Loc Truong

I build, break, and learn with AI.
Documenting my dev journey into machine learning.

[Loc Truong @loctrg]

Listening Multi Task Learning
https://t.co/blSxNopofg

[Loc Truong @loctrg]

Transfer learning shines when:
Both tasks use the same input type.
You have much more data for Task A than Task B.
Features learned from Task A help Task B.
#LearnAI 🚀🤖

[Loc Truong @loctrg]

Transfer learning 🚀 lets you reuse a pre-trained neural network (like for image recognition) on new tasks (like radiology diagnosis) with less data & faster training! It's a game changer for #DeepLearning in real-world AI! 🔄🤖 #LearnAI

[Loc Truong @loctrg]

After 10+ years of dev, finally learning Vim serious 🤣🤣

[Loc Truong @loctrg]

Why does AI always have to be this massive, generalized, 90%+ accuracy beast?
What if we just trained tiny, fast models built for one user only? Or train model using user data
#learnai

[Loc Truong @loctrg]

Battling high bias? ➡️ Try bigger models, better optimizers (Adam, RMSprop), or new architectures (CNN/RNN).
Fighting variance? ➡️ Add more data, or regularize (L2, dropout, data augmentation).
#LearnAI 🧠🔬 Which is your challenge right now?

[Loc Truong @loctrg]

🚀 When tuning ML models, always compare training/dev error with human-level error! If training error ≫ human error → fix bias. If dev error ≫ train error → fix variance. Human-level error = your north star for realistic model performance! #LearnAI 🧠🐱

[Loc Truong @loctrg]

Just Learnt:
If training error ≫ human error, fix bias;
if dev error ≫ training error, fix variance.
Always compare to human-level for realistic goals! #LearnAI 🐱🤖

[Loc Truong @loctrg]

Not all errors matter equally! 🚩 Even with just 3% error, Algorithm A is riskier if it lets pornographic images slip by. Always tune your loss function to penalize critical mistakes more heavily, not just minimize error rate. #LearnAI

[Loc Truong @loctrg]

Cluely isn't just for hacking meetings — it's how AI understands your screen. A top tool for hands-on learning and AI research. 🧠💻 #LearnAI #Cluely #ResearchTools

[Loc Truong @loctrg]

Listening Soft Max Regression
https://t.co/w7Xt8BKG8o

[Loc Truong @loctrg]

BatchNorm isn't just about speeding up training—it adds noise to each mini-batch by normalizing with its own mean/variance. This acts like a gentle regularizer, a bit like Dropout, making your model more robust! 🚀 #LearnAI

[Loc Truong @loctrg]

🚨 Just learnt - Neural networks can struggle due to internal covariate shift! Batch normalization (BN) helps stabilize and speed up training by normalizing activations layer by layer. 🔥 #LearnAI #MachineLearning

[Loc Truong @loctrg]

🚀 Mini-batch & BatchNorm in deep learning: Using mini-batches boosts efficiency, and BatchNorm replaces bias b with learnable β for better normalization & training stability! #LearnAI

[Loc Truong @loctrg]

🚀 Implementing Batch Norm in Neural Nets:

Compute μ & σ² from the batch

Normalize: zᵢ → (zᵢ - μ) / √(σ² + ε)

Scale & shift: ẑᵢ = γ * z_norm + β
→ γ, β = learnable params
It stabilizes training & speeds it up.
Normalize, then learn! ⚡🧠

#DeepLearning #AI #ML

[Loc Truong @loctrg]

Normalizing inputs in neural networks accelerates learning by ensuring all features have similar scales. This reduces training time and improves convergence. Always normalize your data before training for optimal results! 🚀 #LearnAI

[Loc Truong @loctrg]

Listening https://t.co/TUYoWmsiL2
#Deeplearning #LearnAI

[Loc Truong @loctrg]

🚀 Hyperparameter tuning: Don’t “babysit” one model like a panda 🐼. Run many experiments in parallel like caviar—test lots of models, find the best fast! 🔥 #LearnAI #deeplearning

[Loc Truong @loctrg]

#Learningenglish by listen #Deeplearning
How to pick Hyperparameters:
https://t.co/H1fK4b972E

Listen your favorite, not just for study

[Loc Truong @loctrg]

💡Hyperparameter tuning tip:
Sampling uniformly isn’t always smart.
Learning_rate: use log scale
→ 10^r, with r ~ U[-4, 0]
Beta for moving average:
→ beta = 1 - 10^r, with r ~ U[-3, -1]
Why?
🔍 Sensitivity is nonlinear
📉 You waste resources if you don’t scale right
#LearnAI