Backpropogation: WHERE is Derivative of Transfer Function - neural-network

First off: I understand derivatives and the chain rule. I'm not great with math, but I have an understanding.
Numerous tutorials on Backpropogation (let's use this and this) using gradient descent state that we use the derivative of the transfer function (sigmoid or tanh) to calculate the gradient and hence which way to travel next. In these tutorials I see (t-o)(1-o)(o) as the formula for calculating the error for the output neurons, which seems to be the derivative of the error calculation (1/2)(t-o)^2 * (1-o). (t = target, o = output_actual)
Why do I not see the derivative of the transfer function (assuming sigmoid): e^x/((e^x + 1)^2) anywhere? Or for when tanh is used as the transfer function: sech^2(x) ... where
x = weighted input?
Also, some tutorials use (target - actual) , (target - actual)^2 [ Sum of Squares - useful for negative outputs] or the squared error function: (1/2)(target - actual)^2.
Where is the derivative of the transfer function and which is the correct error formula to use?

Why do I not see the derivative of the transfer function (assuming sigmoid): e^x/((e^x + 1)^2) anywhere?
You do, its expressed as with the derivative in the wiki page you link. If we expand the latter we get
(1/(1+e^-x))*(1-1/(1+e^-x)) = e^x/(e^x+1)^2
which is the original form you noted.
Or for when tanh is used as the transfer function: sech^2(x) ... where x = weighted input?
Well, in this case its because the page doesn't mention the tanh as a potential activation function. But in real life it is expressed in a similar way so that we can avoid any unnecessary computations.
(target - actual)^2 [ Sum of Squares - useful for negative outputs] or the squared error function: (1/2)(target - actual)^2.
The difference is only a constant factor. The math comes out a little nicer if you keep the division by 2. In practice the only thing that would change is your learning rate gets implicitly multiplied/divided by 2 depending on which perspective you look at.
Also, some tutorials use (target - actual)
You probably misread. (t-a) would be the derivative of (t-a)^2/2 . Just (t-a) would have a derivative of -1, which I'm fairly sure would hinder learning for a nn.

It is a common calculus topic to find derivatives for functions.
You may also use online mathematica to do so, here: http://www.wolframalpha.com/
The code to enter is
D[ 1/(1+e^(-x)), x ]
You can enter any function using Mathematica notation: http://integrals.wolfram.com/about/input/.
With the derivative, you can plug it in the general formula for error functions. When the derivative is too complex, you can try to use the function Simplify[...] to find better analytic forms.
As for choosing which transfer function to use, you can consider their domains and ranges. The logistic function (1 / (1 + exp(-x)) has range (0,1) but the atan(x) function has range (-1, 1). If you perform mathematical analysis on the learning algorithms, the choice of the transfer function may matter a lot. However, if you are running simulations, the choice of transfer functions should not be critical, as long as they have the S-shape (sigmoidal).
Another thing to pint out, the logistic function (1 / (1 + exp(-x) ) is only one instance of sigmoidal functions. atan(x) is also sigmoidal.

Related

Dirac probability measure in Matlab

I want to implement densities of probability measures in Matlab. For that I define density as a function handle such that the integral of some function f (given as a function handle) on the interval [a,b] can be computed by
syms x
int(f(x)*density(x),x,a,b)
When it comes to the Dirac measure the problem is that
int(dirac(x),x,0,b)
delivers the value 1/2 instead of 1 for all b>0. However if I type
int(dirac(x),x,a,b)
where a<0 and b>0 the returned value is 1 as it should be. For this reason multiplying by 2 will not suffice as I want my density to be valid for all intervals [a,b]. I also dont want to distinct cases before integrating so that the code remains valid for a large class of densities.
Does someone know how I can implement the Dirac probability measure (as defined here) in Matlab?
There's no unique, accepted definition for int(delta, 0, b). The problem here isn't that you're getting the "wrong" answer as that you want to impose a different convention somehow on what the delta function is than what was provided by Matlab. (Their choice is defensible but not unique.) If you evaluate this in Wolfram Alpha, for example, it will give you theta(0) - which is not defined as anything in particular. Here theta is the Heaviside function. If you want to impose your own convention, implement your own delta function.
EDIT
I see you wrote a comment on the question, while I was writing this answer, so.... Keep in mind that the Dirac measure or Dirac delta function is not a function at all. The problem you're having, along with what's described below, all relate to trying to give a functional form to something that is inherently not a function. What you are doing is not well-defined in the framework that you have in Matlab.
END OF EDIT
To put the point about conventions in context, the delta function can be defined by different properties. One is that int(delta(x) f(x), a, b) = f(0) when a < 0 < b. This doesn't tell you anything about the integral that you want. Another, which probably leads to an answer as you're getting from Matlab, is to define it as a limit. One (but not the only choice) is the limit of the zero-mean Gaussian as the variance goes to 0.
If you want to use a convention int(delta(x) f(x), a, b) = f(0) when a <= 0 < b, that probably won't get you in much trouble, but just keep in mind that it's a convention you've chosen more than a "right" or
"wrong" answer relative to what you got from Matlab.
As related note, there is a similar choice to be made on the step function (Heaviside function) at x=0. There are conventions in which is is (a) undefined, (b) -1, (c) +1, and (d) 1/2. None is "wrong." This probably corresponds roughly to the choice on the Dirac function since the Heaviside is (roughly) the integral of the Dirac.

Confused by the notation (a and z) and usage of backpropagation equations used in neural networks gradient decent training

I’m writing a neural network but I have trouble training it using backpropagation so I suspect there is a bug/mathematical mistake somewhere in my code. I’ve spent ours reading different literature on how the equations of backpropagation should look but I’m a bit confused by it since different books say different things, or at least use wildly confusing and contradictory notation. So, I was hoping that someone who knows with a 100% certainty how it works could clear it out for me.
There are two steps in the backpropagation that confuse me. Let’s assume for simplicity that I only have a three layer feed forward net, so we have connections between input-hidden and hidden-output. I call the weighted sum that reaches a node z and the same value but after it has passed the activation function of the node a.
Apparently I’m not allowed to embed an image with the equations that my question concern so I will have to link it like this: https://i.stack.imgur.com/CvyyK.gif
Now. During backpropagation, when calculating the error in the nodes of the output layer, is it:
[Eq. 1] Delta_output = (output-target) * a_output through the derivative of the activation function
Or is it
[Eq. 2] Delta_output = (output-target) * z_output through the derivative of the activation function
And during the error calculation of the nodes in the hidden layer, same thing, is it:
[Eq. 3] Delta_hidden = a_h through the derivative of the activation function * sum(w_h*Delta_output)
Or is it
[Eq. 4] Delta_hidden = z_h through the derivative of the activation function * sum(w_h*Delta_output)
So the question is basically; when running a node's value through the derivative version of the activation function during backpropagation, should the value be expressed as it was before or after it passed the activation function (z or a)?
Is the first or the second equation in the image correct and similarly is the third or fourth equation in the image correct?
Thanks.
You have to compute the derivatives with the values before it have passed through the activation function. So the answer is "z".
Some activation functions simplify the computation of the derivative, like tanh:
a = tanh(z)
derivative on z of tanh(z) = 1.0 - tanh(z) * tanh(z) = 1.0 - a * a
This simplification can lead to the confusion you was talking about, but here is another activation function without possible confusion:
a = sin(z)
derivative on z of sin(z) = cos(z)
You can find a list of activation functions and their derivatives on wikipedia: activation function.
Some networks doesn't have an activation function on the output nodes, so the derivative is 1.0, and delta_output = output - target or delta_output = target - output, depending if you add or substract the weight change.
If you are using and activation function on the output nodes, the you'll have to give targets that are in the range of the activation function like [-1,1] for tanh(z).

Goodness of fit with MATLAB and chi-square test

I would like to measure the goodness-of-fit to an exponential decay curve. I am using the lsqcurvefit MATLAB function. I have been suggested by someone to do a chi-square test.
I would like to use the MATLAB function chi2gof but I am not sure how I would tell it that the data is being fitted to an exponential curve
The chi2gof function tests the null hypothesis that a set of data, say X, is a random sample drawn from some specified distribution (such as the exponential distribution).
From your description in the question, it sounds like you want to see how well your data X fits an exponential decay function. I really must emphasize, this is completely different to testing whether X is a random sample drawn from the exponential distribution. If you use chi2gof for your stated purpose, you'll get meaningless results.
The usual approach for testing the goodness of fit for some data X to some function f is least squares, or some variant on least squares. Further, a least squares approach can be used to generate test statistics that test goodness-of-fit, many of which are distributed according to the chi-square distribution. I believe this is probably what your friend was referring to.
EDIT: I have a few spare minutes so here's something to get you started. DISCLAIMER: I've never worked specifically on this problem, so what follows may not be correct. I'm going to assume you have a set of data x_n, n = 1, ..., N, and the corresponding timestamps for the data, t_n, n = 1, ..., N. Now, the exponential decay function is y_n = y_0 * e^{-b * t_n}. Note that by taking the natural logarithm of both sides we get: ln(y_n) = ln(y_0) - b * t_n. Okay, so this suggests using OLS to estimate the linear model ln(x_n) = ln(x_0) - b * t_n + e_n. Nice! Because now we can test goodness-of-fit using the standard R^2 measure, which matlab will return in the stats structure if you use the regress function to perform OLS. Hope this helps. Again I emphasize, I came up with this off the top of my head in a couple of minutes, so there may be good reasons why what I've suggested is a bad idea. Also, if you know the initial value of the process (ie x_0), then you may want to look into constrained least squares where you bind the parameter ln(x_0) to its known value.

How to overcome singularities in numerical integration (in Matlab or Mathematica)

I want to numerically integrate the following:
where
and a, b and β are constants which for simplicity, can all be set to 1.
Neither Matlab using dblquad, nor Mathematica using NIntegrate can deal with the singularity created by the denominator. Since it's a double integral, I can't specify where the singularity is in Mathematica.
I'm sure that it is not infinite since this integral is based in perturbation theory and without the
has been found before (just not by me so I don't know how it's done).
Any ideas?
(1) It would be helpful if you provide the explicit code you use. That way others (read: me) need not code it up separately.
(2) If the integral exists, it has to be zero. This is because you negate the n(y)-n(x) factor when you swap x and y but keep the rest the same. Yet the integration range symmetry means that amounts to just renaming your variables, hence it must stay the same.
(3) Here is some code that shows it will be zero, at least if we zero out the singular part and a small band around it.
a = 1;
b = 1;
beta = 1;
eps[x_] := 2*(a-b*Cos[x])
n[x_] := 1/(1+Exp[beta*eps[x]])
delta = .001;
pw[x_,y_] := Piecewise[{{1,Abs[Abs[x]-Abs[y]]>delta}}, 0]
We add 1 to the integrand just to avoid accuracy issues with results that are near zero.
NIntegrate[1+Cos[(x+y)/2]^2*(n[x]-n[y])/(eps[x]-eps[y])^2*pw[Cos[x],Cos[y]],
{x,-Pi,Pi}, {y,-Pi,Pi}] / (4*Pi^2)
I get the result below.
NIntegrate::slwcon:
Numerical integration converging too slowly; suspect one of the following:
singularity, value of the integration is 0, highly oscillatory integrand,
or WorkingPrecision too small.
NIntegrate::eincr:
The global error of the strategy GlobalAdaptive has increased more than
2000 times. The global error is expected to decrease monotonically after a
number of integrand evaluations. Suspect one of the following: the
working precision is insufficient for the specified precision goal; the
integrand is highly oscillatory or it is not a (piecewise) smooth
function; or the true value of the integral is 0. Increasing the value of
the GlobalAdaptive option MaxErrorIncreases might lead to a convergent
numerical integration. NIntegrate obtained 39.4791 and 0.459541
for the integral and error estimates.
Out[24]= 1.00002
This is a good indication that the unadulterated result will be zero.
(4) Substituting cx for cos(x) and cy for cos(y), and removing extraneous factors for purposes of convergence assessment, gives the expression below.
((1 + E^(2*(1 - cx)))^(-1) - (1 + E^(2*(1 - cy)))^(-1))/
(2*(1 - cx) - 2*(1 - cy))^2
A series expansion in cy, centered at cx, indicates a pole of order 1. So it does appear to be a singular integral.
Daniel Lichtblau
The integral looks like a Cauchy Principal Value type integral (i.e. it has a strong singularity). That's why you can't apply standard quadrature techniques.
Have you tried PrincipalValue->True in Mathematica's Integrate?
In addition to Daniel's observation about integrating an odd integrand over a symmetric range (so that symmetry indicates the result should be zero), you can also do this to understand its convergence better (I'll use latex, writing this out with pen and paper should make it easier to read; it took a lot longer to write than to do, it's not that complicated):
First, epsilon(x)-\epsilon(y)\propto\cos(y)-\cos(x)=2\sin(\xi_+)\sin(\xi_-) where I have defined \xi_\pm=(x\pm y)/2 (so I've rotated the axes by pi/4). The region of integration then is \xi_+ between \pi/\sqrt{2} and -\pi/\sqrt{2} and \xi_- between \pm(\pi/\sqrt{2}-\xi_-). Then the integrand takes the form \frac{1}{\sin^2(\xi_-)\sin^2(\xi_+)} times terms with no divergences. So, evidently, there are second-order poles, and this isn't convergent as presented.
Perhaps you could email the persons who obtained an answer with the cos term and ask what precisely it is they did. Perhaps there's a physical regularisation procedure being employed. Or you could have given more information on the physical origin of this (some sort of second order perturbation theory for some sort of bosonic system?), had that not been off-topic here...
May be I am missing something here, but the integrand
f[x,y]=Cos^2[(x+y)/2]*(n[x]-n[y])/(eps[x]-eps[y]) with n[x]=1/(1+Exp[Beta*eps[x]]) and eps[x]=2(a-b*Cos[x]) is indeed a symmetric function in x and y: f[x,-y]= f[-x,y]=f[x,y].
Therefore its integral over any domain [-u,u]x[-v,v] is zero. No numerical integration seems to be needed here. The result is just zero.

How to find minimum of nonlinear, multivariate function using Newton's method (code not linear algebra)

I'm trying to do some parameter estimation and want to choose parameter estimates that minimize the square error in a predicted equation over about 30 variables. If the equation were linear, I would just compute the 30 partial derivatives, set them all to zero, and use a linear-equation solver. But unfortunately the equation is nonlinear and so are its derivatives.
If the equation were over a single variable, I would just use Newton's method (also known as Newton-Raphson). The Web is rich in examples and code to implement Newton's method for functions of a single variable.
Given that I have about 30 variables, how can I program a numeric solution to this problem using Newton's method? I have the equation in closed form and can compute the first and second derivatives, but I don't know quite how to proceed from there. I have found a large number of treatments on the web, but they quickly get into heavy matrix notation. I've found something moderately helpful on Wikipedia, but I'm having trouble translating it into code.
Where I'm worried about breaking down is in the matrix algebra and matrix inversions. I can invert a matrix with a linear-equation solver but I'm worried about getting the right rows and columns, avoiding transposition errors, and so on.
To be quite concrete:
I want to work with tables mapping variables to their values. I can write a function of such a table that returns the square error given such a table as argument. I can also create functions that return a partial derivative with respect to any given variable.
I have a reasonable starting estimate for the values in the table, so I'm not worried about convergence.
I'm not sure how to write the loop that uses an estimate (table of value for each variable), the function, and a table of partial-derivative functions to produce a new estimate.
That last is what I'd like help with. Any direct help or pointers to good sources will be warmly appreciated.
Edit: Since I have the first and second derivatives in closed form, I would like to take advantage of them and avoid more slowly converging methods like simplex searches.
The Numerical Recipes link was most helpful. I wound up symbolically differentiating my error estimate to produce 30 partial derivatives, then used Newton's method to set them all to zero. Here are the highlights of the code:
__doc.findzero = [[function(functions, partials, point, [epsilon, steps]) returns table, boolean
Where
point is a table mapping variable names to real numbers
(a point in N-dimensional space)
functions is a list of functions, each of which takes a table like
point as an argument
partials is a list of tables; partials[i].x is the partial derivative
of functions[i] with respect to 'x'
epilson is a number that says how close to zero we're trying to get
steps is max number of steps to take (defaults to infinity)
result is a table like 'point', boolean that says 'converged'
]]
-- See Numerical Recipes in C, Section 9.6 [http://www.nrbook.com/a/bookcpdf.php]
function findzero(functions, partials, point, epsilon, steps)
epsilon = epsilon or 1.0e-6
steps = steps or 1/0
assert(#functions > 0)
assert(table.numpairs(partials[1]) == #functions,
'number of functions not equal to number of variables')
local equations = { }
repeat
if Linf(functions, point) <= epsilon then
return point, true
end
for i = 1, #functions do
local F = functions[i](point)
local zero = F
for x, partial in pairs(partials[i]) do
zero = zero + lineq.var(x) * partial(point)
end
equations[i] = lineq.eqn(zero, 0)
end
local delta = table.map(lineq.tonumber, lineq.solve(equations, {}).answers)
point = table.map(function(v, x) return v + delta[x] end, point)
steps = steps - 1
until steps <= 0
return point, false
end
function Linf(functions, point)
-- distance using L-infinity norm
assert(#functions > 0)
local max = 0
for i = 1, #functions do
local z = functions[i](point)
max = math.max(max, math.abs(z))
end
return max
end
You might be able to find what you need at the Numerical Recipes in C web page. There is a free version available online. Here (PDF) is the chapter containing the Newton-Raphson method implemented in C. You may also want to look at what is available at Netlib (LINPack, et. al.).
As an alternative to using Newton's method the Simplex Method of Nelder-Mead is ideally suited to this problem and referenced in Numerical Recpies in C.
Rob
You are asking for a function minimization algorithm. There are two main classes: local and global. Your problem is least squares so both local and global minimization algorithms should converge to the same unique solution. Local minimization is far more efficient than global so select that.
There are many local minimization algorithms but one particularly well suited to least squares problems is Levenberg-Marquardt. If you don't have such a solver to hand (e.g. from MINPACK) then you can probably get away with Newton's method:
x <- x - (hessian x)^-1 * grad x
where you compute the inverse matrix multiplied by a vector using a linear solver.
Since you already have the partial derivatives, how about a general gradient-descent approach?
Maybe you think you have a good-enough solution, but for me, the easiest way to think about this is to understand it in the 1-variable case first, and then extend it to the matrix case.
In the 1-variable case, if you divide the first derivative by the second derivative, you get the (negative) step size to your next trial point, e.g. -V/A.
In the N-variable case, the first derivative is a vector and the second derivative is a matrix (the Hessian). You multiply the derivative vector by the inverse of the second derivative, and the result is the negative step-vector to your next trial point, e.g. -V*(1/A)
I assume you can get the 2nd-derivative Hessian matrix. You will need a routine to invert it. There are plenty of these around in various linear algebra packages, and they are quite fast.
(For readers who are not familiar with this idea, suppose the two variables are x and y, and the surface is v(x,y). Then the first derivative is the vector:
V = [ dv/dx, dv/dy ]
and the second derivative is the matrix:
A = [dV/dx]
[dV/dy]
or:
A = [ d(dv/dx)/dx, d(dv/dy)/dx]
[ d(dv/dx)/dy, d(dv/dy)/dy]
or:
A = [d^2v/dx^2, d^2v/dydx]
[d^2v/dxdy, d^2v/dy^2]
which is symmetric.)
If the surface is parabolic (constant 2nd derivative) it will get to the answer in 1 step. On the other hand, if the 2nd derivative is very not-constant, you could encounter oscillation. Cutting each step in half (or some fraction) should make it stable.
If N == 1, you'll see that it does the same thing as in the 1-variable case.
Good luck.
Added: You wanted code:
double X[N];
// Set X to initial estimate
while(!done){
double V[N]; // 1st derivative "velocity" vector
double A[N*N]; // 2nd derivative "acceleration" matrix
double A1[N*N]; // inverse of A
double S[N]; // step vector
CalculateFirstDerivative(V, X);
CalculateSecondDerivative(A, X);
// A1 = 1/A
GetMatrixInverse(A, A1);
// S = V*(1/A)
VectorTimesMatrix(V, A1, S);
// if S is small enough, stop
// X -= S
VectorMinusVector(X, S, X);
}
My opinion is to use a stochastic optimizer, e.g., a Particle Swarm method.