Lesson · Hinglish · ⏱ 8 min padhne mein
⛰️ Gradients aur gradient descent: models behtar kaise hote hain
EnglishHinglish
🏠 Asli zindagi ki misaal
Socho aap raat ko dhundh wali pahaadi pe ho aur neeche pahunchna hai. Door tak dikh nahi raha, isliye paon ke neeche dhalaan mehsoos karte ho aur neeche ki taraf chhota kadam lete ho. Phir yahi repeat karte ho. Dhalaan gradient hai, har kadam ek update hai, aur kadam ka size learning rate hai. Neural networks bilkul aise hi seekhte hain.
💡 Aasaan shabdon mein
Model ka 'loss' ek number hai jo batata hai ki uski predictions kitni galat hain. Gradient batata hai ki loss kam karne ke liye har parameter ko kis direction mein aur kitna hilna chahiye. Gradient descent baar-baar saare parameters ko us dhalaan ki taraf thoda-thoda khiskaata hai. Learning rate bahut bada ho to aap aage nikal jaate ho aur idhar-udhar uchhalte ho; bahut chhota ho to training mein bahut time lagta hai. Backpropagation woh efficient tareeka hai, calculus ke chain rule se, jo crore parameters ke gradients ek saath nikaalta hai. Adam jaise optimisers gradient descent ke smart versions hain.
🎬 Dekho kaise kaam karta hai
Training ek ball jaisi hai jo loss curve pe neeche lud rahi hai: har step dhalaan pe neeche jaata hai, aur learning rate step ka size tay karta hai. Woh sabse kam loss pe rukti hai.
🪜 Step by step
- Random weights se shuru karo; model ke predictions kharab hote hain.
- Loss function se naapo kitne kharab (jaise average squared error).
- Gradient nikalo: har weight ke liye, kis direction mein loss badhta hai aur kitni tezi se.
- Har weight ko ulti direction mein chhota step do: weight = weight − learning_rate × gradient.
- Bahut saare mini batches pe repeat karo jab tak validation data pe loss sudharna band na ho.
🔑 Zaroori shabd
- Loss
- Number jo batata hai model kitna galat hai. Kam behtar hai.
- Learning rate
- Har update pe kitna bada kadam lena hai.
- Backpropagation
- Woh tareeka jo network ki saari layers ke through gradients nikaalta hai.
👀 Example dekho
# Find w so that w * 3 ≈ 12 (answer: 4)
w, lr = 0.0, 0.05
for step in range(20):
pred = w * 3
loss = (pred - 12) ** 2
grad = 2 * (pred - 12) * 3 # slope of the loss
w -= lr * grad # step downhill
print(round(w, 3)) # → 4.0Training ka poora idea 7 lines mein: predict karo, galti naapo, dhalaan pe neeche kadam lo, repeat karo.
⚠️ Aam galtiyaan jinse bachna hai
- Learning rate bahut zyaada: loss uchhalta hai ya phat jaata hai. Bahut kam: training rengti hai. Kuch values try karo.
- Loss curve plot na karna. Training chal rahi hai ya nahi dekhne ka sabse tez tarika yahi hai.
✅ Jaldi se check karo
Training loss bahut upar-neeche ho raha hai aur kabhi settle nahi hota. Shayad fix kya hai?
- Learning rate badhao
- Test set delete karo
- Learning rate ghatao
Jawaab dekho
C. Learning rate ghatao: Bahut bade kadam neeche wale point se aage nikal jaate hain. Chhote kadam settle ho jaate hain.
🙋 Seekhne wale yeh poochhte hain
Kya gradients haath se nikalne padenge?
Nahi. PyTorch jaise frameworks inhe khud nikaalte hain (autograd). Idea samajhna zaroori hai.
Adam kya hai?
Ek optimiser jo recent gradients se har weight ka step size adjust karta hai. Neural networks ke liye mazboot default hai.
📌 Yaad rakho
Gradient = dhalaan kis taraf hai. Seekhna = loss ki pahaadi pe bahut saare chhote kadam neeche.
📖 Is lesson ke shabd (AI dictionary)
Poora course, aapke roadmap ke hisaab se
Progress tracking, quick checks aur step-by-step order ke saath. Yeh lesson in courses ka hissa hai:
Course mein seekho