Gate: features for objects nobody showed it (AI off)
Why it matters Pretraining is only worth it if the features carry over to things the pretraining never saw. Here you test that yourself, with your own tools, on six new kinds of object.
Gate rules: AI tutor off. The new objects (star, bar, arrow, L-shape, hourglass, drop) never appeared in the 4000 photos the encoders were pretrained on. You have 400 labelled photos of them to train a probe and 500 to test it.
Compare three candidates with your own linear probe from lesson 3 (copy it in: the gate does not provide one): the random encoder and the two recorded SimCLR encoders. Their features are computed for you in F. Then name the best and say how far it beats the random encoder.
The check recomputes everything with the course's probe and passes when your numbers agree, your best is the real best, and it beats the random encoder by at least 8 points.
Words and symbols in this lesson
- linear probe
- a linear classifier trained on frozen features: a test of what the features already say
- SimCLR
- self-supervised training with two views of each picture and InfoNCE (Chen et al. 2020)
Fill accuracy, best and margin.
AI off: no tutor, hints or worked solution on this one. Your own thinking only.
Spotted a problem in this lesson? Report it · Ask about this lesson
This exercise opens soon
Finish the lesson before it and this one unlocks. The whole course is free, no account needed, and you can read this lesson right now.
This browser is blocking cookies for learn.floati.life, so lessons you finished here cannot open in order. Allow cookies for this site and reload.