I've been trying to understand how Lightgbm handless L1 loses (MAE, MAPE, HUBER)
According to this article, the gain during a split should depend only on the first and second derivatives of the loss function. This is due to the fact that Lightgbm uses a second order approximation to the loss function and consequently we can approximate the loss as follows
For L1 losses however, the absolute value of the gradient of the loss is constant and its hessian 0. I've also read that to deal with this, for loss functions with hessian = 0 we should rather use 1 as the Hessian:
"For these objective function with first_order_gradient is constant, LightGBM has a special treatment for them: (...) it will use the constant gradient for the tree structure learning, but use the residual for the leaf output calculation, with percentile function, e.g. 50% for MAE. This solution is from sklearn, and is proven to work in many benchmarks."
However, even using constant hessian doesn't make sense to me: if for instance when using MAE the gradient is the sign of the error, the squared gradient doesn't give us information. Does it mean that when the gradient is constant, LightGbm does not use the second order approximation, and defaults to traditional gradient boosting?
On the other hand, when reading about GOSS boosting the original lightgbm paper
for the GOSS boosting strategy, the authors consider the square of the sum of the gradients. I see the same problem as above: if the gradient of the MAE is the sign of the error, how does taking the square of the gradient reflect a gain? Does it mean that also GOSS won't work with loss functions with constant gradient?
Thanks in advance,
I've asked this in the Lightgbm repo and got this answer:
Before this version, we use the second-order approximation, but its performance actually is not good.
And we switch back to 1) use first-order gradient to find split point; 2) then use the median of residuals for leaf outputs, as shown in the above code.
So it seems Lightgbm will treat the already implemented L1 losses using gradient descent. For custom loss functions, it will still try to do the 2nd order approx.
Related
I understand that both LinearRegression class and SGDRegressor class from scikit-learn performs linear regression. However, only SGDRegressor uses Gradient Descent as the optimization algorithm.
Then what is the optimization algorithm used by LinearRegression, and what are the other significant differences between these two classes?
LinearRegression always uses the least-squares as a loss function.
For SGDRegressor you can specify a loss function and it uses Stochastic Gradient Descent (SGD) to fit. For SGD you run the training set one data point at a time and update the parameters according to the error gradient.
In simple words - you can train SGDRegressor on the training dataset, that does not fit into RAM. Also, you can update the SGDRegressor model with a new batch of data without retraining on the whole dataset.
To understand the algorithm used by LinearRegression, we must have in mind that there is (in favorable cases) an analytical solution (with a formula) to find the coefficients which minimize the least squares:
theta = (X'X)^(-1)X'Y (1)
where X' is the the transpose matrix of X.
In the case of non-invertibility, the inverse can be replaced by the Moore-Penrose pseudo-inverse calculated using "singular value decomposition" (SVD). And even in the case of invertibility, the SVD method is faster and more stable than applying the formula (1).
PS - No LaTeX (MathJaX) in Stackoverflow ???
--
Pierre (from France)
I'm studying Markov Random Fields, and, apparently, inference in MRF is hard / computationally expensive. Specifically, Kevin Murphy's book Machine Learning: A Probabilistic Perspective says the following:
"In the first term, we fix y to its observed values; this is sometimes called the clamped term. In the second term, y is free; this is sometimes called the unclamped term or contrastive term. Note that computing the unclamped term requires inference in the model, and this must be done once per gradient step. This makes training undirected graphical models harder than training directed graphical models."
Why are we performing inference here? I understand that we're summing over all y's, which seems expensive, but I don't see where we're actually estimating any parameters. Wikipedia also talks about inference, but only talks about calculating the conditional distribution, and needing to sum over all non-specified nodes.. but.. that's not what we're doing here, is it?
Alternatively, any have good intuition on why inference in MRF is difficult?
Sources:
Chapter 19 of ML:PP: https://www.cs.ubc.ca/~murphyk/MLbook/pml-print3-ch19.pdf
Specific section seen below
When training your CRF, you want to estimate your parameters, \theta.
In order to do this, you can differentiate your loss function (Equation 19.38) with respect to \theta, set it to 0, and solve for \theta.
You can't analytically solve the equation for \theta if you do this though. You can, however, minimise Equation 19.38 by gradient descent. Since the loss function is convex, it is guaranteed that gradient descent will get you the globally optimal solution when it converges.
Equation 19.41 is the actual gradient which you need to compute in order to be able to do gradient descent. The first term is easy (and computationally cheap) to compute as you are summing up over the observed values of y. However, the second term requires you to do inference. In this term, you are not summing up over the observed value of y as in the first term. Instead, you need to compute the configuration of y (inference), and then calculate the value of the potential function under this configuration.
The Keras implementation of dropout references this paper.
The following excerpt is from that paper:
The idea is to use a single neural net at test time without dropout.
The weights of this network are scaled-down versions of the trained
weights. If a unit is retained with probability p during training, the
outgoing weights of that unit are multiplied by p at test time as
shown in Figure 2.
The Keras documentation mentions that dropout is only used at train time, and the following line from the Dropout implementation
x = K.in_train_phase(K.dropout(x, level=self.p), x)
seems to indicate that indeed outputs from layers are simply passed along during test time.
Further, I cannot find code which scales down the weights after training is complete as the paper suggests. My understanding is this scaling step is fundamentally necessary to make dropout work, since it is equivalent to taking the expected output of intermediate layers in an ensemble of "subnetworks." Without it, the computation can no longer be considered sampling from this ensemble of "subnetworks."
My question, then, is where is this scaling effect of dropout implemented in Keras, if at all?
Update 1: Ok, so Keras uses inverted dropout, though it is called dropout in the Keras documentation and code. The link http://cs231n.github.io/neural-networks-2/#reg doesn't seem to indicate that the two are equivalent. Nor does the answer at https://stats.stackexchange.com/questions/205932/dropout-scaling-the-activation-versus-inverting-the-dropout. I can see that they do similar things, but I have yet to see anyone say they are exactly the same. I think they are not.
So a new question: Are dropout and inverted dropout equivalent? To be clear, I'm looking for mathematical justification for saying they are or aren't.
Yes. It is implemented properly. From the time when Dropout was invented - folks improved it also from the implementation point of view. Keras is using one of this techniques. It's called inverted dropout and you may read about it here.
UPDATE:
To be honest - in the strict mathematical sense this two approaches are not equivalent. In inverted case you are multiplying every hidden activation by a reciprocal of dropout parameter. But due to that derivative is linear it is equivalent to multiplying all gradient by the same factor. To overcome this difference you must set different learning weight then. From this point of view this approaches differ. But from a practical point view - this approaches are equivalent because:
If you use a method which automatically sets the learning rate (like RMSProp or Adagrad) - it will make almost no change in algorithm.
If you use a method where you set your learning rate automatically - you must take into account the stochastic nature of dropout and that due to the fact that some neurons will be turned off during training phase (what will not happen during test / evaluation phase) - you must to rescale your learning rate in order to overcome this difference. Probability theory gives us the best rescalling factor - and it is a reciprocal of dropout parameter which makes the expected value of a loss function gradient length the same in both train and test / eval phases.
Of course - both points above are about inverted dropout technique.
Excerpted from the original Dropout paper (Section 10):
In this paper, we described dropout as a method where we retain units with probability p at training time and scale down the weights by multiplying them by a factor of p at test time. Another way to achieve the same effect is to scale up the retained activations by multiplying by 1/p at training time and not modifying the weights at test time. These methods are equivalent with appropriate scaling of the learning rate and weight initializations at each layer.
Note though, that while keras's dropout layer is implemented using inverted dropout. The rate parameter the opposite of keep_rate.
keras.layers.Dropout(rate, noise_shape=None, seed=None)
Dropout consists in randomly setting a fraction rate of input units to
0 at each update during training time, which helps prevent
overfitting.
That is, rate sets the rate of dropout and not the rate to keep which you would expect with inverted dropout
Keras Dropout
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.
I'm going through the ML Class on Coursera on Logistic Regression and also the Manning Book Machine Learning in Action. I'm trying to learn by implementing everything in Python.
I'm not able to understand the difference between the cost function and the gradient. There are examples on the net where people compute the cost function and then there are places where they don't and just go with the gradient descent function w :=w - (alpha) * (delta)w * f(w).
What is the difference between the two if any?
Whenever you train a model with your data, you are actually producing some new values (predicted) for a specific feature. However, that specific feature already has some values which are real values in the dataset. We know the closer the predicted values to their corresponding real values, the better the model.
Now, we are using cost function to measure how close the predicted values are to their corresponding real values.
We also should consider that the weights of the trained model are responsible for accurately predicting the new values. Imagine that our model is y = 0.9*X + 0.1, the predicted value is nothing but (0.9*X+0.1) for different Xs.
[0.9 and 0.1 in the equation are just random values to understand.]
So, by considering Y as real value corresponding to this x, the cost formula is coming to measure how close (0.9*X+0.1) is to Y.
We are responsible for finding the better weight (0.9 and 0.1) for our model to come up with a lowest cost (or closer predicted values to real ones).
Gradient descent is an optimization algorithm (we have some other optimization algorithms) and its responsibility is to find the minimum cost value in the process of trying the model with different weights or indeed, updating the weights.
We first run our model with some initial weights and gradient descent updates our weights and find the cost of our model with those weights in thousands of iterations to find the minimum cost.
One point is that gradient descent is not minimizing the weights, it is just updating them. This algorithm is looking for minimum cost.
A cost function is something you want to minimize. For example, your cost function might be the sum of squared errors over your training set. Gradient descent is a method for finding the minimum of a function of multiple variables. So you can use gradient descent to minimize your cost function. If your cost is a function of K variables, then the gradient is the length-K vector that defines the direction in which the cost is increasing most rapidly. So in gradient descent, you follow the negative of the gradient to the point where the cost is a minimum. If someone is talking about gradient descent in a machine learning context, the cost function is probably implied (it is the function to which you are applying the gradient descent algorithm).
It's strange to think about it, but there is more than one measure for how "accurately" a line fits to data points.
To access how accurately a line fits the data, we have a "cost" function which which can compare predicted vs. actual values and provide a "penalty" for how wrong it is.
penalty = cost_funciton(predicted, actual)
A naive cost function might just take the difference between the predicted and actual.
More sophisticated functions will square the value, since we'd rather have many small errors than one large error.
Additionally, each point has a different "sensitivity" to moving the line. Some points react very strongly to movement. Others react less strongly.
Often, you can make a tradeoff, and move TOWARD a point that is sensitive, and AWAY from a point that is NOT sensitive. In that scenario , you get more than you give up.
The "gradient" is a way of measuring how sensitive each point is to moving the line.
This article does a good job of describing WHY there is more than one measure, and WHY some points are more sensitive than others:
https://towardsdatascience.com/wrapping-your-head-around-gradient-descent-with-pictures-3fbd810235f5?source=friends_link&sk=7117e5de8c66bd4a4c2bb2a87a928773
Let's take an example of logistic regression model for binary classification. Output(Predicted Value) of the model for any given input will be offset(deviation) with respect to the actual output(Expected Value) while training. So, the model needs to be trained with minimal error(loss) so that model can perform well with high accuracy.
The function used to find the parameters(m and c in case of linear equation, y = mx+c) value at which the minimal error(loss) occurs is called Cost Function/Loss Function. Loss function is a term used to find the loss for single row/record of the training sample and Cost function is a term used to find the loss for the entire training dataset.
Now, How do we find the parameter(m and c in our case) values at which the minimum loss occurs? Its by using gradient descent algorithm using the equation, which helps us to find the points at which the minimum loss occurs and the parameters values at this points are considered for model building (let say y = 0.5x + 2) where m=.5 and c=2 are the points at which the loss is minimum.
Cost function is something is like at what cost you are building your model for a good model that cost should be minimum. To find the minimum cost function we use gradient descent method. That give value of coefficients to determine minimum cost function