1. Home
  2. Detail

@janainakize: #magnesio #viraltiktokshop

Jana Kize
Jana Kize
Open In TikTok:
Region: BR
Saturday 15 August 2026 03:14:53 GMT
554
17
2
0

Music

Download

No Watermark .mp4 (10.97MB) No Watermark(HD) .mp4 (10.97MB) Watermark .mp4 (0MB) Music .mp3

Comments

janetekizema
janetekizema :
♥️♥️♥️
2026-08-15 16:07:00
0
janetekizema
janetekizema :
👍👍👍
2026-08-15 16:07:16
1
To see more videos from user @janainakize, please go to the Tikwm homepage.

Other Videos
#duoctaitro #quangcao #thoitrangnu
#duoctaitro #quangcao #thoitrangnu
➡️ Machine Learning from Scratch, part 39: mini batch gradient descent. For each step downhill, the network measures its error on only a small random part of its training data (a mini batch) and adjusts its weights from that. Picture the training data as a stack of flashcards, each with an input and the correct answer: a batch is a small handful of cards, and once the network has gone through the whole stack, that is one epoch. The full batch uses all the data for every step, so its slope is exact. A mini batch uses a small random part, so each step is a quick and noisy guess. Big networks learn from millions of examples, and the exact slope over all of them for every step would cost far too much time and memory, so they train on mini batches (often between 32 and 512 examples). Our own run, nothing downloaded. minibatch.py (15 lines of numpy) makes 1,000 points around a line and trains one neuron with a weight and a bias, so the whole error surface is one bowl we can draw in 3D. Full batch, batch of 100 and batch of 10, learning rate 0.1, same start, 100 steps each. After one pass through the data (same work), the full batch has made one step and its error is 7.869; the batch of 10 has made 100 steps and is at 0.993. After 100 steps each, the full batch sits at the lowest point (0.956) and the batch of 10 at 0.993, with a hundredth of the work. 41 of its 100 steps made the error on all the data worse, and it still ended near the bottom. Honest notes: the error is measured on the same 1,000 points we train on. With a fixed learning rate, the batch of 10 keeps jittering about 4 % above the lowest point. Smaller batches give cheaper but noisier steps, bigger batches smoother but more expensive ones, so batch size is a dial next to the learning rate. #machinelearning #ai #neuralnetwork #python #gradientdescent
➡️ Machine Learning from Scratch, part 39: mini batch gradient descent. For each step downhill, the network measures its error on only a small random part of its training data (a mini batch) and adjusts its weights from that. Picture the training data as a stack of flashcards, each with an input and the correct answer: a batch is a small handful of cards, and once the network has gone through the whole stack, that is one epoch. The full batch uses all the data for every step, so its slope is exact. A mini batch uses a small random part, so each step is a quick and noisy guess. Big networks learn from millions of examples, and the exact slope over all of them for every step would cost far too much time and memory, so they train on mini batches (often between 32 and 512 examples). Our own run, nothing downloaded. minibatch.py (15 lines of numpy) makes 1,000 points around a line and trains one neuron with a weight and a bias, so the whole error surface is one bowl we can draw in 3D. Full batch, batch of 100 and batch of 10, learning rate 0.1, same start, 100 steps each. After one pass through the data (same work), the full batch has made one step and its error is 7.869; the batch of 10 has made 100 steps and is at 0.993. After 100 steps each, the full batch sits at the lowest point (0.956) and the batch of 10 at 0.993, with a hundredth of the work. 41 of its 100 steps made the error on all the data worse, and it still ended near the bottom. Honest notes: the error is measured on the same 1,000 points we train on. With a fixed learning rate, the batch of 10 keeps jittering about 4 % above the lowest point. Smaller batches give cheaper but noisier steps, bigger batches smoother but more expensive ones, so batch size is a dial next to the learning rate. #machinelearning #ai #neuralnetwork #python #gradientdescent
#🌍🔐❤️🥹____💤
#🌍🔐❤️🥹____💤
on top x forget me now | #nino🎧🎶 #xh #lyrics #music #fyp
on top x forget me now | #nino🎧🎶 #xh #lyrics #music #fyp


About

Legal