User shimao

103 votes

Accepted

Impractical question: is it possible to find the regression line using a ruler and compass?

answered Feb 1, 2021 at 2:07

61 votes

Accepted

Dropout makes performance worse

answered Aug 23, 2017 at 1:58

53 votes

Accepted

Dice-coefficient loss function vs cross-entropy

answered May 3, 2018 at 15:35

49 votes

Why not just dump the neural networks and deep learning?

answered Aug 11, 2017 at 4:22

48 votes

Accepted

What is the "capacity" of a machine learning model?

answered Nov 8, 2017 at 6:49

33 votes

Accepted

Neural networks vs everything else

answered May 3, 2018 at 16:41

32 votes

Accepted

When to "add" layers and when to "concatenate" in neural networks?

answered Aug 7, 2018 at 0:42

30 votes

Accepted

What are attention mechanisms exactly?

answered May 10, 2018 at 4:40

29 votes

Accepted

What does "vanilla" mean?

answered Jul 30, 2017 at 1:21

24 votes

Accepted

Difference in log base for cross entropy calcuation

answered Jul 30, 2017 at 2:58

23 votes

Accepted

How can a statistician who has the data for a non-normal distribution guess better than one who only has the mean?

answered Jul 18, 2021 at 19:24

17 votes

Accepted

How do Variational Auto Encoders backprop past the sampling step

answered Apr 26, 2018 at 3:23

17 votes

Can a GAN be used for data augmentation?

answered Apr 17, 2018 at 20:59

17 votes

How do backbone and head architecture work in Mask R-CNN?

answered Apr 4, 2019 at 5:08

16 votes

Log probabilities in reference to softmax classifier

answered Jul 8, 2017 at 0:05

16 votes

Accepted

When generating samples using variational autoencoder, we decode samples from $N(0,1)$ instead of $\mu + \sigma N(0,1)$

answered Oct 1, 2018 at 20:00

15 votes

Accepted

Loss not decreasing but performance is improving

answered Nov 15, 2017 at 13:20

15 votes

Overview over Reinforcement Learning Algorithms

answered Jan 24, 2018 at 13:45

15 votes

Why do we use Gaussian distributions in Variational Autoencoder?

answered Apr 11, 2019 at 23:31

15 votes

Accepted

When calculating self-attention for Transformer ML architectures, why do we need both a key and a query weight matrix?

answered Mar 25, 2021 at 4:27

14 votes

Lack of Batch Normalization Before Last Fully Connected Layer

answered Apr 4, 2019 at 4:54

14 votes

What should I do when my neural network doesn't generalize well?

answered Sep 9, 2018 at 1:52

13 votes

Why do we use loss functions to estimate a model instead of evaluation metrics like accuracy?

answered Nov 28, 2018 at 20:02

13 votes

Can't deep learning models now be said to be interpretable? Are nodes features?

answered May 4, 2018 at 18:33

13 votes

Is Proximal Policy Optimization (PPO) an on-policy reinforcement learning algorithm?

answered Nov 29, 2019 at 22:31

13 votes

Accepted

Foundation models : Is it a new paradigm for statistics and machine learning?

answered Sep 19, 2021 at 15:27

12 votes

Accepted

Do multiple deep descents exist?

answered Jul 10, 2020 at 16:48

12 votes

Why care so much about expected utility?

answered Jan 14, 2019 at 20:01

11 votes

Accepted

Wasserstein Loss is very sensitive to model architecture

answered Mar 27, 2018 at 12:06

11 votes

Accepted

Is manually tuning learning rate during training redundant with optimization methods like Adam?

answered Jul 7, 2017 at 5:58

Stack Exchange Network

665 Answers

Impractical question: is it possible to find the regression line using a ruler and compass?

Dropout makes performance worse

Dice-coefficient loss function vs cross-entropy

Why not just dump the neural networks and deep learning?

What is the "capacity" of a machine learning model?

Neural networks vs everything else

When to "add" layers and when to "concatenate" in neural networks?

What are attention mechanisms exactly?

What does "vanilla" mean?

Difference in log base for cross entropy calcuation

How can a statistician who has the data for a non-normal distribution guess better than one who only has the mean?

How do Variational Auto Encoders backprop past the sampling step

Can a GAN be used for data augmentation?

How do backbone and head architecture work in Mask R-CNN?

Log probabilities in reference to softmax classifier

When generating samples using variational autoencoder, we decode samples from $N(0,1)$ instead of $\mu + \sigma N(0,1)$

Loss not decreasing but performance is improving

Overview over Reinforcement Learning Algorithms

Why do we use Gaussian distributions in Variational Autoencoder?

When calculating self-attention for Transformer ML architectures, why do we need both a key and a query weight matrix?

Lack of Batch Normalization Before Last Fully Connected Layer

What should I do when my neural network doesn't generalize well?

Why do we use loss functions to estimate a model instead of evaluation metrics like accuracy?

Can't deep learning models now be said to be interpretable? Are nodes features?

Is Proximal Policy Optimization (PPO) an on-policy reinforcement learning algorithm?

Foundation models : Is it a new paradigm for statistics and machine learning?

Do multiple deep descents exist?

Why care so much about expected utility?

Wasserstein Loss is very sensitive to model architecture

Is manually tuning learning rate during training redundant with optimization methods like Adam?