New answers tagged

0

Correlation between inputs of neural network does not matter. The neural network will learn which input values are associated with the target labels. Depending on the architecture of the neural network, it will learn how different combinations of features can predict the target labels. If the high correlation is predictive, it will help the model learn.


0

One way is to construct the problem as a classification problem, where you break the week into 30m slots. You'll need to carefully define the label of success - after normalizing the channel activity and looking for a high z-score. This way you can still use your desired features. To further improve the model you'd like to use seasonality data and holidays. (...


0

These are great first attempts! However, neural networks are notoriously bad at working with tabular data. You'd might be better served using a traditional ML model (e.g., linear regression, SVM). Regardless of whether you're using a neural net or otherwise, you should normalize/transform your input features and the output feature (i.e., your closing price). ...


0

Selecting the order of training data is often called curriculum learning. Typically, the model is trained on easier to learn data and then successively trained on more difficult to learn data.


0

This sounds as a supervised change point detection task. I believe your approach with the 0/1 dummy variable is fine. Alternatively to the neural network approach you could try out some classification algorithms. You might find helpful this review on change point detection (A Survey of Methods for Time Series Change Point Detection), where some alternative ...


2

As I understand this, what I've described so far is stochastic gradient descent. Without any way to generate a gradient or any mention of using gradients, this is not gradient descent. Your choice of the word "mutate" plus terms like "parents" and "breed" would lead me to believe that you initially want to train your neural ...


0

The only case where I would consider resampling data is when there is a requirement to improve recall for a particular class. Thus the goal would be to force the classifier to predict this class more often, even though it usually means decreasing performance in general. Resampling is an easy method but rarely the optimal one. In general I'd first do an ...


0

Your understanding is not correct. The 2D convolution is indeed a volume convolution. The filter is a tensor of dimensions 7x7x3. The depth of the output equals to the number of filters in the convolution; yours has 1 filter, so the depth of the output is 1.


3

I think it is important to think about the application of that classifier and get the negative class images to be from a similar distribution as will be your application. For example if you want to classify blog images get the negative examples from blogs, if you want to classify facebook photos, get the facebook photos. Note that this should apply also for ...


0

One of the most common ways to compare performance of different models is predictive ability on a hold-out data set. Slice out data that that models did not see during training. Then compare performance of different models on the same dataset using the same evaluation metric. Root mean square error (RMSE) is an example of an evaluation metric.


0

The problem is often framed in the inverse - find bivariate features with high correlation which are then removed from a model to increase interpretability and allow certain models to be fit. This is commonly called multicollinearity.


0

The original problem might be too complex for neuroevolution to learn. You can train in easier version then progressive make the examples more difficult. This is commonly called curriculum learning.


0

If Kernel Least Mean Square (KLMS) is to be applied, it would be useful to most augment traditional machine learning (e.g., linear regression). KLMS are not common in machine learning because kernel methods require more both memory space and computational power. Kernel methods definitely increase the training time while probably not significantly increasing ...


0

The LSA community seems to have first used the word “embedding” in Landauer et al. (1997), in a variant of its mathematical meaning as a mapping from one space or mathematical structure to another. In LSA, the word embedding seems to have described the mapping from the space of sparse count vectors to the latent space of SVD dense vectors. Although the word ...


1

Your manual calculation uses log with base $10$: IN: -(0*np.log10(0.1) + 0.5*np.log10(0.3) + 0.5*np.log10(0.6)) OUT: 0.372363747448347 However, np.log is the natural logarithm (from the documentation): numpy.log numpy.log(x, /, out=None, *, where=True, casting='same_kind', order='K', dtype=None, subok=True[, signature, extobj]) = <ufunc 'log'> ...


0

If the classes are mutually exclusive then ideally you should use categorical cross entropy for the loss function. Binary loss should still work after a fashion though, since it will still encourage incorrect classes to predict low probability, and correct classes high. You do not appear to be renormalising the value of aiOutPossible before using ...


0

The goal of supervised machine learning is automatically learn the features weights to predict target values. If you have target values, you can fit a machine learning algorithm (e.g., a k-nearest neighbors or a neural networks). There is no need to pick the weights yourself, the algorithm will do it for you.


1

For those who are interested in this question, I finally found out some useful metrics that perform well in comparing distributions (and different from KL-divergence): Wasserstein metric, Energy metric, Shannon-Entropy metric, Maximum Mean Discrepancy metric. They are metrics in the sense that they satisfy the properties of a mathematical metric(symmetric ...


1

I think you have an issue with the number of elements in your dataset for analyzing a neural network with 257 features. Consider reducing the number of features. Are all of them mandatory? What is the correlation between them? What is the mutual information between all these variables? Consider adding more data to you dataset. Is that possible? Could you add ...


0

There are several differences: CBOW is not equivalent to matrix factorisation of a PMI matrix, it is SkipGram whose loss function is minimised when $W^\top C=PMI$ (see Goldberg & Levy paper, 2014) SkipGram is not equivalent to PCA of the PMI matrix since the loss function is not the least squares loss (again see Goldberg & Levy, 2014 or "What ...


0

It could also be that the problem is very hard to learn. I've had this and actually after 6 hours of identical outputs in each batch (which happens because the 'average' answer is the easiest to minimize loss), the network finally started learning: Some things I plan on doing to make the learning happen earlier: changes in learning rate using features from ...


0

Classification can totally be done using LSTM. While this is a duplicate of another question, I can provide an example article that demonstrates how this can be done through code.


0

Another approach using Keras Sequence class: class DataGenerator(keras.utils.Sequence): def __init__(self, x_data, y_data, batch_size): self.x, self.y = x_data, y_data self.batch_size = batch_size self.num_batches = np.ceil(len(x_data) / batch_size) self.batch_idx = np.array_split(range(len(x_data)), self.num_batches) def __len__(self): ...


Top 50 recent answers are included