#is-it-einstein — Neural networks

Is it Einstein? Two faces go through the scanner of a network that learned only Einstein: Haythem → NO, Einstein → YES. Watching is free; touching is Premium.

What you'll play with

Is it Einstein? 🧠 The film starts on its own, nothing to do. A face-recognition network has learned one single face: Einstein. We show it two heads, one after the other, and it decides: YES or NO.
Scan 1: the photo of Haythem, the founder. Crop, then his pixels really go through 16 3×3 filters (Sobel, Laplacian, emboss…), fly toward the 512-number embedding, cross the small 64-neuron head… and the stamp drops: NO. Nice smile, wrong hair.
Scan 2: a caricature of Einstein sticking his tongue out. Same pipeline, same network, nothing retrained — only the embedding changes. The stamp: YES! That is the whole idea of transfer learning: FaceNet stays frozen, we only compare the end.
Watching is free, and the film loops. Touching the network — /photo to enrol your face instead of Einstein's, /threshold 0.85 for a stricter judge, /seed 12 for another draw, or simply /pause and /scan — is Premium. That is why the box below is greyed out. Next: #cnn-filters to see convolutions up close, #transfer-learning to understand why we don't retrain everything.

Channel commands

/scan <haythem|einstein|next> — Jump to a scan (haythem, einstein, or the next one) and play it from the start.
/pause — Pause the film.
/play — Resume the film.
/speed <0.5|1|2|4> — Change the playback speed.
/replay — Restart the current scan from the first frame.
/subtitles — Hide or show the subtitles.
/autoplay — Chain the scans, or stop at the end of the current one.
/reset — Put the film back to its original settings.
/photo — Enrol your face instead of Einstein (soon: real inference).
/threshold <0.4..0.95> — Threshold above which the verdict says YES.
/seed <1..999> — Another draw of the activations (same pipeline, other texture).

Glossary

Cropping (MTCNN): Before identifying, we find the face and crop it. MTCNN is a small cascade detector: first "is there a head?", then the landmarks (eyes, nose, mouth) to straighten it.
FaceNet: A network trained once and for all to bring two photos of the same person together and push two strangers apart. We freeze it: no retraining. We only read its vector of 512 numbers.
512-d embedding: The face compressed into 512 numbers. Two portraits of Einstein give close vectors (high cosine); Einstein and Haythem do not. It is a fingerprint, not a stored photo.
Classification head: A tiny network (here 64 neurons then 2 outputs) learned on top of frozen FaceNet. It is the one saying YES / NO. Changing person means relearning this head, not the whole network.
Transfer learning: Reuse a network trained on millions of faces, and learn only the last layer on one person. Less data, less compute, and the norm in vision.
Decision threshold: Above the threshold we say YES; below, NO. Raising it makes the judge stricter (fewer false acceptances, more refusals).
Caricature, not photo: Einstein's face here is a generated caricature, not an archive photograph: the famous shots remain under copyright. The network does not care — edges, eyes, a moustache.

Other channels in Neural networks

#neuron — Let us break a neuron apart: inputs, weights, sum, activation.
#dropout — Dropout regularization: the best friend of deep networks.
#activation — Why an activation? ReLU, sigmoid, tanh, Leaky ReLU, GELU and the vanishing gradient.
#gradient-descent — Loss is a landscape. The gradient says which way is up, so we step the other way.
#backpropagation — The computation graph replayed backwards: each node receives ∂L/∂(itself) and the chain rule does the rest.
#overfitting — A big model on few data points: the decision boundary twists until it memorizes the noise.
#cnn-filters — An image is a grid of numbers. A 3×3 filter slides over it, multiplies, adds up: that is a convolution.
#is-it-einstein — Is it Einstein? Two faces go through the scanner of a network that learned only Einstein: Haythem → NO, Einstein → YES. Watching is free; touching is Premium.
#embeddings-3d — A word becomes a vector: close in space = close in meaning, and you can do math on them.