Derivative of activation function vs partial derivative wrt. loss function - machine-learning

Some terms in AI are confusing me. The derivative function used in backpropagation is the derivative of activation function or the derivative of loss function?
These terms are confusing: derivative of act. function, partial derivative wrt. loss function??
I'm still not getting it correct.

When you optimize a model, you define a loss function. This typically represents the error with respect to some training data.
It is common to use gradient based optimization to minimize this error. Typically, stochastic gradient descent (SGD) and related approaches (Adam, Adagrad, etc.).
The gradient of the loss function, is a vector composed of the partial derivatives of the loss with respect to each of the weights in the model.
In each iteration, weights are updated against the direction of the gradient (remember we are minimizing).
I guess the reason you might be confused is because due to the chain rule, when calculating the gradient of the loss function, you are required to differentiate activation functions. But keep in mind that this is because of the chain rule.

Related

ReLU weight initialization?

I have read that the " He weight Initialization" (He et al., 2015) built on the Lecun weight initialization and suggested a zero-mean Gaussian distribution where the standard deviation is
enter image description here
and this function should be used with ReLU to solve the vanishing/exploding gradient problem. For me, it does make sense because the way ReLu was built makes it no bothered with vanishing/exploding gradient problem. Since, if the input is less than 0 the derivative would be zero otherwise the derivative would be one. So, whatever the variance is, the gradient would be zero or one. Therefore, the He weight Initialization is useless. I know that I am missing something, that's why I am asking if anyone would tell me the usefulness of that weight initialization?
Weight initialization is applied, in general terms, to weights of layers that have learnable / trainable parameters, just like dense layers, convolutional layers, and other layers. ReLU is an activation function, fully deterministic, and has no initialization.
Regarding to the vanishing gradient problem, the backpropagation step is funded by computing the gradients by the chain rule (partial derivatives) for each weight (see here):
(...) each of the neural network's weights receive an update
proportional to the partial derivative of the error function with
respect to the current weight in each iteration of training.
The more deep a network is, the smaller these gradients get, and when a network becomes deep enough, the backprop step is less effective (in the worst case, it stops learning) and this becomes a problem:
This has the effect of multiplying n of these small numbers to compute
gradients of the "front" layers in an n-layer network, meaning that
the gradient (error signal) decreases exponentially with n while the
front layers train very slowly.
Choosing a proper activation function, like ReLU, help avoiding this to happen, as you mentioned in the OP, by making partial derivatives of this activation not too small:
Rectifiers such as ReLU suffer less from the vanishing gradient
problem, because they only saturate in one direction.
Hope this helps!

Will the shape of the Loss function change during training?

I have some problem understanding the theory of loss function and hope some one can help me.
Usually when people try to explain gradient descent to you, they will show you a loss function that looks like the very first image in this post gradient descent: all you need to know. I understand the entire theory of gradient descent is to adjust the weights and minimize the loss function.
My question is, will the shape of the Loss function change during the training or it will just stay remain as the image shown in the above post? I know that the weights are something that we are always tuning so the parameters that determines the shape of the Loss function should be the inputs X={x1,x2,...xn}. Let's make an easy example: suppose our inputs are [[1,2,3,4,5],[5,4,3,2,1]] and labels are [1,0] (Only two training sample for ease, and we are setting the batch size to 1). Then the loss function should be some thing like this for the first training sample
L = (1-nonlinear(1*w1+2*w2+3*w3+4*w4+5*w5+b))^2
and for the second training sample the loss function should be:
L = (0-nonlinear(5*w1+4*w2+3*w3+2*w4+1*w5+b))^2
Apparently, these two loss functions doesn't looks like the same if we plot them so does that mean the shape of the Loss function are changing during training? Then why are people still using that one image ( A point that slides down from the Loss function and finds the global minima) to explain the gradient descent theory?
Note: I'm not changing the loss function, the loss function are still mean square error. I'm trying to say that the shape of the Loss function seems to be changing.
I know where my problem comes from! I thought that we are not able to plot a function such as f(x,y) = xy without any constant in it, but we actually could! I searched the graph on google for f(x,y)=xy and truly we can plot them out! So now I understand, as long as we get the lost function, we can get the plot! Thanks guys
The function stays the same. The point of Gradient Decent is to find the lowest point on a given loss function that you define.
Generally, the loss function you are training to minimize does not change throughout the course of a training session. The flaw in reasoning is that you are assuming that the loss function is characterized by weights of the network, when in fact the weights of that network are a sort-of input to the loss function.
To clarify, let us assume we are predicting some N-dimensional piece of information and we have a ground truth vector, call it p, and a loss function L taking in a prediction vector p_hat which we define as
L(p_hat) := norm(p - p_hat).
This is a very primitive (and quite ineffective) loss function, but it is one nonetheless. Once we begin training, this loss function will be the function that we will try to minimize to get our network to perform the best with respect to. Notice that this loss function will attain different values for different inputs of p_hat, this does not mean the loss function is changing! In the end, the loss function will be an N-dimensional hypersurface in an N+1-dimensional hyperspace that stays the same no matter what (similar to the thing you see in the image where it is a 2-dimensional surface in a 3-dimensional space).
Gradient descent tries to find a minimum on this surface that is constructed by the loss function, but we do not really know what the surface looks like as a whole, instead, we find out small things about the surface by evaluating the loss function as the values of p_hat we give it.
Note, this is all a huge oversimplification, but can be a useful way to think about it getting started.
A Loss Function is a metric that measures the distance from your predictions to your targets.
The ideia is to choose the weighs so your predictions are close to your targets, that is, your model learned/memorized the input.
The loss function should usually not be changed during training, because the minimum in the original function might not coincide with the new one, so the gradient descent's work is lost.

Is there any benefit to having linear activation functions at the last layer vs an activation function like tanh?

I understand this decision depends on the task, but let me explain.
I'm designing a model that predicts steering angles from a given dashboard video frame using a convolutional neural network with dense layers at end. In my final dense layer, I only have a single unit that predicts a steering angle.
My question here is, for my task would either option below show a boost in performance?
a. Get ground truth steering angles, convert to radians, and squash them using tanh so they are between -1 and 1. In the final dense layer of my network, use a tanh activation function.
b. Get ground truth steering angles. These raw angles are between -420 and 420 degrees. In the final layer, use a linear activation.
I'm trying to think about it logically, where in option A the loss will likely be much smaller since the network is dealing with much smaller numbers. This would lead to smaller changes in weights.
Let me know your thoughts!
There are two types of variables in neural networks: weights and biases (mostly, there are additional variables, e.g. the moving mean and moving variance required for batchnorm). They behave a bit differently, for instance biases are not penalized by a regularizer as a result they don't tend to get small. So an assumption that the network is dealing only with small numbers is not accurate.
Still, biases need to be learned, and as can be seen from ResNet performance, it's easier to learn smaller values. In this sense, I'd rather pick [-1, 1] target range over [-420, 420]. But tanh is probably not an optimal activation function:
With tahn (just like with sigmoid), a saturated neuron kills the gradient during backprop. Choosing tahn with no specific reason is likely to hurt your training.
Forward and backward passes with tahn need to compute exp, which is also relatively expensive.
My option would be (at least initially, until some other variant proves to work better) to squeeze the ground truth values and have no activation at all (I think that's what you mean by a linear activation): let the network learn [-1, 1] range by itself.
In general, if you have any activation functions in the hidden layers, ReLu has proven to work better than sigmoid, though other modern functions have been proposed recently, e.g. leaky ReLu, PRelu, ELU, etc. You might try any of those.

Backpropagation in Gradient Descent for Neural Networks vs. Linear Regression

I'm trying to understand "Back Propagation" as it is used in Neural Nets that are optimized using Gradient Descent. Reading through the literature it seems to do a few things.
Use random weights to start with and get error values
Perform Gradient Descent on the loss function using these weights to arrive at new weights.
Update the weights with these new weights until the loss function is minimized.
The steps above seem to be the EXACT process to solve for Linear Models (Regression for e.g.)? Andrew Ng's excellent course on Coursera for Machine Learning does exactly that for Linear Regression.
So, I'm trying to understand if BackPropagation does anything more than gradient descent on the loss function.. and if not, why is it only referenced in the case of Neural Nets and why not for GLMs (Generalized Linear Models). They all seem to be doing the same thing- what might I be missing?
The main division happens to be hiding in plain sight: linearity. In fact, extend to question to continuity of the first derivative, and you'll encapsulate most of the difference.
First of all, take note of one basic principle of neural nets (NN): a NN with linear weights and linear dependencies is a GLM. Also, having multiple hidden layers is equivalent to a single hidden layer: it's still linear combinations from input to output.
A "modern' NN has non-linear layers: ReLUs (change negative values to 0), pooling (max, min, or mean of several values), dropouts (randomly remove some values), and other methods destroy our ability to smoothly apply Gradient Descent (GD) to the model. Instead, we take many of the principles and work backward, applying limited corrections layer by layer, all the way back to the weights at layer 1.
Lather, rinse, repeat until convergence.
Does that clear up the problem for you?
You got it!
A typical ReLU is
f(x) = x if x > 0,
0 otherwise
A typical pooling layer reduces the input length and width by a factor of 2; in each 2x2 square, only the maximum value is passed through. Dropout simply kills off random values to make the model retrain those weights from "primary sources". Each of these is a headache for GD, so we have to do it layer by layer.
So, I'm trying to understand if BackPropagation does anything more than gradient descent on the loss function.. and if not, why is it only referenced in the case of Neural Nets
I think (at least originally) back propagation of errors meant less than what you describe: the term "backpropagation of errors" only refered to the method of calculating derivatives of the loss function, instead of e.g. automatic differentiation, symbolic differentiation, or numerical differentiation. No matter what the gradient was then used for (e.g. Gradient Descent, or maybe Levenberg/Marquardt).
They all seem to be doing the same thing- what might I be missing?
They're using different models. If your neural network used linear neurons, it would be equivalent to linear regression.

What should be a generic enough convergence criteria of Stochastic Gradient Descent

I am implementing a generic module for Stochastic Gradient Descent. That takes arguments: training dataset, loss(x,y), dw(x,y) - per sample loss and per sample gradient change.
Now, for the convergence criteria, I have thought of :-
a) Checking loss function after every 10% of the dataset.size, averaged over some window
b) Checking the norm of the differences between weight vector, after every 10-20% of dataset size
c) Stabilization of error on the training set.
d) Change in the sign of the gradient (again, checked after every fixed intervals) -
I have noticed that these checks (precision of check etc.) depends on other stuff also, like step size, learning rate.. and the effect can vary from one training problem to another.
I can't seem to make up mind on, what should be the generic stopping criterion, regardless of the training set, fx,df/dw thrown at the SGD module. What do you guys do?
Also, for (d), what would be the meaning of "change in sign" for a n-dimensional vector? As, in - given dw_i, dw_i+1, how do I detect the change of sign, does it even have a meaning in more than 2 dimensions?
P.S. Apologies for non-math/latex symbols..still getting used to the stuff.
First, stochastic gradient descent is the on-line version of gradient descent method. The update rule is using a single example at a time.
Suppose, f(x) is your cost function for a single example, the stopping criteria of SGD for N-dimensional vector is usually:
See this1, or this2 for details.
Second, there is a further twist on stochastic gradient descent using so-called “minibatches”. It works identically to SGD, except that it uses more than one training example to make each estimate of the gradient. This technique reduces variance in the estimate of the gradient, and often makes better use of the hierarchical memory organization in modern computers. See this3.

Resources