Explore by topic
AI, programming, data, computing, and the internet.
A capacity-controlled threshold separates approximation from interpolation: once a network has enough parameters to fit every training point exactly, the optimization problem shifts from finding shared patterns to finding any solution, and the learned function changes character accordingly.
A comparison-based detector only flags events that differ from a stored baseline, so it catches new visual anomalies but misses anything the baseline already allows, anything outside its field of view, and anything the camera cannot see.
A real-time traffic system detects road changes by comparing live movement signals against the recent past, not by watching the road itself.
地铁 App 用从起点按乘车时间逐层向外扩展的贪心搜索，每个站点只保留一次最优到达方式，因此几秒内就能锁定全程最短的换乘路线。
道路ネットワークは静的な地図ではなく、変化し続ける「動的な地図」であり、封鎖や事故といったイベントは「必要な場所だけ書き換える」ことでシステム全体に伝わっていく、という仕組み。
动态路网在突发封路场景下，主要通过更早的封路感知、更准的绕行代价和更稳的派单重算，把派单+路径的端到端响应时间压缩到一个有限区间；而静态路网只有在司机实地反馈后才开始反应，且代价估计偏置。
A node is locked the first time it is extracted as the minimum from the priority queue, because every still-unlocked node is at least as far away as the just-locked one, so no future relaxation can produce a shorter path to it.
逐层扩展按路径长度递增的顺序发现路径，因此第一次命中终点时，所走路径长度即为全局最短。
动态路网在实时系统中通过'事件触发→增量更新→局部重算'三步机制维持准确，外卖派单和城市规划都用同一套思想应对突发事件。
Navigation runs a layer-by-layer greedy search from the start point, so every intersection is only locked in once with its shortest arrival — and the answer falls out the instant the destination is reached.
导航用从起点逐层向外扩展的贪心搜索，只检查每一个路口最优的一次到达方式，因此几秒内即可锁定全城最短路线。
Interpolation forces the network to pass exactly through every training point, producing high-frequency oscillations between them; approximation sacrifices exact fit for a smoother function that captures the underlying trend and behaves reasonably on unseen inputs.
Dimension mismatch between queries and keys distorts the attention score distribution — specifically inflating its variance — and the sqrt(d_k) normalization in scaled dot-product attention exists precisely to keep that distribution stable regardless of dimension.
The nearest‑centroid rule defines cluster membership by which centroid is closest, so cluster edges are the straight perpendicular bisectors between centroids — independent of the actual spread, shape, or density of the points.
Big-O describes how running time grows with input size, but it hides three things that can dominate real performance: the multiplicative constant, the cost of each primitive operation, and how the data interacts with memory hierarchy. Algorithm choice should weight all three, not just the asymptotic class.
How keys are converted into numeric values, how those values are squeezed into a fixed array range with modulo arithmetic, and why collisions are an expected consequence.
A network that can fit unseen noise isn't memorizing exceptions — it's because high-capacity models can interpolate the training data exactly, which forces the learned function to be smooth enough to generalize on real patterns.
Attention compares a query vector to a bank of key vectors via a dot product, and the resulting scores weight how much of each value vector is read — a similarity operation over learned representations, not over surface words.
A cluster's edge is the set of points whose nearest centroid switches, and the centroid itself is the mean of its assigned points.
Big-O ignores constants because it describes how runtime scales with input size, and any fixed multiplier becomes irrelevant once the input grows.
Hash table lookup is constant time because the hash function maps a key directly to a slot index, turning a search across n entries into a single computed jump — provided the table is sized and the hash is uniform enough that collisions stay rare and amortized constant.
解释筛法通过预先标记倍数来实现批量淘汰的机制，并说明它在候选质数生成中相对于试除法的效率优势。
Gradient noise and parameter noise shape optimization through fundamentally different mechanisms: gradient noise produces an implicit regularization that scales with learning rate, while parameter noise produces an explicit perturbation that scales with weight magnitude.
A reusable design pattern for AI-allowed assignments that separates automatable production from non-delegable inquiry, defended by a worked example and a boundary check.