Im personally studying theories of neural network and got some questions.
In many books and references, for activation function of hidden layer, hyper-tangent functions were used.
Books came up with really simple reason that linear combinations of tanh functions can describe nearly all shape of functions with given error.
But, there came a question.
Is this a real reason why tanh function is used?
If then, is it the only reason why tanh function is used?
if then, is tanh function the only function that can do that?
if not, what is the real reason?..
I stock here keep thinking... please help me out of this mental(?...) trap!
Most of time tanh is quickly converge than sigmoid and logistic function, and performs better accuracy [1]. However, recently rectified linear unit (ReLU) is proposed by Hinton [2] which shows ReLU train six times fast than tanh [3] to reach same training error. And you can refer to [4] to see what benefits ReLU provides.
Accordining to about 2 years machine learning experience. I want to share some stratrgies the most paper used and my experience about computer vision.
Normalizing input is very important
Normalizing well could get better performance and converge quickly. Most of time we will subtract mean value to make input mean to be zero to prevent weights change same directions so that converge slowly [5] .Recently google also points that phenomenon as internal covariate shift out when training deep learning, and they proposed batch normalization [6] so as to normalize each vector having zero mean and unit variance.
More data more accuracy
More training data could generize feature space well and prevent overfitting. In computer vision if training data is not enough, most of used skill to increase training dataset is data argumentation and synthesis training data.
Choosing a good activation function allows training better and efficiently.
ReLU nonlinear acitivation worked better and performed state-of-art results in deep learning and MLP. Moreover, it has some benefits e.g. simple to implementation and cheaper computation in back-propagation to efficiently train more deep neural net. However, ReLU will get zero gradient and do not train when the unit is zero active. Hence some modified ReLUs are proposed e.g. Leaky ReLU, and Noise ReLU, and most popular method is PReLU [7] proposed by Microsoft which generalized the traditional recitifed unit.
Others
choose large initial learning rate if it will not oscillate or diverge so as to find a better global minimum.
shuffling data
In truth both tanh and logistic functions can be used. The idea is that you can map any real number ( [-Inf, Inf] ) to a number between [-1 1] or [0 1] for the tanh and logistic respectively. In this way, it can be shown that a combination of such functions can approximate any non-linear function.
Now regarding the preference for the tanh over the logistic function is that the first is symmetric regarding the 0 while the second is not. This makes the second one more prone to saturation of the later layers, making training more difficult.
To add up to the the already existing answer, the preference for symmetry around 0 isn't just a matter of esthetics. An excellent text by LeCun et al "Efficient BackProp" shows in great details why it is a good idea that the input, output and hidden layers have mean values of 0 and standard deviation of 1.
Update in attempt to appease commenters: based purely on observation, rather than the theory that is covered above, Tanh and ReLU activation functions are more performant than sigmoid. Sigmoid also seems to be more prone to local optima, or a least extended 'flat line' issues. For example, try limiting the number of features to force logic into network nodes in XOR and sigmoid rarely succeeds whereas Tanh and ReLU have more success.
Tanh seems maybe slower than ReLU for many of the given examples, but produces more natural looking fits for the data using only linear inputs, as you describe. For example a circle vs a square/hexagon thing.
http://playground.tensorflow.org/ <- this site is a fantastic visualisation of activation functions and other parameters to neural network. Not a direct answer to your question but the tool 'provides intuition' as Andrew Ng would say.
Many of the answers here describe why tanh (i.e. (1 - e^2x) / (1 + e^2x)) is preferable to the sigmoid/logistic function (1 / (1 + e^-x)), but it should noted that there is a good reason why these are the two most common alternatives that should be understood, which is that during training of an MLP using the back propagation algorithm, the algorithm requires the value of the derivative of the activation function at the point of activation of each node in the network. While this could generally be calculated for most plausible activation functions (except those with discontinuities, which is a bit of a problem for those), doing so often requires expensive computations and/or storing additional data (e.g. the value of input to the activation function, which is not otherwise required after the output of each node is calculated). Tanh and the logistic function, however, both have very simple and efficient calculations for their derivatives that can be calculated from the output of the functions; i.e. if the node's weighted sum of inputs is v and its output is u, we need to know du/dv which can be calculated from u rather than the more traditional v: for tanh it is 1 - u^2 and for the logistic function it is u * (1 - u). This fact makes these two functions more efficient to use in a back propagation network than most alternatives, so a compelling reason would usually be required to deviate from them.
In theory I in accord with above responses. In my experience, some problems have a preference for sigmoid rather than tanh, probably due to the nature of these problems (since there are non-linear effects, is difficult understand why).
Given a problem, I generally optimize networks using a genetic algorithm. The activation function of each element of the population is choosen randonm between a set of possibilities (sigmoid, tanh, linear, ...). For a 30% of problems of classification, best element found by genetic algorithm has sigmoid as activation function.
In deep learning the ReLU has become the activation function of choice because the math is much simpler from sigmoid activation functions such as tanh or logit, especially if you have many layers. To assign weights using backpropagation, you normally calculate the gradient of the loss function and apply the chain rule for hidden layers, meaning you need the derivative of the activation functions. ReLU is a ramp function where you have a flat part where the derivative is 0, and a skewed part where the derivative is 1. This makes the math really easy. If you use the hyperbolic tangent you might run into the fading gradient problem, meaning if x is smaller than -2 or bigger than 2, the derivative gets really small and your network might not converge, or you might end up having a dead neuron that does not fire anymore.
Related
I have a question. I watched a really detailed tutorial on implementing an artificial neural network in C++. And now I have more than a basic understanding of how a neural network works and how to actually program and train one.
So in the tutorial a hyperbolic tangent was used for calculating outputs, and obviously its derivative for calculating gradients. However I wanted to move on to a different function. Specifically Leaky RELU (to avoid dying neurons).
My question is, it specifies that this activation function should be used for the hidden layers only. For the output layers a different function should be used (either a softmax or a linear regression function). In the tutorial the guy taught the neural network to be an XOR processor. So is this a classification problem or a regression problem?
I tried to google the difference between the two, but I can't quite grasp the category for the XOR processor. Is it a classification or a regression problem?
So I implemented the Leaky RELU function and its derivative but I don't know whether I should use a softmax or a regression function for the output layer.
Also for recalculating the output gradients I use the Leaky RELU's derivative(for now) but in this case should I use the softmax's/regression derivative as well?
Thanks in advance.
I tried to google the difference between the two, but I can't quite grasp the category for the XOR processor. Is it a classification or a regression problem?
In short, classification is for discrete target, regression is for continuous target. If it were a floating point operation, you had a regression problem. But here the result of XOR is 0 or 1, so it's a binary classification (already suggested by Sid). You should use a softmax layer (or a sigmoid function, which works particularly for 2 classes). Note that the output will be a vector of probabilities, i.e. real valued, which is used to choose the discrete target class.
Also for recalculating the output gradients I use the Leaky RELU's derivative(for now) but in this case should I use the softmax's/regression derivative as well?
Correct. For the output layer you'll need a cross-entropy loss function, which corresponds to the softmax layer, and it's derivative for the backward pass.
If there will be hidden layers that still use Leaky ReLu, you'll also need Leaky ReLu's derivative accordingly, for these particular layers.
Highly recommend this post on backpropagation details.
Most examples of neural networks for classification tasks I've seen use the a softmax layer as output activation function. Normally, the other hidden units use a sigmoid, tanh, or ReLu function as activation function. Using the softmax function here would - as far as I know - work out mathematically too.
What are the theoretical justifications for not using the softmax function as hidden layer activation functions?
Are there any publications about this, something to quote?
I haven't found any publications about why using softmax as an activation in a hidden layer is not the best idea (except Quora question which you probably have already read) but I will try to explain why it is not the best idea to use it in this case :
1. Variables independence : a lot of regularization and effort is put to keep your variables independent, uncorrelated and quite sparse. If you use softmax layer as a hidden layer - then you will keep all your nodes (hidden variables) linearly dependent which may result in many problems and poor generalization.
2. Training issues : try to imagine that to make your network working better you have to make a part of activations from your hidden layer a little bit lower. Then - automaticaly you are making rest of them to have mean activation on a higher level which might in fact increase the error and harm your training phase.
3. Mathematical issues : by creating constrains on activations of your model you decrease the expressive power of your model without any logical explaination. The strive for having all activations the same is not worth it in my opinion.
4. Batch normalization does it better : one may consider the fact that constant mean output from a network may be useful for training. But on the other hand a technique called Batch Normalization has been already proven to work better, whereas it was reported that setting softmax as activation function in hidden layer may decrease the accuracy and the speed of learning.
Actually, Softmax functions are already used deep within neural networks, in certain cases, when dealing with differentiable memory and with attention mechanisms!
Softmax layers can be used within neural networks such as in Neural Turing Machines (NTM) and an improvement of those which are Differentiable Neural Computer (DNC).
To summarize, those architectures are RNNs/LSTMs which have been modified to contain a differentiable (neural) memory matrix which is possible to write and access through time steps.
Quickly explained, the softmax function here enables a normalization of a fetch of the memory and other similar quirks for content-based addressing of the memory. About that, I really liked this article which illustrates the operations in an NTM and other recent RNN architectures with interactive figures.
Moreover, Softmax is used in attention mechanisms for, say, machine translation, such as in this paper. There, the Softmax enables a normalization of the places to where attention is distributed in order to "softly" retain the maximal place to pay attention to: that is, to also pay a little bit of attention to elsewhere in a soft manner. However, this could be considered like to be a mini-neural network that deals with attention, within the big one, as explained in the paper. Therefore, it could be debated whether or not Softmax is used only at the end of neural networks.
Hope it helps!
Edit - More recently, it's even possible to see Neural Machine Translation (NMT) models where only attention (with softmax) is used, without any RNN nor CNN: http://nlp.seas.harvard.edu/2018/04/03/attention.html
Use a softmax activation wherever you want to model a multinomial distribution. This may be (usually) an output layer y, but can also be an intermediate layer, say a multinomial latent variable z. As mentioned in this thread for outputs {o_i}, sum({o_i}) = 1 is a linear dependency, which is intentional at this layer. Additional layers may provide desired sparsity and/or feature independence downstream.
Page 198 of Deep Learning (Goodfellow, Bengio, Courville)
Any time we wish to represent a probability distribution over a discrete variable with n possible values, we may use the softmax function. This can be seen as a generalization of the sigmoid function which was used to represent a probability
distribution over a binary variable.
Softmax functions are most often used as the output of a classifier, to represent the probability distribution over n different classes. More rarely, softmax functions can be used inside the model itself, if we wish the model to choose between one of n different options for some internal variable.
Softmax function is used for the output layer only (at least in most cases) to ensure that the sum of the components of output vector is equal to 1 (for clarity see the formula of softmax cost function). This also implies what is the probability of occurrence of each component (class) of the output and hence sum of the probabilities(or output components) is equal to 1.
Softmax function is one of the most important output function used in deep learning within the neural networks (see Understanding Softmax in minute by Uniqtech). The Softmax function is apply where there are three or more classes of outcomes. The softmax formula takes the e raised to the exponent score of each value score and devide it by the sum of e raised the exponent scores values. For example, if I know the Logit scores of these four classes to be: [3.00, 2.0, 1.00, 0.10], in order to obtain the probabilities outputs, the softmax function can be apply as follows:
import numpy as np
def softmax(x):
z = np.exp(x - np.max(x))
return z / z.sum()
scores = [3.00, 2.0, 1.00, 0.10]
print(softmax(scores))
Output: probabilities (p) = 0.642 0.236 0.087 0.035
The sum of all probabilities (p) = 0.642 + 0.236 + 0.087 + 0.035 = 1.00. You can try to substitute any value you know in the above scores, and you will get a different values. The sum of all the values or probabilities will be equal to one. That’s makes sense, because the sum of all probability is equal to one, thereby turning Logit scores to probability scores, so that we can predict better. Finally, the softmax output, can help us to understand and interpret Multinomial Logit Model. If you like the thoughts, please leave your comments below.
While digging through the topic of neural networks and how to efficiently train them, I came across the method of using very simple activation functions, such as the rectified linear unit (ReLU), instead of the classic smooth sigmoids. The ReLU-function is not differentiable at the origin, so according to my understanding the backpropagation algorithm (BPA) is not suitable for training a neural network with ReLUs, since the chain rule of multivariable calculus refers to smooth functions only.
However, none of the papers about using ReLUs that I read address this issue. ReLUs seem to be very effective and seem to be used virtually everywhere while not causing any unexpected behavior. Can somebody explain to me why ReLUs can be trained at all via the backpropagation algorithm?
To understand how backpropagation is even possible with functions like ReLU you need to understand what is the most important property of derivative that makes backpropagation algorithm works so well. This property is that :
f(x) ~ f(x0) + f'(x0)(x - x0)
If you treat x0 as actual value of your parameter at the moment - you can tell (knowing value of a cost function and it's derivative) how the cost function will behave when you change your parameters a little bit. This is most crucial thing in backpropagation.
Because of the fact that computing cost function is crucial for a cost computation - you will need your cost function to satisfy the property stated above. It's easy to check that ReLU satisfy this property everywhere except a small neighbourhood of 0. And this is the only problem with ReLU - the fact that we cannot use this property when we are close to 0.
To overcome that you may choose the value of ReLU derivative in 0 to either 1 or 0. On the other hand most of researchers don't treat this problem as serious simply because of the fact, that being close to 0 during ReLU computations is relatively rare.
From the above - of course - from the pure mathematical point of view it's not plausible to use ReLU with backpropagation algorithm. On the other hand - in practice it usually doesn't make any difference that it has this weird behaviour around 0.
I am finding a very hard time to visualize how the activation function actually manages to classify non-linearly separable training data sets.
Why does the activation function (e.g tanh function) work for non-linear cases? What exactly happens mathematically when the activation function projects the input to output? What separates training samples of different classes, and how does this work if one had to plot this process graphically?
I've tried looking for numerous sources, but what exactly makes the activation function actually work for classifying training samples in a neural network, I just cannot grasp easily and would like to be able to picture this in my mind.
Mathematical result behind neural networks is Universal Approximation Theorem. Basically, sigmoidal functions (those which saturate on both ends, like tanh) are smooth almost-piecewise-constant approximators. The more neurons you have – the better your approximation is.
This picture was taked from this article: A visual proof that neural nets can compute any function. Make sure to check that article, it has other examples and interactive applets.
NNs actually, at each level, create new features by distorting input space. Non-linear functions allow you to change "curvature" of target function, so further layers have chance to make it linear-separable. If there were no non-linear functions, any combination of linear function is still linear, thus no benefit from multi-layerness. As a graphical example consider
this animation
This pictures where taken from this article. Also check out that cool visualization applet.
Activation functions have very little to do with classifying non-linearly separable sets of data.
Activation functions are used as a way to normalize signals at every step in your neural network. They typically have an infinite domain and a finite range. Tanh, for example, has a domain of (-∞,∞) and a range of (-1,1). The sigmoid function maps the same domain to (0,1).
You can think of this as a way of enforcing equality across all of your learned features at a given neural layer (a.k.a. feature scaling). Since the input domain is not known before hand it's not as simple as regular feature scaling (for linear regression) and thusly activation functions must be used. The effects of the activation function are compensated for when computing errors during back-propagation.
Back-propagation is a process that applies error to the neural network. You can think of this as a positive reward for the neurons that contributed to the correct classification and a negative reward for the neurons that contributed to an incorrect classification. This contribution is often known as the gradient of the neural network. The gradient is, effectively, a multi-variable derivative.
When back-propagating the error, each individual neuron's contribution to the gradient is the activations function's derivative at the input value for that neuron. Sigmoid is a particularly interesting function because its derivative is extremely cheap to compute. Specifically s'(x) = 1 - s(x); it was designed this way.
Here is an example image (found by google image searching: neural network classification) that demonstrates how a neural network might be superimposed on top of your data set:
I hope that gives you a relatively clear idea of how neural networks might classify non-linearly separable datasets.
I understand neural networks with any number of hidden layers can approximate nonlinear functions, however, can it approximate:
f(x) = x^2
I can't think of how it could. It seems like a very obvious limitation of neural networks that can potentially limit what it can do. For example, because of this limitation, neural networks probably can't properly approximate many functions used in statistics like Exponential Moving Average, or even variance.
Speaking of moving average, can recurrent neural networks properly approximate that? I understand how a feedforward neural network or even a single linear neuron can output a moving average using the sliding window technique, but how would recurrent neural networks do it without X amount of hidden layers (X being the moving average size)?
Also, let us assume we don't know the original function f, which happens to get the average of the last 500 inputs, and then output a 1 if it's higher than 3, and 0 if it's not. But for a second, pretend we don't know that, it's a black box.
How would a recurrent neural network approximate that? We would first need to know how many timesteps it should have, which we don't. Perhaps a LSTM network could, but even then, what if it's not a simple moving average, it's an exponential moving average? I don't think even LSTM can do it.
Even worse still, what if f(x,x1) that we are trying to learn is simply
f(x,x1) = x * x1
That seems very simple and straightforward. Can a neural network learn it? I don't see how.
Am I missing something huge here or are machine learning algorithms extremely limited? Are there other learning techniques besides neural networks that can actually do any of this?
The key point to understand is compact:
Neural networks (as any other approximation structure like, polynomials, splines, or Radial Basis Functions) can approximate any continuous function only within a compact set.
In other words the theory states that, given:
A continuous function f(x),
A finite range for the input x, [a,b], and
A desired approximation accuracy ε>0,
then there exists a neural network that approximates f(x) with an approximation error less than ε, everywhere within [a,b].
Regarding your example of f(x) = x2, yes you can approximate it with a neural network within any finite range: [-1,1], [0, 1000], etc. To visualise this, imagine that you approximate f(x) within [-1,1] with a Step Function. Can you do it on paper? Note that if you make the steps narrow enough you can achieve any desired accuracy. The way neural networks approximate f(x) is not much different than this.
But again, there is no neural network (or any other approximation structure) with a finite number of parameters that can approximate f(x) = x2 for all x in [-∞, +∞].
The question is very legitimate and unfortunately many of the answers show how little practitioners seem to know about the theory of neural networks. The only rigorous theorem that exists about the ability of neural networks to approximate different kinds of functions is the Universal Approximation Theorem.
The UAT states that any continuous function on a compact domain can be approximated by a neural network with only one hidden layer provided the activation functions used are BOUNDED, continuous and monotonically increasing. Now, a finite sum of bounded functions is bounded by definition.
A polynomial is not bounded so the best we can do is provide a neural network approximation of that polynomial over a compact subset of R^n. Outside of this compact subset, the approximation will fail miserably as the polynomial will grow without bound. In other words, the neural network will work well on the training set but will not generalize!
The question is neither off-topic nor does it represent the OP's opinion.
I am not sure why there is such a visceral reaction, I think it is a legitimate question that is hard to find by googling it, even though I think it is widely appreciated and repeated outloud. I think in this case you are looking for the actually citations showing that a neural net can approximate any function. This recent paper explains it nicely, in my opinion. They also cite the original paper by Barron from 1993 that proved a less general result. The conclusion: a two-layer neural network can represent any bounded degree polynomial, under certain (seemingly non-restrictive) conditions.
Just in case the link does not work, it is called "Learning Polynomials with Neural Networks" by Andoni et al., 2014.
I understand neural networks with any number of hidden layers can approximate nonlinear functions, however, can it approximate:
f(x) = x^2
The only way I can make sense of that question is that you're talking about extrapolation. So e.g. given training samples in the range -1 < x < +1 can a neural network learn the right values for x > 100? Is that what you mean?
If you had prior knowledge, that the functions you're trying to approximate are likely to be low-order polynomials (or any other set of functions), then you could surely build a neural network that can represent these functions, and extrapolate x^2 everywhere.
If you don't have prior knowledge, things are a bit more difficult: There are infinitely many smooth functions that fit x^2 in the range -1..+1 perfectly, and there's no good reason why we would expect x^2 to give better predictions than any other function. In other words: If we had no prior knowledge about the function we're trying to learn, why would we want to learn x -> x^2? In the realm of artificial training sets, x^2 might be a likely function, but in the real world, it probably isn't.
To give an example: Let's say the temperature on Monday (t=0) is 0°, on Tuesday it's 1°, on Wednesday it's 4°. We have no reason to believe temperatures behave like low-order polynomials, so we wouldn't want to infer from that data that the temperature next Monday will probably be around 49°.
Also, let us assume we don't know the original function f, which happens to get the average of the last 500 inputs, and then output a 1 if it's higher than 3, and 0 if it's not. But for a second, pretend we don't know that, it's a black box.
How would a recurrent neural network approximate that?
I think that's two questions: First, can a neural network represent that function? I.e. is there a set of weights that would give exactly that behavior? It obviously depends on the network architecture, but I think we can come up with architectures that can represent (or at least closely approximate) this kind of function.
Question two: Can it learn this function, given enough training samples? Well, if your learning algorithm doesn't get stuck in a local minimum, sure: If you have enough training samples, any set of weights that doesn't approximate your function gives a training error greater that 0, while a set of weights that fit the function you're trying to learn has a training error=0. So if you find a global optimum, the network must fit the function.
A network can learn x|->x * x if it has a neuron that calculates x * x. Or more generally, a node that calculates x**p and learns p. These aren't commonly used, but the statement that "no neural network can learn..." is too strong.
A network with ReLUs and a linear output layer can learn x|->2*x, even on an unbounded range of x values. The error will be unbounded, but the proportional error will be bounded. Any function learnt by such a network is piecewise linear, and in particular asymptotically linear.
However, there is a risk with ReLUs: once a ReLU is off for all training examples it ceases learning. With a large domain, it will turn on for some possible test examples, and give an erroneous result. So ReLUs are only a good choice if test cases are likely to be within the convex hull of the training set. This is easier to guarantee if the dimensionality is low. One work around is to prefer LeakyReLU.
One other issue: how many neurons do you need to achieve the approximation you want? Each ReLU or LeakyReLU implements a single change of gradient. So the number needed depends on the maximum absolute value of the second differential of the objective function, divided by the maximum error to be tolerated.
There are theoretical limitations of Neural Networks. No neural network can ever learn the function f(x) = x*x
Nor can it learn an infinite number of other functions, unless you assume the impractical:
1- an infinite number of training examples
2- an infinite number of units
3- an infinite amount of time to converge
NNs are good in learning low-level pattern recognition problems (signals that in the end have some statistical pattern that can be represented by some "continuous" function!), but that's it!
No more!
Here's a hint:
Try to build a NN that takes n+1 data inputs (x0, x1, x2, ... xn) and it will return true (or 1) if (2 * x0) is in the rest of the sequence. And, good luck.
Infinite functions especially those that are recursive cannot be learned. They just are!