让整个班底的配合越来越好。Every connection owns a weight (the little knob on the line), and every neuron carries a bias (the base under its feet). A network with millions of parameters is mostly these two — and training does exactly one thing: gently tweak those millions-to-billions of numbers until the whole cast performs better.
🎮 你来当调线师（1 分钟）🎮 Your turn: line fitter (1 minute)
道理讲完了。4 颗靶点等着你，两个滑块：一个管倾斜（权重），一个管高低（偏置）。看看几秒能通关。Theory done. Four targets await, two sliders: one tilts (weight), one shifts (bias). See how fast you can clear it.
一句话记住它：权重管「倾斜」（乘法），偏置管「高低」（加法）；网络里它们以百万计，训练就是微调它们。Remember it in one line: weights set the tilt (a multiply), biases set the height (an add); a network holds millions of them, and training just tweaks them.
权重 = 每条信号线自己的音量旋钮（乘法）Weight = each signal line's own volume knob (a multiply)偏置 = 每个神经元自己的保底分（加法）：z = w·x + bBias = each neuron's own head start (an add): z = w·x + b训练 = 微调这几百万到几十亿个数字，没有别的Training = tweaking those millions-to-billions of numbers — nothing else
内容参考 OpenStax 等公开教材，多来源核对 · AI 生成、人工审核 · 发现错误欢迎指正，帮这片叶子长得更好。Based on OpenStax and other open textbooks, cross-checked across sources · AI-generated, human-reviewed · Spotted a mistake? Tell us — help this leaf grow.
Back to tree
