18

There is no direct way to find the optimal number of them: people empirically try and see (e.g., using cross-validation). The most common search techniques are random, manual, and grid searches. There exist more advanced techniques such as Gaussian processes, e.g. Optimizing Neural Network Hyperparameters with Gaussian Processes for Dialog Act ...


14

It seems to me that you already understand the shortcomings of ReLUs and sigmoids (like dead neurons in the case of plain ReLU). You may want to look at ELU (exponential linear units) and SELU (self-normalising version of ELU). Under some mild assumptions, the latter has the nice property of self-normalisation, which mitigates the problem of vanishing and ...


13

I have read somewhere on the web (I lost the reference) that the number of units (or neurons) in a hidden layer should be a power of 2 because it helps the learning algorithm to converge faster. I would quite like to see a reference to this suggestion, in case it has been misunderstood. As far as I know, there is no such effect in normal neural networks. In ...


7

I have an idea to find the optimal number of hidden neurons required in a neural network but I'm not sure how accurate it is. It's a complete non-starter, and there is a no such calculation possible in the general case (real-valued inputs to a neural network). Even with one input neuron it is not possible. That is because even with one input, the output ...


7

For a more intelligent approach than random or exhaustive searches, you could try a genetic algorithm such as NEAT http://nn.cs.utexas.edu/?neat. However, this has no guarantee to find a global optima, it is simply an optimization algorithm based on performance and is therefore vulnerable to getting stuck in a local optima.


6

You could say that NAS fits into the domain of Meta Learning or Meta Machine learning. I've pulled the NAS papers from my notes, this is a collection of papers/lectures that I personally found very interesting. It's sorted in rough chronological descending order, and *** means influential / must read. Quoc V. Le and Barret Zoph are to good authors on the ...


6

There is a technique called Pruning in neural networks, which is used just for this same purpose. The pruning is done on the number of hidden layers. The process is very similar to the pruning process of decision trees. The pruning process is done as follows: Train a large, densely connected, network with a standard training algorithm Examine the trained ...


5

Paper Szegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the inception architecture for computer vision[J]. arXiv preprint arXiv:1512.00567, 2015. gives some general design principles: Avoid representational bottlenecks, especially early in the network; Balance the width and depth of the network. Optimal performance of the network can be reached ...


5

The difference between the validation and test set in my opinion should be explained in this way: the validation set is meant to be used multiple times. the test set is meant to be used only once. I think that the misunderstanding here arise because machine learning is mostly taught focusing only on a specific part of a large pipeline, which is the model ...


5

In reinforcement learning, exploration has a specific meaning, which is in contrast with the meaning of exploitation, hence the so-called exploration-exploitation dilemma (or trade-off). You explore when you decide to visit states that you have not yet visited or to take actions you have not yet taken. On the other hand, you exploit when you decide to take ...


4

Yes. Usually you would use cross validation to avoid overfitting during parameter tuning. If your dataset is large enough, and you don't try too many parameter combinations, this will work well, because to "get lucky" and overfit, a parameter combination will need to work very well on many variations of the problem, which is less likely than working well on ...


4

The book Deep Learning by Goodfellow, Bengio, and Courville says (Sec 8.3.3, p 292 in my copy) states that Unfortunately, in the stochastic gradient case, Nesterov momentum does not improve the rate of convergence. I'm not sure why this is, but the theoretical advantage depends on a convex problem, and from this, it sounds like the practical advantage ...


4

What is Bayesian optimization? Introduction Bayesian optimization (BO) is an optimization technique used to model an unknown (usually continuous) function $f: \mathbb{R}^d \rightarrow Y$, where typically $d \leq 20$, so it can be used to solve regression and classification problems, where you want to find an approximation of $f$. In this sense, BO is ...


4

A surrogate model is a simplified model. It is a mapping $y_S=f_S(x)$ that approximates the original model $y=f(x)$, in a given domain, reasonably well. Source: Engineering Design via Surrogate Modelling: A Practical Guide In the context of Bayesian optimization, one wants to optimize a function $y=f(x)$ which is expensive (very time consuming) to evaluate, ...


4

The number of dimensions is a hyperparameter of your model, and you should do a hyperparameter search, like with any other parameters. There's also a tradeoff between dimension and training speed, so it should be small enough to be trainable in a reasonable time.


3

The correct number of child processes will depend on the hardware available to you. Simplifying a bit, child processes can be in one of two states: waiting for memory or disk access, or running. If your problem fits nicely in your computers' memory, then processes will spend almost all of their time running. If it's too big for memory, they will ...


3

***Take my answer as a side note to that given by cantordust: If one can verify that an activation function perform well in some cases, that good behavior often extrapolates to other problems. Thus, by testing activation functions on a few different problems, one can often infer how well (or badly) it will perform on most problems. The following video shows ...


3

In principle, yes, you can also evolve the genetic algorithm (or, in general, evolutionary algorithm), i.e. you can evolve its operations (such as the mutation and cross-over) and hyper-parameters (such as the size of the population or mutation rate). For example, you could use genetic programming to evolve the cross-over operation of a genetic algorithm. ...


3

In general, it is definitely very computationally expensive, so an exhaustive search is not performed in practice. However, there are some recent approaches for determining whether the architecture is "fine" without training the neural network first - by looking at the covariance matrix after forwarding the data, for example, in a recent paper ...


3

Personally, I would choose the following two as the most important: epsilon: When using an epsilon-greedy policy, epsilon determines how often the agent should explore and how often it should exploit. Balancing exploration and exploitation is crucial for the success of the learning agent. Too little exploration might not teach anything to the agent and too ...


2

Take a look at this article. It give tools to actually understand what your filters have learn and show what you can do next to optimize your hyper-parameters. Also check more recent articles that seek to provide interpretations of what NN learn.


2

How to find the best configuration for an algorithm is an open research question in AI. The topic in general is known as `hyper-parameter optimization' and there are a range of possible methods: One of the most popular is IRace, but other possibilities include: Spearmint: uses wrappers in Matlab or Python. It uses MongoDb, and Bayesian optimisation ...


2

You can take a look at bayesian hyperparameter optimization as a general method of optimizing loss (or anything) as a function of the hyperparameters. But note that in general the deeper your network the better, so optimizing loss as a function of number of layers isn't a very fun thing to do. Grid search and a bit of common sense (as learnt by seeing many ...


2

Such a system can and does solve the problem of hyperparameter tuning. Google's AutoML does this. Here is another example that uses a Genetic Algorithm to breed new neural network structures. AutoML has been shown to outperform humans in the rate that it improves network designs. It seems to favour Residual Network style topologies.


2

Yes, your understanding of the hidden state is correct. But the size of the hidden state is a hyperparameter that needs to found by trial-and-error. There is no closed-form formula or solution which links the size of the hidden state and the problem at hand. But, there are some rules of thumb like to start out with the size of the hidden state to be a power ...


2

The theory behind hyper-parameter optimization (HPO) is not well developed. Nonetheless, there are several hyper-parameter optimization approaches, such as Bayesian optimization (using Gaussian processes), random search, grid search, genetic algorithms, etc. See, for example, the paper Hyperparameter Search in Machine Learning (2015), which attempts to ...


2

After you've computed $h^{1}_{optimal}$ the only thing you can be sure is that this is the best (assuming constrained case) value of $h^1$ (with respect to some model performance metric) given your initial values for $h^2, ..., h^n$. If you change a bit any of $h^2, ..., h^n$ you're no longer certain that the value $h^1$ you found is the optimal one. So yes, ...


2

If you used your five $X_{test}$ sets multiple times (to measure the average AUC) to decide on the best set of hyperparameters (i.e. optimizer, learning rate, batch size, dropout, activation) then yes, you successfully conducted hyper-parameter optimization. However, the AUC you received for the best set of hyperparameters found (by manual tuning) is not ...


2

Simply stated, you use your validation set to regularize your model for unseen data. Test data is completely unseen data, on which you evaluate your model. Various validation strategies are used to improve your model to perform for unseen data. So strategies like k-fold cross-validation are used. Also, the validation set helps you in tuning your ...


Only top voted, non community-wiki answers of a minimum length are eligible