How Does Facial Recognition Work?
Facial recognition works by detecting a face, mapping its features into a numerical embedding vector, and matching identities by computing similarity in that vector space.
Explore by topic
AI, programming, data, computing, and the internet.
Facial recognition works by detecting a face, mapping its features into a numerical embedding vector, and matching identities by computing similarity in that vector space.
How QR code version controls module count, how mask patterns sculpt the dark/light texture of the data area without changing the data, and how data mode rewrites which characters are converted into bits.
A QR code stores data as a grid of black and white modules wrapped in finder patterns, timing strips, and Reed–Solomon error correction, letting a camera locate the code, sample each module, and reconstruct the message despite damage.
Recommending one title from a catalog of thousands comes down to comparing patterns across millions of viewers, scoring each candidate against your taste, and ranking the results.
Wi-Fi transmits data by encoding digital information onto radio waves on specific frequency bands (2.4 GHz and 5 GHz); a router and device take turns broadcasting on shared channels, using protocols like CSMA/CA to avoid stepping on each other's signals.
That your phone pins down your position by timing signals from multiple satellites and solving for the one point in 3D space where those timing measurements agree.
Spotify's recommendations come from blending collaborative filtering (users similar to you liked these songs), content-based analysis (the audio features of songs you enjoy), and contextual signals — with collaborative filtering doing most of the heavy lifting.
The Friendship Paradox follows directly from averaging over a network: high-degree nodes are sampled more often, so the mean degree of a random friend exceeds the mean degree of a random person.
The birthday paradox connects a counterintuitive probability result to the practical sizing of hash outputs, showing why cryptographers treat 'find any collision' attacks as roughly square-root cheaper than 'find a specific match.'
The story of how neural networks survived a bitter AI winter to become the engine of today's artificial intelligence.
e is the unique base in which a function equals its own derivative, so it is the natural language for any process whose rate of change equals its current size.
i is √(−1), and when you treat it as code rather than algebra its meaning becomes a visible 90° rotation that gives complex numbers their 2D geometry.
Water's hydrogen bonds force molecules into a spacious hexagonal lattice when they freeze, so solid water is less dense than its liquid form — and density is what decides what floats.
That a digital street is a chosen distance metric, and that the clusters you find are the shapes that metric carves out — not an objective truth about the data.
推荐算法通过不断强化你过去的点击偏好来提高准确率,但这种'利用'机制会主动压缩探索空间,导致兴趣范围越来越窄,从而产生熟悉又无聊的矛盾体验。
中文拼音输入法“猜下一个词”的本质,是把“当前上下文片段”和“个人历史词频”同时折算成每个候选词的得分,再按分排序把前几位推上候选栏。
A stress-strain curve plots stress against strain and reveals the four stages of a solid's mechanical response: elastic deformation, yielding, plastic deformation, and fracture, each marked by a recognizable feature on the curve.
The investigation establishes what an experimentally prepared one-bit system must release when it is erased, how the measurement tests the bound, and why this does not imply that today’s computers operate near it.
Landauer's principle: the minimum energy cost of information is set by the thermodynamic cost of reducing physical possibilities, not by the speed or technology of the device.
The hot-spot pattern inside a microwave is a standing electromagnetic wave whose nodes are set by cavity geometry and the dielectric objects inside it; adding a mug redistributes the field so new hot and cold zones appear.
Bending a rod stores elastic strain energy in the tension side of the cross-section, and a sudden fracture releases that energy once stress crosses the material's failure threshold.
The famous 42-move result is a boundary case that exposes how combinatorial explosion separates tractable problems from intractable ones — and why such boundaries define the outer edge of any predictive model.
Walks are translations and turns are rotations, encoded as matrices that act on position vectors; applying a rotation matrix then a translation matrix moves the point differently than translating first then rotating, because matrix multiplication is non-commutative and the rotation literally rotates the direction the translation points in.
Refrigerators don't reverse nature; they pay an energy cost — using a phase-changing refrigerant and a compressor to pump heat uphill from cold to hot, obeying the second law of thermodynamics.