Let's say I have a set of samples, which consists of a non-stationary stochastic process with a uniform probability distribution (Gaussian). I need an adaptive linear regression over the set of samples. Basically I want the 'best-fit' line to behave a certain way. I have a separate signal, and I know the 'best-fit' line of the form Y=Mx+B will have a slope M proportional to that other signal. So I need the optimization problem to minimize the distance between the points BUT giving me a slope proportional to the other signal. What's the simplest machine learning/stats approach to use for this problem?
If i understand your question correctly, you can just use normal regression, or a gradient descent type algorithm, but instead of having the degrees of freedom as M and B, you can use a proportionality constant to M of the known data, and a separate B.
ie. the known signal:
Y1 = M1*x + B1
Y2 = k*M1*x + B2
solve for k and B2 such that the mean difference to x and y is minimised.
In theory, this seems to be intrinsic anyway. If you solved the problem for a linear solution in the first place. k would be M2 / M1 ....
Related
I am tasked to perform a prediction analysis. This requires performing a linear regression on several (~10) predictor variables and coming up with intercepts for all and a constant.
so final equation will be of this format y = c + c1x1 + c2x2 + c3x3....
Now I know that you can use fitlm function in MATLAB that is available with Statistics and Machine Learning Toolbox however at this point I don't know if we will be purchasing it. How do I perform linear regression on them ?
You can use the closed form solution of linear least squares.
C=inv(transpose(X)*X)*transpose(X)*y
In the above, make the first row of X all ones, and the following rows are x1, x2,...
C will contain the corresponding constants. The first entry in C is c.
From: https://www.mathworks.com/help/matlab/data_analysis/linear-regression.html
You can write your predictor variables as a matrix X using X = [ones(length(x1),1),x1,x2,x3,...,xn] and formulating the response variables Y as the equation Y = XB and doing a matrix inverse operation using mldivide as B = X\Y to find your regression coefficients.
I am working on a MR-physic simulation written in Matlab which simulates bloch's equations on an defined object. The magnetisation in the object is updated every time-step with the following functions.
function Mt = evolveMtrans(gamma, delta_B, G, T2, Mt0, delta_t)
% this function calculates precession and relaxation of the
% transversal component, Mt, of M
delta_phi = gamma*(delta_B + G)*delta_t;
Mt = Mt0 .* exp(-delta_t*1./T2 - 1i*delta_phi);
end
This function is a very small part of the entire code but is called upon up to 250.000 times and thus slows down the code and the performance of the entire simulation. I have thought about how I can speed up the calculation but haven't come up with a good solution. There is one line that is VERY time consuming and stands for approximately 50% - 60% of the overall simulation time. This is the line,
Mt = Mt0 .* exp(-delta_t*1./T2 - 1i*delta_phi);
where
Mt0 = 512x512 matrix
delta_t = a scalar
T2 = 512x512 matrix
delta_phi = 512x512 matrix
I would be very grateful for any suggestion to speed up this calculation.
More info below,
The function evovleMtrans is called every timestep during the simulation.
The parameters that are used for calling the function are,
gamma = a constant. (gyramagnetic constant)
delta_B = the magnetic field value
G = gradientstrength
T2 = a 512x512 matrix with T2-values for the object
Mstart.r = a 512x512 matrix with the values M.r had the last timestep
delta_t = a scalar with the difference in time since the last calculated M.r
The only parameters of these that changed during the simulation are,
G, Mstart.r and delta_t. The rest do not change their values during the simulation.
The part below is the part in the main code that calls the function.
% update phase and relaxation to calcTime
delta_t = calcTime - Mstart_t;
delta_B = (d-d0)*B0;
G = Sq.Gx*Sq.xGxref + Sq.Gz*Sq.zGzref;
% Precession around B0 (z-axis) and B1 (+-x-axis or +-y-axis)
% is defined clock-wise in a right hand system x, y, z and
% x', y', z (see the Bloch equation, Bloch 1946 and Levitt
% 1997). The x-axis has angle zero and the y-axis has angle 90.
% For flipping/precession around B1 in the xy-plane, z-axis has
% angle zero.
% For testing of precession direction:
% delta_phi = gamma*((ones(size(d)))*1e-6*B0)*delta_t;
M.r = evolveMtrans(gamma, delta_B, G, T2, Mstart.r, delta_t);
M.l = evolveMlong(T1, M0.l, Mstart.l, delta_t);
This is not a surprise.
That "single line" is a matrix equation. It's really 1,024 simultaneous equations.
Per Jannick, that first term means element-wise division, so "delta_t/T[i,j]". Multiplying a matrix by a scalar is O(N^2). Matrix addition is O(N^2). Evaluating exponential of a matrix will be O(N^2).
I'm not sure if I saw a complex argument in there as well. Does that mean complex matricies with real and imaginary entries? Does your equation simplify to real and imaginary parts? That means twice the number of computations.
Your best hope is to exploit symmetry as much as possible. If all your matricies are symmetric, you cut your calculations roughly in half.
Use parallelization if you can.
Algorithm choice can make a big difference, too. If you're using explicit Euler integration, you may have time step limitations due to stability concerns. Is that why you have 250,000 steps? Maybe a larger time step is possible with a more stable integration schema. Think about a higher order adaptive scheme with error correction, like 5th order Runge Kutta.
There are several possibilities to improve the speed of the code but all that I see come with a caveat.
Numerical ode integration
The first possibility would be to change your analytical solution by numerical differential equation solver. This has several advantages
The analytical solution includes the complex exponential function, which is costly to calculate, while the differential equation contains only multiplication and addition. (d/dt u = -a u => u=exp(-at))
There are plenty of built-in solvers for matlab available and they are typically pretty fast (e.g. ode45). The built-ins however all use a variable step size. This improves speed and accuracy but would be a problem if you really need a fixed equally spaced grid of time points. Here are unofficial fixed step solvers.
As a start you could also try to use just an euler step by replacing
M.r = evolveMtrans(gamma, delta_B, G, T2, Mstart.r, delta_t);
by
delta_phi = gamma*(delta_B + G)*t_step;
M.r += M.r .* (1-t_step*1./T2 - 1i*delta_phi);
You can then further improve that by precalculating all constant values, e.g. one_over_T1=1/T1, moving delta_phi out of the loop.
Caveat:
You are bound to a minimum step size or the accuracy suffers. Therefore this is only a good idea if you time-spacing is quite fine.
Less points in time
You should carfully analyze whether you really need so many points in time. It seems somewhat puzzling to me that you need so many points. As you know the full analytical solution you can freely choose how to sample the time and maybe use this to your advantage.
Going fortran
This might seem like a grand step but in my experience basic (simple loops, matrix operations etc.) matlab code can be relatively easily translated to fortran line-by-line. This would be especially helpful in addition to my first point. If you still want to use the full analytical solution probably there is not much to gain here because exp is already pretty fast in matlab.
I have the following equation:
I want to do a exponential curve fitting using MATLAB for the above equation, where y = f(u,a). y is my output while (u,a) are my inputs. I want to find the coefficients A,B for a set of provided data.
I know how to do this for simple polynomials by defining states. As an example, if states= (ones(size(u)), u u.^2), this will give me L+Mu+Nu^2, with L, M and N being regression coefficients.
However, this is not the case for the above equation. How could I do this in MATLAB?
Building on what #eigenchris said, simply take the natural logarithm (log in MATLAB) of both sides of the equation. If we do this, we would in fact be linearizing the equation in log space. In other words, given your original equation:
We get:
However, this isn't exactly polynomial regression. This is more of a least squares fitting of your points. Specifically, what you would do is given a set of y and set pair of (u,a) points, you would build a system of equations and solve for this system via least squares. In other words, given the set y = (y_0, y_1, y_2,...y_N), and (u,a) = ((u_0, a_0), (u_1, a_1), ..., (u_N, a_N)), where N is the number of points that you have, you would build your system of equations like so:
This can be written in matrix form:
To solve for A and B, you simply need to find the least-squares solution. You can see that it's in the form of:
Y = AX
To solve for X, we use what is called the pseudoinverse. As such:
X = A^{*} * Y
A^{*} is the pseudoinverse. This can eloquently be done in MATLAB using the \ or mldivide operator. All you have to do is build a vector of y values with the log taken, as well as building the matrix of u and a values. Therefore, if your points (u,a) are stored in U and A respectively, as well as the values of y stored in Y, you would simply do this:
x = [u.^2 a.^3] \ log(y);
x(1) will contain the coefficient for A, while x(2) will contain the coefficient for B. As A. Donda has noted in his answer (which I embarrassingly forgot about), the values of A and B are obtained assuming that the errors with respect to the exact curve you are trying to fit to are normally (Gaussian) distributed with a constant variance. The errors also need to be additive. If this is not the case, then your parameters achieved may not represent the best fit possible.
See this Wikipedia page for more details on what assumptions least-squares fitting takes:
http://en.wikipedia.org/wiki/Least_squares#Least_squares.2C_regression_analysis_and_statistics
One approach is to use a linear regression of log(y) with respect to u² and a³:
Assuming that u, a, and y are column vectors of the same length:
AB = [u .^ 2, a .^ 3] \ log(y)
After this, AB(1) is the fit value for A and AB(2) is the fit value for B. The computation uses Matlab's mldivide operator; an alternative would be to use the pseudo-inverse.
The fit values found this way are Maximum Likelihood estimates of the parameters under the assumption that deviations from the exact equation are constant-variance normally distributed errors additive to A u² + B a³. If the actual source of deviations differs from this, these estimates may not be optimal.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 3 years ago.
Improve this question
I'm going to write a program where the input is a data set of 2D points and the output is the regression coefficients of the line of best fit by minimizing the minimum MSE error.
I have some sample points that I would like to process:
X Y
1.00 1.00
2.00 2.00
3.00 1.30
4.00 3.75
5.00 2.25
How would I do this in MATLAB?
Specifically, I need to get the following formula:
y = A + Bx + e
A is the intercept and B is the slope while e is the residual error per point.
Judging from the link you provided, and my understanding of your problem, you want to calculate the line of best fit for a set of data points. You also want to do this from first principles. This will require some basic Calculus as well as some linear algebra for solving a 2 x 2 system of equations. If you recall from linear regression theory, we wish to find the best slope m and intercept b such that for a set of points ([x_1,y_1], [x_2,y_2], ..., [x_n,y_n]) (that is, we have n data points), we want to minimize the sum of squared residuals between this line and the data points.
In other words, we wish to minimize the cost function F(m,b,x,y):
m and b are our slope and intercept for this best fit line, while x and y are a vector of x and y co-ordinates that form our data set.
This function is convex, so there is an optimal minimum that we can determine. The minimum can be determined by finding the derivative with respect to each parameter, and setting these equal to 0. We then solve for m and b. The intuition behind this is that we are simultaneously finding m and b such that the cost function is jointly minimized by these two parameters. In other words:
OK, so let's find the first quantity :
We can drop the factor 2 from the derivative as the other side of the equation is equal to 0, and we can also do some distribution of terms by multiplying the -x_i term throughout:
Next, let's tackle the next parameter :
We can again drop the factor of 2 and distribute the -1 throughout the expression:
Knowing that is simply n, we can simplify the above to:
Now, we need to simultaneously solve for m and b with the above two equations. This will jointly minimize the cost function which finds the best line of fit for our data points.
Doing some re-arranging, we can isolate m and b on one side of the equations and the rest on the other sides:
As you can see, we can formulate this into a 2 x 2 system of equations to solve for m and b. Specifically, let's re-arrange the two equations above so that it's in matrix form:
With regards to above, we can decompose the problem by solving a linear system: Ax = b. All you have to do is solve for x, which is x = A^{-1}*b. To find the inverse of a 2 x 2 system, given the matrix:
The inverse is simply:
Therefore, by substituting our quantities into the above equation, we solve for m and b in matrix form, and it simplifies to this:
Carrying out this multiplication and solving for m and b individually, this gives:
As such, to find the best slope and intercept to best fit your data, you need to calculate m and b using the above equations.
Given your data specified in the link in your comments, we can do this quite easily:
%// Define points
X = 1:5;
Y = [1 2 1.3 3.75 2.25];
%// Get total number of points
n = numel(X);
% // Define relevant quantities for finding quantities
sumxi = sum(X);
sumyi = sum(Y);
sumxiyi = sum(X.*Y);
sumxi2 = sum(X.^2);
sumyi2 = sum(Y.^2);
%// Determine slope and intercept
m = (sumxi * sumyi - n*sumxiyi) / (sumxi^2 - n*sumxi2);
b = (sumxiyi * sumxi - sumyi * sumxi2) / (sumxi^2 - n*sumxi2);
%// Display them
disp([m b])
... and we get:
0.4250 0.7850
Therefore, the line of best fit that minimizes the error is:
y = 0.4250*x + 0.7850
However, if you want to use built-in MATLAB tools, you can use polyfit (credit goes to Luis Mendo for providing the hint). polyfit determines the line (or nth order polynomial curve rather...) of best fit by linear regression by minimizing the sum of squared errors between the best fit line and your data points. How you call the function is so:
coeff = polyfit(x,y,order);
x and y are the x and y points of your data while order determines the order of the line of best fit you want. As an example, order=1 means that the line is linear, order=2 means that the line is quadratic and so on. Essentially, polyfit fits a polynomial of order order given your data points. Given your problem, order=1. As such, given the data in the link, you would simply do:
X = 1:5;
Y = [1 2 1.3 3.75 2.25];
coeff = polyfit(X,Y,1)
coeff =
0.4250 0.7850
The way coeff works is that these are the coefficients of the regression line, starting from the highest order in decreasing value. As such, the above coeff variable means that the regression line was fitted as:
y = 0.4250*x + 0.7850
The first coefficient is the slope while the second coefficient is the intercept. You'll also see that this matches up with the link you provided.
If you want a visual representation, here's a plot of the data points as well as the regression line that best fits these points:
plot(X, Y, 'r.', X, polyval(coeff, X));
Here's the plot:
polyval takes an array of coefficients (usually produced by polyfit), and you provide a set of x co-ordinates and it calculates what the y values are given the values of x. Essentially, you are evaluating what the points are along the best fit line.
Edit - Extending to higher orders
If you want to extend so that you're finding the best fit for any nth order polynomial, I won't go into the details, but it boils down to constructing the following linear system. Given the relationship for the ith point between (x_i, y_i):
You would construct the following linear system:
Basically, you would create a vector of points y, and you would construct a matrix X such that each column denotes taking your vector of points x and applying a power operation to each column. Specifically, the first column is the zero-th power, the first column is the first power, the second column is the second power and so on. You would do this up until m, which is the order polynomial you want. The vector of e would be the residual error for each point in your set.
Specifically, the formulation of the problem can be written in matrix form as:
Once you construct this matrix, you would find the parameters by least-squares by calculating the pseudo-inverse. How the pseudo-inverse is derived, you can read it up on the Wikipedia article I linked to, but this is the basis for minimizing a system by least-squares. The pseudo-inverse is the backbone behind least-squares minimization. Specifically:
(X^{T}*X)^{-1}*X^{T} is the pseudo-inverse. X itself is a very popular matrix, which is known as the Vandermonde matrix and MATLAB has a command called vander to help you compute that matrix. A small note is that vander in MATLAB is returned in reverse order. The powers decrease from m-1 down to 0. If you want to have this reversed, you'd need to call fliplr on that output matrix. Also, you will need to append one more column at the end of it, which is the vector with all of its elements raised to the mth power.
I won't go into how you'd repeat your example for anything higher order than linear. I'm going to leave that to you as a learning exercise, but simply construct the vector y, the matrix X with vander, then find the parameters by applying the pseudo-inverse of X with the above to solve for your parameters.
Good luck!
I have a neural network with N input nodes and N output nodes, and possibly multiple hidden layers and recurrences in it but let's forget about those first. The goal of the neural network is to learn an N-dimensional variable Y*, given N-dimensional value X. Let's say the output of the neural network is Y, which should be close to Y* after learning. My question is: is it possible to get the inverse of the neural network for the output Y*? That is, how do I get the value X* that would yield Y* when put in the neural network? (or something close to it)
A major part of the problem is that N is very large, typically in the order of 10000 or 100000, but if anyone knows how to solve this for small networks with no recurrences or hidden layers that might already be helpful. Thank you.
If you can choose the neural network such that the number of nodes in each layer is the same, and the weight matrix is non-singular, and the transfer function is invertible (e.g. leaky relu), then the function will be invertible.
This kind of neural network is simply a composition of matrix multiplication, addition of bias and transfer function. To invert, you'll just need to apply the inverse of each operation in the reverse order. I.e. take the output, apply the inverse transfer function, multiply it by the inverse of the last weight matrix, minus the bias, apply the inverse transfer function, multiply it by the inverse of the second to last weight matrix, and so on and so forth.
This is a task that maybe can be solved with autoencoders. You also might be interested in generative models like Restricted Boltzmann Machines (RBMs) that can be stacked to form Deep Belief Networks (DBNs). RBMs build an internal model h of the data v that can be used to reconstruct v. In DBNs, h of the first layer will be v of the second layer and so on.
zenna is right.
If you are using bijective (invertible) activation functions you can invert layer by layer, subtract the bias and take the pseudoinverse (if you have the same number of neurons per every layer this is also the exact inverse, under some mild regularity conditions).
To repeat the conditions: dim(X)==dim(Y)==dim(layer_i), det(Wi) not = 0
An example:
Y = tanh( W2*tanh( W1*X + b1 ) + b2 )
X = W1p*( tanh^-1( W2p*(tanh^-1(Y) - b2) ) -b1 ), where W2p and W1p represent the pseudoinverse matrices of W2 and W1 respectively.
The following paper is a case study in inverting a function learned from Neural Networks. It is a case study from the industry and looks a good beginning for understanding how to go about setting up the problem.
An alternate way of approaching the task of getting the desired x that yields desired y would be start with random x (or input as seed), then through gradient decent (similar algorithm to back propagation, difference being that instead of finding derivatives of weights and biases, you find derivatives of x. Also, mini batching is not needed.) repeatedly adjust x until it yields a y that is close to the desired y. This approach has an advantage that it allows an input of a seed (starting x, if not randomly selected). Also, I have a hypothesis that the final x will have some similarity to initial x(seed), which would imply that this algorithm has the ability to transpose, depending on the context of the neural network application.