Skip to main content

New answers tagged

0 votes

local minima vs saddle points in deep learning

<Global and Local Minima> A global minimum is the globally minimal point whose gradient is zero but not a local minimum. A local minimum is the locally minimal point whose gradient is zero but ...
Super Kai - Kazuya Ito's user avatar
1 vote
Accepted

Is it appropriate to use KL Divergence as a loss function for a 1x3 regression model?

KL divergence is defined as the number of bits required to convert one distribution into another. The lower bound value is zero and is achieved when the distributions under observation are identical. ...
xabash's user avatar
  • 86
7 votes
Accepted

How do regression loss functions like MAE and MSE work although they remove the plus/minus sign?

You're absolutely right that these loss functions seemingly discard the sign information by taking the absolute value. The key insight, however, is that while the loss function itself discards the ...
Ansh Tandon's user avatar
3 votes

How do regression loss functions like MAE and MSE work although they remove the plus/minus sign?

But the loss itself has removed this plus/minus information. From the loss value, you can't tell which direction to go in. However, we also have the loss gradient, which does include directional ...
MuhammedYunus's user avatar
0 votes

Interpretation of PPO learning curve, value loss, policy loss

Hard to say much without knowing the specifics of the problem setting, reward function. Seems like trying out a couple of different hyper-parameters may help you get better performance, for example ...
xabash's user avatar
  • 86
0 votes

What is the "fast version" of ZFNet referenced in SPPNet and Faster R-CNN papers?

The "fast" version uses all convolutional layers and fully connected layers in comparison to the other versions, in which they remove several different layers. This version is likely called ...
Oxbowerce's user avatar
  • 7,592
0 votes

Batch Normalization vs Layer Normalization

Batch Normalization layer: can normalize each feature within a mini-batch of samples(layers) to be similar scale to accelerate(speed up) training. is unstable with small batch sizes, then it leads to ...
Super Kai - Kazuya Ito's user avatar
0 votes

wierd neural network approache

I think you don't need to use a neural network for the seating arrangement. Seating guests is a known problem in operational research modelled as a Quadratic Multiknapsack Problem. You can find some ...
Tomasz Witkowski's user avatar
0 votes

Average loss is 0 when training dataset with darknet yolov4

follow this link to configure your configuration parameters rightly: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects
MichaelMM's user avatar
0 votes
Accepted

Predicted output is only 0s

A ROCAUC of $0.7487$ is decent, certainly enough to say that the model is finding some way to distinguish between the two categories (at least in-sample). The fact that the categorical predictions are ...
Dave's user avatar
  • 3,960
0 votes

Conv2d with time series

It works the same as in image-to-image NN. It recognizes patterns in the data and is trained to forecast future patterns from the past patterns. It can also support related data and attributes. Amazon ...
Yair Beer's user avatar
2 votes

Noob question - which NLP/deep learning technique shoud I use

Noob question - which NLP/deep learning technique shoud I use String matching + regex.
Franck Dernoncourt's user avatar
2 votes

Points to remember when embarking on an organization-wide turn to AI solutions

I don't have great references for this answer, but I noticed it from my own experience of starting a team at a large corporation. I think it is often difficult for people at the executive level to ...
healthydata's user avatar
0 votes
Accepted

Linear warmup always result in a significant drop in accuracy

As you've pointed out, the train loss goes up at the same time the validation loss drops. To me, this suggests that the learning rate increase has led to an instability, as the model momentarily fails ...
MuhammedYunus's user avatar
1 vote

balancing and imbalancing in supervised anomaly detection probelm

Well, I'll put forth my answer from a research perspective and from my experience in anomaly detection Truth is, it is hard to evaluate an anomaly detection model under supervised condition due to the ...
Sahan Dissanayaka's user avatar
0 votes

Dicussion for X-vectors

512 is the size of the output of the layer. The input size of the frame2 is a result of splicing together frames at the three time points, $t-2, \ t, \ t+2$, each of output size 512 and $512\cdot3=...
mcas's user avatar
  • 1
2 votes
Accepted

Diffusion Models: Conditioning on Time vs. Noise Level

They should have phrased it a bit differently... Each timestep t is associated with some noise level for a particular schedule, but the noise levels sampled during training do not have to be used when ...
Parlance's user avatar
0 votes

How to update first layer weights?

I think your biggest error is updating weights with the model's error and not the error's derivatives. It's not backpropagation. To update a parameter of the network you must compute the derivative of ...
Tomasz Witkowski's user avatar
-3 votes

Is it legal to use a model found on github for a personal project and uploading the personal project onto github?

Yes -- I do this all the time, even in commercial contexts. For a personal project you are fine. Happy building! Even my AI product says it's fine :). https://us.idyllic.app/question/Is-it-legal-to-...
MaximilianP's user avatar
0 votes

How to deal with variable length input in the architecture of deep learning methods?

If you are able to crop consistently, that's a great step to do to throw away useless information. For example, if you are expecting a single bright object on a dark background. If this is not ...
Zwerchhau's user avatar
  • 131

Top 50 recent answers are included