I've looked everywhere and can't find anything that explains the actual derivation of backprop for residual layers. Here's my best attempt and where I'm stuck. It is worth mentioning that the derivation that I'm hoping for is from a generic perspective that need not be limited to convolutional NNs.
If the formula for calculating the output of a normal hidden layer is F(x) then the formula for a hidden layer with a residual connection is F(x) + o, where x is the weight adjusted output of a previous layer, o is the output of a previous layer, and F is the activation function. To get the delta for a normal layer during back-propagation one needs to calculate the gradient of the output ∂F(x)/∂x. For a residual layer this is ∂(F(x) + o)/∂x which is separable into ∂F(x)/∂x + ∂o/∂x (1).
If all of this is correct, how does one deal with ∂o/∂x? It seems to me that it depends on how far back in the network o comes from.
If o is just from the previous layer then o*w=x where w are the weights connecting the previous layer to the layer for F(x). Taking the derivative of each side relative to o gives ∂(o*w)/∂o = ∂x/∂o, and the result is w = ∂x/do which is just the inverse of the term that comes out at (1) above. Does it make sense that in this case the gradient of the residual layer is just ∂F(x)/∂x + 1/w ? Is it accurate to interpret 1/w as a matrix inverse? If so then is that actually getting computed by NN frameworks that use residual connections or is there some shortcut that is for adding in the error from the residual?
If o is from further back in the network then, I think, the derivation becomes slightly more complicated. Here is an example where the residual comes from one layer further back in a network. The network architecture is Input--w1--L1--w2--L2--w3--L3--Out, having a residual connection from the L1 to L3 layers. The symbol o from the first example is replaced by the layer output L1 for unambiguity. We are trying to calculate the gradient at L3 during back-prop which has a forward function of F(x)+L1 where x=F(F(L1*w2)*w3). The derivative of this relationship is ∂x/∂L1=∂F(F(L1*w2)*w3/∂L1, which is more complicated but doesn't seem too difficult to solve numerically.
If the above derivation is reasonable then it's worth noting that there is a case when the derivation fails, and that is when a residual connection originates from the Input layer. This is because the input cannot be broken down into a o*w=x expression (where x would be the input values). I think this must suggest that residual layers cannot originate from from the input layer, but since I've seen network architecture diagrams that have residual connections that originate from the input, this casts my above derivations into doubt. I can't see where I've gone wrong though. If anyone can provide a derivation or code sample for how they calculate the gradient at residual merge points correctly, I would be deeply grateful.
EDIT:
The core of my question is, when using residual layers and doing vanilla back-propagation, is there any special treatment of the error at the layers where residuals are added? Since there is a 'connection' between the layer where the residual comes from and the layer where it is added, does the error need to get distributed backwards over this 'connection'? My thinking is that since residual layers provide raw information from the beginning of the network to deeper layers, the deeper layers should provide raw error to the earlier layers.
Based on what I've seen (reading the first few pages of googleable forums, reading the essential papers, and watching video lectures) and Maxim's post down below, I'm starting to think that the answer is that ∂o/∂x = 0 and that we treat o as a constant.
Does anyone do anything special during back-prop through a NN with residual layers? If not, then does that mean residual layers are an 'active' part of the network on only the forward pass?
I think you've over-complicated residual networks a little bit. Here's the link to the original paper by Kaiming He at al.
In section 3.2, they describe the "identity" shortcuts as y = F(x, W) + x, where W are the trainable parameters. You can see why it's called "identity": the value from the previous layer is added as is, without any complex transformation. This makes two things:
F now learns the residual y - x (discussed in 3.1), in short: it's easier to learn.
The network gets an extra connection to the previous layer, which improves the gradients flow.
The backward flow through the identity mapping is trivial: the error message is passed unchanged, no inverse matrices are involved (in fact, they are not involved in any linear layer).
Now, paper authors go a bit further and consider a slightly more complicated version of F, which changes the output dimensions (which probably you had in mind). They write it generally as y = F(x, W) + Ws * x, where Ws is the projection matrix. Note that, though it's written as matrix multiplication, this operation is in fact very simple: it adds extra zeros to x to make its shape larger. You can read a discussion of this operation in this question. But this does very few changes the backward: the error message is simply clipped to the original shape of x.
Related
I read in a book, where the author mentioned that the bias bk is used to produce an affine transform to the output uk (The summation of weighted input signals).
Also, the author mentioned that due to this bias that gives a constant value of, say 'k' makes the neuron not connected to the previous layer.
I am in a confused state. Can someone please tell me what the above two points mean, and if there are any other uses of a bias to the network?
Thanks in advance!
If the neurons activation is z(a) = wa + b, b is the bias. It's a bias because the larger it is, the more this neuron is biased, or in other words it doesnt care much about what was passed to it (a) from the last layer. I'm assuming the second point is referring to the fact that if a bias is large enough (positive or negative) it is like the neuron no longer cares what is passed to it, it's always going to pass the same thing to the next layer. I would need to see it in context to be certain about what the author is saying, but overall you just need to understand that it is a constant that can add bias (doesnt care about what the last layer gave it). Dont fret too much about its implications though, because the learning (or optimization) process is going to adjust these automatically so you're not going to have to choose proper bias values for the network. As you become more familiar with the concepts it will start to make more sense
I'm trying to get a better understanding of neural networks by trying to programm a Convolution Neural Network by myself.
So far, I'm going to make it pretty simple by not using max-pooling and using simple ReLu-activation. I'm aware of the disadvantages of this setup, but the point is not making the best image detector in the world.
Now, I'm stuck understanding the details of the error calculation, propagating it back and how it interplays with the used activation-function for calculating the new weights.
I read this document (A Beginner's Guide To Understand CNN), but it doesn't help me understand much. The formula for calculating the error already confuses me.
This sum-function doesn't have defined start- and ending points, so i basically can't read it. Maybe you can simply provide me with the correct one?
After that, the author assumes a variable L that is just "that value" (i assume he means E_total?) and gives an example for how to define the new weight:
where W is the weights of a particular layer.
This confuses me, as i always stood under the impression the activation-function (ReLu in my case) played a role in how to calculate the new weight. Also, this seems to imply i simply use the error for all layers. Doesn't the error value i propagate back into the next layer somehow depends on what i calculated in the previous one?
Maybe all of this is just uncomplete and you can point me into the direction that helps me best for my case.
Thanks in advance.
You do not backpropagate errors, but gradients. The activation function plays a role in caculating the new weight, depending on whether or not the weight in question is before or after said activation, and whether or not it is connected. If a weight w is after your non-linearity layer f, then the gradient dL/dw wont depend on f. But if w is before f, then, if they are connected, then dL/dw will depend on f. For example, suppose w is the weight vector of a fully connected layer, and assume that f directly follows this layer. Then,
dL/dw=(dL/df)*df/dw //notations might change according to the shape
//of the tensors/matrices/vectors you chose, but
//this is just the chain rule
As for your cost function, it is correct. Many people write these formulas in this non-formal style so that you get the idea, but that you can adapt it to your own tensor shapes. By the way, this sort of MSE function is better suited to continous label spaces. You might want to use softmax or an svm loss for image classification (I'll come back to that). Anyway, as you requested a correct form for this function, here is an example. Imagine you have a neural network that predicts a vector field of some kind (like surface normals). Assume that it takes a 2d pixel x_i and predicts a 3d vector v_i for that pixel. Now, in your training data, x_i will already have a ground truth 3d vector (i.e label), that we'll call y_i. Then, your cost function will be (the index i runs on all data samples):
sum_i{(y_i-v_i)^t (y_i-vi)}=sum_i{||y_i-v_i||^2}
But as I said, this cost function works if the labels form a continuous space (here , R^3). This is also called a regression problem.
Here's an example if you are interested in (image) classification. I'll explain it with a softmax loss, the intuition for other losses is more or less similar. Assume we have n classes, and imagine that in your training set, for each data point x_i, you have a label c_i that indicates the correct class. Now, your neural network should produce scores for each possible label, that we'll note s_1,..,s_n. Let's note the score of the correct class of a training sample x_i as s_{c_i}. Now, if we use a softmax function, the intuition is to transform the scores into a probability distribution, and maximise the probability of the correct classes. That is , we maximse the function
sum_i { exp(s_{c_i}) / sum_j(exp(s_j))}
where i runs over all training samples, and j=1,..n on all class labels.
Finally, I don't think the guide you are reading is a good starting point. I recommend this excellent course instead (essentially the Andrew Karpathy parts at least).
What is the concept behind taking the derivative? It's interesting that for somehow teaching a system, we have to adjust its weights. But why are we doing this using a derivation of the transfer function. What is in derivation that helps us. I know derivation is the slope of a continuous function at a given point, but what does it have to do with the problem.
You must already know that the cost function is a function with the weights as the variables.
For now consider it as f(W).
Our main motive here is to find a W for which we get the minimum value for f(W).
One of the ways for doing this is to plot function f in one axis and W in another....... but remember that here W is not just a single variable but a collection of variables.
So what can be the other way?
It can be as simple as changing values of W and see if we get a lower value or not than the previous value of W.
But taking random values for all the variables in W can be a tedious task.
So what we do is, we first take random values for W and see the output of f(W) and the slope at all the values of each variable(we get this by partially differentiating the function with the i'th variable and putting the value of the i'th variable).
now once we know the slope at that point in space we move a little further towards the lower side in the slope (this little factor is termed alpha in gradient descent) and this goes on until the slope gives a opposite value stating we already reached the lowest point in the graph(graph with n dimensions, function vs W, W being a collection of n variables).
The reason is that we are trying to minimize the loss. Specifically, we do this by a gradient descent method. It basically means that from our current point in the parameter space (determined by the complete set of current weights), we want to go in a direction which will decrease the loss function. Visualize standing on a hillside and walking down the direction where the slope is steepest.
Mathematically, the direction that gives you the steepest descent from your current point in parameter space is the negative gradient. And the gradient is nothing but the vector made up of all the derivatives of the loss function with respect to each single parameter.
Backpropagation is an application of the Chain Rule to neural networks. If the forward pass involves applying a transfer function, the gradient of the loss function with respect to the weights will include the derivative of the transfer function, since the derivative of f(g(x)) is f’(g(x))g’(x).
Your question is a really good one! Why should I move the weight more in one direction when the slope of the error wrt. the weight is high? Does that really make sense? In fact it does makes sense if the error function wrt. the weight is a parabola. However it is a wild guess to assume it is a parabola. As rcpinto says, assuming the error function is a parabola, make the derivation of the a updates simple with the Chain Rule.
However, there are some other parameter update rules that actually addresses this, non-intuitive assumption. You can make update rule that takes the weight a fixed size step in the down-slope direction, and then maybe later in the training decrease the step size logarithmic as you train. (I'm not sure if this method has a formal name.)
There are also som alternative error function that can be used. Look up Cross Entropy in you neural network text book. This is an adjustment to the error function such that the derivative (of the transfer function) factor in the update rule cancels out. Just remember to pick the right cross entropy function based on you output transfer function.
When I first started getting into Neural Nets, I had this question too.
The other answers here have explained the math which makes it pretty clear that a derivative term will appear in your calculations while you are trying to update the weights. But all of those calculations are being done in order to implement Back-propagation, which is just one of the ways of updating weights! Now read on...
You are correct in assuming that at the end of the day, all a neural network tries to do is update its weights to fit the data you feed into it. Within this statement lies your answer too. What you are getting confused with here is the idea of the Back-propagation algorithm. Many textbooks use backprop to update neural nets by default but do not mention that there are other ways to update weights too. This leads to the confusion that neural nets and backprop are the same thing and are inherently connected. This also leads to the false belief that neural nets need backprop to train.
Please remember that Back-propagation is just ONE of the ways out there to train your neural network (although it is the most famous one). Now, you must have seen the math involved in backprop, and hence you can see where the derivative term comes in from (some other answers have also explained that). It is possible that other training methods won't need the derivatives, although most of them do. Read on to find out why....
Think about this intuitively, we are talking about CHANGING weights, the direct mathematical operation related to change is a derivative, makes sense that you should need to evaluate derivatives to change weights.
Do let me know if you are still confused and I'll try to modify my answer to make it better. Just as a parting piece of information, another common misconception is that gradient descent is a part of backprop, just like it is assumed that backprop is a part of neural nets. Gradient descent is just one way to minimize your cost function, there are plenty of others you can use. One of the answers above makes this wrong assumption too when it says "Specifically Gradient Descent". This is factually incorrect. :)
Training a neural network means minimizing an associated "error" function wrt the networks weights. Now there are optimization methods that use only function values (Simplex method of Nelder and Mead, Hooke and Jeeves, etc), methods that in addition use first derivatives (steepest descend, quasi Newton, conjugate gradient) and Newton methods using second derivatives as well. So if you want to use a derivative method, you have to calculate the derivatives of the error function, which in return involves the derivatives of the transfer or activation function.
Back propagation is just a nice algorithm to calculate the derivatives, and nothing more.
Yes, the question was really good, this question was also came in my head while i am understanding the Backpropagation. After doing ForwordPropagation on neural network we do back propagation in network to minimize the total error. And there also many other way to minimize the error.your question is why we are doing derivative in backpropagation, the reason is that, As we all know the meaning of derivative is to find the slope of a function or in other words we can find change of particular thing with respect to particular thing. So here we are doing derivative to minimize the total error with respect to the corresponding weights of the network.
and here by doing the derivation of total error with respect to weights we can find it's slope or in other words we can find what is the change in total error with respect to the small change of the weight, so that we can update the weight to minimize the error with the help of this Gradient Descent formula, that is, Weight= weight-Alpha*(del(Total error)/del(weight)).Or in other words New Weights = Old Weights - learning-rate x Partial derivatives of loss function w.r.t. parameters.
Here Alpha is the learning rate which is control the weight update, means if the derivative the - ve than Alpha make it +ve(Becouse of -Alpha in formula) and if +ve it's remain +ve so that weight update goes in +ve direction and it's reflected to minimize the Total error.And also the as derivative part is multiples with Alpha, it's decrees the step size of Alpha when the weight converge to the optimal value of weight(minimum error). Thats why we are doing derivative to minimize the error.
I have AR(1) model with data samples $N=500$ that is driven by a random input sequence x. THe observation y is corrupted with measurement noise $v$ of zero mean. The model is
y(t) = 0.195y(t-1) + x(t) + v(t) where x(t) is generated as randn(). I am unsure how to represent this as a state space model and how to estimate the parameters $a$ and the states. I tried the state space representation would be
d(t) = \mathbf{a^T} d(t) + x(t)
y(t) = \mathbf{h^T}d(t) + sigma*v(t)
sigma =2.
I cannot understand how to perform parameter and state estimation. Using the toolbox mentioned below, I checked the Equations of KF to be matching with those in textbooks. However, the approach for parameter estimation is different. I will appreciate a recommendation for the implementation procedure.
Implementation 1:
I am following the implementation here : Learning Kalman Filter. This implementation does not use Expectation Maximization to estimate the parameters of AR model and it finds out the Covariance of the process noise. In my case, I don't have a process noise, but an input $x$.
Implementation 2: Kalman Filter by Kevin Murphy is another toolbox which uses EM for parameter estimation of AR model. Now, it is confusing since both the implementations uses different approach for parameter estimation.
I am having a tough time figuring out the correct approach, the state space model representation and the code. Shall appreciate recommendations on how to proceed.
I ran the first implementation for the KalmanARSquareRoot technique and the result is completely different. There is Exponential Moving Average Smoothing being performed and a MA filer of length 30 being used. The toolbox runs fine if I run the Demo examples. But on changing the model, the result is extremely poor. Maybe I am doing something wrong. Do I need to change the equations of KF for my time series?
In the second implementation, I cannot figure out what and where to change the Equations.
In general, if I have to use these tools, then do I need to change the KF equations for every time series model? How do I write the Equations myself if these toolboxes are inappropriate for all the time series model - AR, MA, ARMA?
I only have a bit of experience with Kalman Filters, so take this with a grain of salt.
It seems you shouldn't need to change the equations at all. Working with the second package (learn_kalman), you can create an A0 matrix of size [length(d(t)) length(d(t))]. C0 is the same, and in your case the initial state probably makes sense to be the Identity matrix (unlike your A0. All you need to do is choose a good initial condition.
However, I took a look at your system (plotted an example) and it seems noise dominates your system. KF is an optimal estimator but I have not known it to reject that much noise. It only guarantees a reduced covariance...which means that if your system is mostly dominated by noise, you will calculate a bad model that estimates your system given the noise!
Try plotting [d f] where d is the initial data and f is calculated using the regressive formula f(t) = C * A * f(t-1) :
f(t) = A * f(t-1)
; y(t) = C * f(t)
That is, pretend as if there is no noise but using the estimated AR model. You will see that it rejects all the noise and 'technically' models the system well (since the only unique behaviour is at the very beginning).
For example, if you have a system with A = 0.195, Q=R=0.l then you will converge to an A = 0.2207 but it still isn't good enough. Here the problem is that your initial state is so low, that within a few steps of data and you are essentially at 0 accounting for noise. Naturally KF can converge to a LOT of model solutions that are similar. Any noise will throw off even the best initial condition.
If you increase the resolution of your data in some way (e.g. larger initial condition, more refined timesteps) you will see a good fit. Ex, changing your initial condition to 110 and you'll find the two curves similar, though the model is still fairly different.
I am not aware of any approach to model your data well. If the noise variance is in fact 1 and your system converges to 0 that quickly, it seems doomed to not be effectively modelled since you just don't capture any unique behaviour in the dataset.
edit:
A more pointed question:
What is the derivative of softmax to be used in my gradient descent?
This is more or less a research project for a course, and my understanding of NN is very/fairly limited, so please be patient :)
I am currently in the process of building a neural network that attempts to examine an input dataset and output the probability/likelihood of each classification (there are 5 different classifications). Naturally, the sum of all output nodes should add up to 1.
Currently, I have two layers, and I set the hidden layer to contain 10 nodes.
I came up with two different types of implementations
Logistic sigmoid for hidden layer activation, softmax for output activation
Softmax for both hidden layer and output activation
I am using gradient descent to find local maximums in order to adjust the hidden nodes' weights and the output nodes' weights. I am certain in that I have this correct for sigmoid. I am less certain with softmax (or whether I can use gradient descent at all), after a bit of researching, I couldn't find the answer and decided to compute the derivative myself and obtained softmax'(x) = softmax(x) - softmax(x)^2 (this returns an column vector of size n). I have also looked into the MATLAB NN toolkit, the derivative of softmax provided by the toolkit returned a square matrix of size nxn, where the diagonal coincides with the softmax'(x) that I calculated by hand; and I am not sure how to interpret the output matrix.
I ran each implementation with a learning rate of 0.001 and 1000 iterations of back propagation. However, my NN returns 0.2 (an even distribution) for all five output nodes, for any subset of the input dataset.
My conclusions:
I am fairly certain that my gradient of descent is incorrectly done, but I have no idea how to fix this.
Perhaps I am not using enough hidden nodes
Perhaps I should increase the number of layers
Any help would be greatly appreciated!
The dataset I am working with can be found here (processed Cleveland):
http://archive.ics.uci.edu/ml/datasets/Heart+Disease
The gradient you use is actually the same as with squared error: output - target. This might seem surprising at first, but the trick is that a different error function is minimized:
(- \sum^N_{n=1}\sum^K_{k=1} t_{kn} log(y_{kn}))
where log is the natural logarithm, N depicts the number of training examples and K the number of classes (and thus units in the output layer). t_kn depicts the binary coding (0 or 1) of the k'th class in the n'th training example. y_kn the corresponding network output.
Showing that the gradient is correct might be a good exercise, I haven't done it myself, though.
To your problem: You can check whether your gradient is correct by numerical differentiation. Say you have a function f and an implementation of f and f'. Then the following should hold:
(f'(x) = \frac{f(x - \epsilon) - f(x + \epsilon)}{2\epsilon} + O(\epsilon^2))
please look at sites.google.com/site/gatmkorn for the open-source Desire simulation program.
For the Windows version, /mydesire/neural folder has several softmax classifiers, some with softmax-specific gradient-descent algorithm.
In the examples, this works nicely for a simplemcharacter-recognition task.
ASee also
Korn, G.A.: Advanced dynamic-system Simulation, Wiley 2007
GAK
look at the link:
http://www.youtube.com/watch?v=UOt3M5IuD5s
the softmax derivative is: dyi/dzi= yi * (1.0 - yi);