We are trying an alteration optimization strategy to solve Lyapunov problems.
We break down our decision variables into two sets, Set 1 and Set 2.
We were perplexed how it was possible that, after getting a solution to Set 1, and plugging in those solved variables into the optimization over Set 2, the transferred variables would not be feasible.
The constraints that fail are those due to the SOS coefficient matching equality constraints.
Here, we print in each row the constraint that failed, and the value of our Initial Guess. We can see that the Initial Guess is off only a very small amount compared to the constraints.
LinearEqualityConstraint
(2 * Symmetric(97,40) + 2 * Symmetric(96,41)) == 9.50028
[9.50027496]
LinearEqualityConstraint
(2 * Symmetric(97,47) + 2 * Symmetric(96,48)) == 234.465
[234.4647013]
LinearEqualityConstraint
(2 * Symmetric(97,54) + 2 * Symmetric(96,55)) == -234.463
[-234.46336504]
LinearEqualityConstraint
(2 * Symmetric(97,61) + 2 * Symmetric(96,62)) == 12.7962
[12.79618825]
LinearEqualityConstraint
(2 * Symmetric(97,68) + 2 * Symmetric(96,69)) == -12.7964
[-12.79637068]
LinearEqualityConstraint
(2 * Symmetric(97,75) + 2 * Symmetric(96,76)) == -51.4061
[-51.40605828]
LinearEqualityConstraint
(2 * Symmetric(97,81) + 2 * Symmetric(96,82)) == 51.406
[51.40604213]
LinearEqualityConstraint
(2 * Symmetric(97,86) + 2 * Symmetric(96,87)) == 192.794
[192.79430158]
LinearEqualityConstraint
(2 * Symmetric(97,90) + 2 * Symmetric(96,91)) == -141.924
[-141.92366183]
LinearEqualityConstraint
(2 * Symmetric(97,93) + 2 * Symmetric(96,94)) == -37.6674
[-37.66740401]
InitialGuess V_sos and
Our guess for what's happening is:
When you extract the solution from one optimization using result.GetSolution(var), you lose some precision.
Or, when you set the previous solution using prog.SetInitialGuess(np_array) you lose some precision.
What's the solution here? Should we just keep feeding the solution back in even though it says infeasible?
This is a partial cookbook when I debug SOS problem, especially when working with Lyapunov problems:
Choose the right monomial basis
The main idea is to remove the 0-th order monomial 1 from the monomial basis of the sos polynomial. Here is a quick explanation:
The mathematical problem is
Find λ(x)
−Vdot − λ(x) * (ρ − V) is sos
λ(x) is sos
Namely you want to prove that V≤ρ ⇒ Vdot ≤ 0
So first I would suggest to re-write your dynamics to make sure that 0 is the goal state (you can always shift your state).
Second you can see that since x=0 is the equilibrium point, then both V(0) = 0 and Vdot(0) = 0 (Because x=0 is the global minimal of V(x), hence ∂V/∂x=0 at x=0, indicating Vdot(0) = 0), now your sos polynomial p(x) = −Vdot − λ(x) * (ρ − V) must satisfy p(0) = -λ(0) * ρ. But λ(x) >= 0 and ρ > 0, so we know λ(0) = 0.
Lemma
If a sos polynomial s(x) satisfies s(0) = 0, then its monomial basis cannot contain the 0-th order monomial (namely 1).
Proof
Remember that s(x) is a sos polynomial, namely
s(x) = m(x)ᵀQm(x)
where m(x) contains the monomial basis, and Q is a psd matrix. Now let's decompose the monomial basis m(x) into two parts, the 0-th order monomial 1 and the remaining monomials mbar(x). For example, if m(x) = [x1, x2, 1], then mbar(x) = [x1, x2]. We also decompose the psd matrix Q accordingly
s(x) = [mbar(x)]ᵀ [Q11 Q10] [mbar(x)]
[ 1] [Q10 Q00] [ 1]
Since s(0) = Q00 = 0, we also know that Q10 = 0, so now we can use a smaller psd matrix Q11 rather than Q. Equivalently we write s(x) = mbar(x)ᵀ * Q11 * mbar(x), where mbar(x) is the monomial basis that doesn't contain the 0-th order monomial, QED.
So why removing the 0-th order monomial from the monomial basis and use the smaller psd matrix Q11 is a good idea when your sos polynomial s(x) satisfies s(0) = 0? The reason is that if the monomial basis contains 1, then your psd matrix Q has to be on the boundary of the psd cone, namely your SDP problem doesn't have a strict interior. This could leads to violation of Slater's condition, which also breaks the strong duality. One example is that if your s(x) = x², by including the 0'th order monomial, it is written as
x² = [x] [1 0] [x]
[1] [0 0] [1]
And you see that the Gram matrix [[1 0], [0, 0]] is on the boundary of the psd cone (with one eigen value equal to 0). But if you remove 1 from the monomial basis, then its Gram matrix is just Q11=1, strictly in the interior of the psd cone.
In Drake, after removing 1 from the monomial basis, you can create your sos polynomial λ(x) as
lambda_poly, lambda_gram = prog.NewSosPolynomial(monomial_basis)
whee monomial_basis doesn't contain the 0-th order monomial.
Backoff during bilinear alternation
This is a typical problem in bilinear alternation. The issue is that when you solve a conic optimization problem with an objective function, the optimal solution always occurs at the boundary of the cone, namely it is very close to being infeasible. Then when you fix some variables to this solution at the cone boundary, in the next iteration the problem is very likely infeasible due to numerical roundoff error.
A typical solution is that after solving the optimization problem on variable Set 1 with an objective, now "backoff" a little bit by solving a feasibility problem on variable Set 1. This new solution is often strictly feasible (namely it is inside the strict interior of the cone), now pass this strictly feasible solution Set 1 to the next iteration and search for Set 2.
More concretely, suppose at one iteration you solve the following optimization problem
min c'*x
s.t constraint_on_x
and denote the optimal cost as p. Now solve a new feasibility problem
find x
s.t c'*x <= p + epsilon
constraint_on_x
where epsilon can be a small positive number. This new solution will be used in the next iteration to search for a different set of variables.
You can check if your solution is on the boundary of the positive semidefinite cone by checking the Eigen value of your psd matrix. Here is the pseudo-code
for binding : prog.positive_semidefinite_constraints():
psd_sol = result.GetSolution(binding.variables())
psd_sol.reshape((binding.evaluator().matrix_rows(), binding.evaluator().matrix_rows()))
print(f"minimal eigenvalue {np.linalg.eig(psd_sol)[0].min()}")
You should see that before doing this "backoff" some of the minimal eigen value is almost 0. After "backoff" the minimal eigen value gets larger.
Related
I am a student working with time-series data which we feed into a neural network for classification (my task is to build and train this NN).
We're told to use a band-pass filter of 10 Hz to 150 Hz since anything outside that is not interesting.
After applying the band-pass, I've also down-sampled the data to 300 samples per second (originally it was 768 Hz). My understanding of the Shannon Nyquist sampling theorem is that, after applying the band-pass, any information in the data will be perfectly preserved at this sample-rate.
However, I got into a discussion with my supervisor who claimed that 300 Hz might not be sufficient even if the signal was band-limited. She says that it is only the minimum sample rate, not necessarily the best sample rate.
My understanding of the sampling theorem makes me think the supervisor is obviously wrong, but I don't want to argue with my supervisor, especially in case I'm actually the one who has misunderstood.
Can anyone help to confirm my understanding or provide some clarification? And how should I take this up with my supervisor (if at all).
The Nyquist-Shannon theorem states that the sampling frequency should at-least be twice of bandwidth, i.e.,
fs > 2B
So, this is the minimal criteria. If the sampling frequency is less than 2B then there will be aliasing. There is no upper limit on sampling frequency, but more the sampling frequency, the better will be the reconstruction.
So, I think your supervisor is right in saying that it is the minimal condition and not the best one.
Actually, you and your supervisor are both wrong. The minimum sampling rate required to faithfully represent a real-valued time series whose spectrum lies between 10 Hz and 150 Hz is 140 Hz, not 300 Hz. I'll explain this, and then I'll explain some of the context that shows why you might want to "oversample", as it is referred to (spoiler alert: Bailian-Low Theorem). The supervisor is mixing folklore into the discussion, and when folklore is not properly-contexted, it tends to telephone tag into fakelore. (That's a common failing even in the peer-reviewed literature, by the way). And there's a lot of fakelore, here, that needs to be defogged.
For the following, I will use the following conventions.
There's no math layout on Stack Overflow (except what we already have with UTF-8), so ...
a^b denotes a raised to the power b.
∫_I (⋯x⋯) dx denotes an integral of (⋯x⋯) taken over all x ∈ I, with the default I = ℝ.
The support supp φ (or supp_x φ(x) to make the "x" explicit) of a function φ(x) is the smallest closed set containing all the x-es for which φ(x) ≠ 0. For regularly-behaving (e.g. continuously differentiable) functions that means a union of closed intervals and/or half-rays or the whole real line, itself. This figures centrally in the Shannon-Nyquist sampling theorem, as its main condition is that a spectrum have bounded support; i.e. a "finite bandwidth".
For the Fourier transform I will use the version that has the 2π up in the exponent, and for added convenience, I will use the convention 1^x = e^{2πix} = cos(2πx) + i sin(2πx) (which I refer to as the Ramanujan Convention, as it is the convention I frequently used in my previous life oops I mean which Ramanujan secretly used in his life to make the math a whole lot simpler).
The set ℤ = {⋯, -2, -1, 0, +1, +2, ⋯ } is the integers, and 1^{x+z} = 1^x for all z∈ℤ - making 1^x the archetype of a periodic function whose period is 1.
Thus, the Fourier transform f̂(ν) of a function f(t) and its inverse are given by:
f̂(ν) = ∫ f(t) 1^{-νt} dt, f(t) = ∫ f̂(ν) 1^{+νt} dν.
The spectrum of the time series given by the function f(t) is the function f̂(ν) of the cyclic frequency ν, which is what is measured in Hertz (Hz.); t, itself, being measured in seconds. A common convention is to use the angular frequency ω = 2πν, instead, but that muddies the picture.
The most important example, with respect to the issue at hand, is the Fourier transform χ̂_Ω of the interval function given by χ_Ω(t) = 1 if t ∈ [-½Ω,+½Ω] and χ_Ω(t) = 0 else:
χ̂_Ω(t) = ∫_[-½Ω,+½Ω] 1^ν dν
= {1^{+½Ω} - 1^{-½Ω}}/{2πi}
= {2i sin πΩ}/{2πi}
= Ω sinc πΩ
which is where the function sinc x = (sin πx)/(πx) comes into play.
The cardinal form of the sampling theorem is that a function f(t) can be sampled over an equally-spaced sampled domain T ≡ { kΔt: k ∈ ℤ }, if its spectrum is bounded by supp f̂ ⊆ [-½Ω,+½Ω] ⊆ [-1/(2Δt),+1/(2Δt)], with the sampling given as
f(t) = ∑_{t'∈T} f(t') Ω sinc(Ω(t - t')) Δt.
So, this generally applies to [over-]sampling with redundancy factors 1/(ΩΔt) ≥ 1. In the special case where the sampling is tight with ΩΔt = 1, then it reduces to the form
f(t) = ∑_{t'∈T} f(t') sinc({t - t'}/Δt).
In our case, supp f̂ = [10 Hz., 150 Hz.] so the tightest fits are with 1/Δt = Ω = 300 Hz.
This generalizes to equally-spaced sampled domains of the form T ≡ { t₀ + kΔt: k ∈ ℤ } without any modification.
But it also generalizes to frequency intervals supp f̂ = [ν₋,ν₊] of width Ω = ν₊ - ν₋ and center ν₀ = ½ (ν₋ + ν₊) to the following form:
f(t) = ∑_{t'∈T} f(t') 1^{ν₀(t - t')} Ω sinc(Ω(t - t')) Δt.
In your case, you have ν₋ = 10 Hz., ν₊ = 150 Hz., Ω = 140 Hz., ν₀ = 80 Hz. with the condition Δt ≤ 1/140 second, a sampling rate of at least 140 Hz. with
f(t) = (140 Δt) ∑_{t'∈T} f(t') 1^{80(t - t')} sinc(140(t - t')).
where t and Δt are in seconds.
There is a larger context to all of this. One of the main places where this can be used is for transforms devised from an overlapping set of windowed filters in the frequency domain - a typical case in point being transforms for the time-scale plane, like the S-transform or the continuous wavelet transform.
Since you want the filters to be smoothly-windowed functions, without sharp corners, then in order for them to provide a complete set that adds up to a finite non-zero value over all of the frequency spectrum (so that they can all be normalized, in tandem, by dividing out by this sum), then their respective supports have to overlap.
(Edit: Generalized this example to cover both equally-spaced and logarithmic-spaced intervals.)
One example of such a set would be filters that have end-point frequencies taken from the set
Π = { p₀ (α + 1)ⁿ + β {(α + 1)ⁿ - 1} / α: n ∈ {0,1,2,⋯} }
So, for interval n (counting from n = 0), you would have ν₋ = p_n and ν₊ = p_{n+1}, where the members of Π are enumerated
p_n = p₀ (α + 1)ⁿ + β {(α + 1)ⁿ - 1} / α,
Δp_n = p_{n+1} - p_n = α p_n + β = (α p₀ + β)(α + 1)ⁿ,
n ∈ {0,1,2,⋯}
The center frequency of interval n would then be ν₀ = p_n + ½ Δp₀ (α + 1)ⁿ and the width would be Ω = Δp₀ (α + 1)ⁿ, but the actual support for the filter would overlap into a good part of the neighboring intervals, so that when you add up the filters that cover a given frequency ν the sum doesn't drop down to 0 as ν approaches any of the boundary points. (In the limiting case α → 0, this produces an equally-spaced frequency domain, suitable for an equalizer, while in the case β → 0, it produces a logarithmic scale with base α + 1, where octaves are equally-spaced.)
The other main place where you may apply this is to time-frequency analysis and spectrograms. Here, the role of a function f and its Fourier transform f̂ are reversed and the role of the frequency bandwidth Ω is now played by the (reciprocal) time bandwidth 1/Ω. You want to break up a time series, given by a function f(t) into overlapping segments f̃(q,λ) = g(λ)* f(q + λ), with smooth windowing given by the functions g(λ) with bounded support supp g ⊆ [-½ 1/Ω, +½ 1/Ω], and with interval spacing Δq much larger than the time sampling Δt (the ratio Δq/Δt is called the "hop" factor). The analogous role of Δt is played, here, by the frequency interval in the spectrogram Δp = Ω, which is now constant.
Edit: (Fixed the numbers for the Audacity example)
The minimum sampling rate for both supp_λ g and supp_λ f(q,λ) is Δq = 1/Ω = 1/Δp, and the corresponding redundancy factor is 1/(ΔpΔq). Audacity, for instance, uses a redundancy factor of 2 for its spectrograms. A typical value for Δp might be 44100/2048 Hz., while the time-sampling rate is Δt = 1/(2×3×5×7)² second (corresponding to 1/Δt = 44100 Hz.). With a redundancy factor of 2, Δq would be 1024/44100 second and the hop factor would be Δq/Δt = 1024.
If you try to fit the sampling windows, in either case, to the actual support of the band-limited (or time-limited) function, then the windows won't overlap and the only way to keep their sum from dropping to 0 on the boundary points would be for the windowing functions to have sharp corners on the boundaries, which would wreak havoc on their corresponding Fourier transforms.
The Balian-Low Theorem makes the actual statement on the matter.
https://encyclopediaofmath.org/wiki/Balian-Low_theorem
And a shout-out to someone I've been talking with, recently, about DSP-related matters and his monograph, which provides an excellent introductory reference to a lot of the issues discussed here.
A Friendly Guide To Wavelets
Gerald Kaiser
Birkhauser 1994
He said it's part of a trilogy, another installment of which is forthcoming.
I'm writing a k-means algorithm. At each step, I want to compute the distance of my n points to k centroids, without a for loop, and for d dimensions.
The problem is I have a hard time splitting on my number of dimensions with the Matlab functions I know. Here is my current code, with x being my n 2D-points and y my k centroids (also 2D-points of course), and with the points distributed along dimension 1, and the spatial coordinates along the dimension 2:
dist = #(a,b) (a - b).^2;
dx = bsxfun(dist, x(:,1), y(:,1)'); % x is (n,1) and y is (1,k)
dy = bsxfun(dist, x(:,2), y(:,2)'); % so the result is (n,k)
dists = dx + dy; % contains the square distance of each points to the k centroids
[_,l] = min(dists, [], 2); % we then argmin on the 2nd dimension
How to vectorize furthermore ?
First edit 3 days later, searching on my own
Since asking this question I made progress on my own towards vectorizing this piece of code.
The code above runs in approximately 0.7 ms on my example.
I first used repmat to make it easy to do broadcasting:
dists = permute(permute(repmat(x,1,1,k), [3,2,1]) - y, [3,2,1]).^2;
dists = sum(dists, 2);
[~,l] = min(dists, [], 3);
As expected it is slightly slower since we replicate the matrix, it runs at 0.85 ms.
From this example it was pretty easy to use bsxfun for the whole thing, but it turned out to be extremely slow, running in 150 ms so more than 150 times slower than the repmat version:
dist = #(a, b) (a - b).^2;
dists = permute(bsxfun(dist, permute(x, [3, 2, 1]), y), [3, 2, 1]);
dists = sum(dists, 2);
[~,l] = min(dists, [], 3);
Why is it so slow ? Isn't vectorizing always an improvement on speed, since it uses vector instructions on the CPU ? I mean of course simple for loops could be optimized to use it aswell, but how can vectorizing make the code slower ? Did I do it wrong ?
Using a for loop
For the sake of completeness, here's the for loop version of my code, surprisingly the fastest running in 0.4 ms, not sure why..
for i=1:k
dists(:,i) = sum((x - y(i,:)).^2, 2);
endfor
[~,l] = min(dists, [], 2);
Note: This answer was written when the question was also tagged MATLAB. Links to Octave documentation added after the MATLAB tag was removed.
You can use the pdist2MATLAB/Octave function to calculate pairwise distances between two sets of observations.
This way, you offload the bother of vectorization to the people who wrote MATLAB/Octave (and they have done a pretty good job of it)
X = rand(10,3);
Y = rand(5,3);
D = pdist2(X, Y);
D is now a 10x5 matrix where the i, jth element is the distance between the ith X and jth Y point.
You can pass it the kind of distance you want as the third argument -- e.g. 'euclidean', 'minkowski', etc, or you could pass a function handle to your custom function like so:
dist = #(a,b) (a - b).^2;
D = pdist2(X, Y, dist);
As saastn mentions, pdist2(..., 'smallest', k) makes things easier in k-means. This returns just the smallest k values from each column of pdist2's result. Octave doesn't have this functionality, but it's easily replicated using sort()MATLAB/Octave.
D_smallest = sort(D);
D_smallest = D_smallest(1:k, :);
I understand well perceptron so put accent only on kernel but I am not familiar with matemathic expressions so please give me an numerical example and a guide on kernel.
For example:
My hyperplane of perceptron is x1*w1+x2*w2+x3*w3+b=0; The RBF kernel formula: k(x,z) = exp((-|x-z|^2)/2*variance^2) where takes action the radial basis function kernel here. Is x an input and what is z variable here?
Or of what I have to calculate variance if it is variance in the formula?
Somewhere I have understood so that I have to plug this formula in perceptron decision function x1*w1+x2*w2+x3*w3+b=0; but how does it look look like If I plug in?
I would like to ask a numerical example to avoid confusion.
Linear Perceptron
As you know linear perceptrons can be trained for binary classification. More precisely, if there is n features, x1, x2, ..., xn in n-dimensional space, Rn, and you want to label them in 2 categories, y1 & y2 (usually -1 and +1), you can use linear perceptron which defines a hyperplane w1*x1 + ... + wn*xn + b = 0 to do so.
w1*x1 + ... + wn*xn + b > 0 or W.X + b > 0 ==> class = y1
w1*x1 + ... + wn*xn + b < 0 or W.X + b < 0 ==> class = y2
Linear perceptron will work well, only if the problem is linearly separable in Rn. For example, in 2D space, this means that one line can separate the 2 sets of points.
Algorithm
One common algorithm to train the perceptron, i.e., find weights and bias, w's & b, based on N data points, X1, ..., XN, and their labels, Y1, ..., YN is the following:
Initialize: W = zeros(n,1); b = 0
For i=1 to N:
Calculate F(Xi) = W.Xi + b
If F(Xi)*Yi <= 0:
W <--- W + Xi*Yi
b <--- b + Yi
This will give the final value for W & b. Besides, based on the training, W will be a linear combination of training points, Xi's, more precisely, the ones that were misclassified. So W = a1*X1 + ... + ...aN*XN where a's are in {0,y1,y2}.
Now, if there is a new point, let's say Z, to label, we check the sign of F(Z) = W.Z + b = a1*(X1.Z) + ... + aN*(XN.Z) + b. It is interesting that only the inner product of new point and training points take part in it.
Kernel Perceptron
Now, if the problem is not linearly separable, one may try to go to a higher dimensional space in which a hyperplane can do the classification. As an example, consider a circle in 2D space. The points inside and outside of the circle can't be separated by a line. However, if you find a transformation that can take the points to 3D space such that the first 2 coordinates remain the same for all points, and the 3rd coordinate become +1 and -1 for the points inside and outside of the circle respectively, then a plane defined as 3rd coordinate = 0 can separate the points.
Finding such transformations can be difficult and computationally heavy, so the kernel trick is introduced. Notice that we only used the inner product of new points with the training points. Kernel trick employs this fact and defines the inner product of the transformed points without actually finding the transformation.
If the unknown transformation is P(X) then Kernel function will be:
K(Xi,Xj) = <P(Xi),P(Xj)>. So instead of finding P, kernel functions are defined which represent the scalar result of the inner product in high-dimensional space. There are also theorems about what functions can be kernel functions, i.e., correspond to inner product in another space.
After choosing a kernel function, the algorithm will be modified as follows:
Initialize: F(X) = 0
For i=1 to N:
Calculate F(Xi)
If F(Xi)*Yi <= 0:
F(.) <--- F(.) + K(.,Xi)*Yi + Yi
At the end, F(.) = a1*K(.,X1) + ... + ...aN*K(.,XN) + b where a's are in {0,y1,y2}.
RBF Kernel
Radial basis function is one type of kernel function that is actually computing the inner product in an infinite-dimensional space. It can be written as
K(Xi,Xj) = exp(- norm2(Xi-Xj)^2 / (2*sigma^2))
Sigma is some parameter that you can work with to find an optimum value for. For example, you can train the model with different values of sigma and then find the best value based on the performance. You can start with sigma = 1
After training the model to find F(.), for a new data Z, the sign of F(Z) = a1*K(Z,X1) + ... + ...aN*K(Z,XN) + b will determine the class.
Remarks:
Regarding to your question about variance, you don't need to find any variance.
About x and z in your question, in each iteration, you should find the kernel output for the current data point and all the previously added points (the points that were misclassified and hence were added to F).
I couldn't come up with a simple instructive numerical example.
References:
I borrowed some notation from
https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=2&cad=rja&uact=8&ved=0ahUKEwjVu-fXo8DOAhVDxCYKHQkcDDAQFggoMAE&url=http%3A%2F%2Falex.smola.org%2Fteaching%2Fpune2007%2Fpune_3.pdf&usg=AFQjCNHlxy9TnY8xNe2-QDERipN_GycSqQ&bvm=bv.129422649,d.eWE
I am working on replicating a neural network. I'm trying to get an understanding of how the standard layer types work. In particular, I'm having trouble finding a description anywhere of how cross-channel normalisation layers behave on the backward-pass.
Since the normalization layer has no parameters, I could guess two possible options:
The error gradients from the next (i.e. later) layer are passed backwards without doing anything to them.
The error gradients are normalized in the same way the activations are normalized across channels in the forward pass.
I can't think of a reason why you'd do one over the other based on any intuition, hence why I'd like some help on this.
EDIT1:
The layer is a standard layer in caffe, as described here http://caffe.berkeleyvision.org/tutorial/layers.html (see 'Local Response Normalization (LRN)').
The layer's implementation in the forward pass is described in section 3.3 of the alexNet paper: http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf
EDIT2:
I believe the forward and backward pass algorithms are described in both the Torch library here: https://github.com/soumith/cudnn.torch/blob/master/SpatialCrossMapLRN.lua
and in the Caffe library here: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/lrn_layer.cpp
Please could anyone who is familiar with either/both of these translate the method for the backward pass stage into plain english?
It uses the chain rule to propagate the gradient backwards through the local response normalization layer. It is somewhat similar to a nonlinearity layer in this sense (which also doesn't have trainable parameters on its own, but does affect gradients going backwards).
From the code in Caffe that you linked to I see that they take the error in each neuron as a parameter, and compute the error for the previous layer by doing following:
First, on the forward pass they cache a so-called scale, that is computed (in terms of AlexNet paper, see the formula from section 3.3) as:
scale_i = k + alpha / n * sum(a_j ^ 2)
Here and below sum is sum indexed by j and goes from max(0, i - n/2) to min(N, i + n/2)
(note that in the paper they do not normalize by n, so I assume this is something that Caffe does differently than AlexNet). Forward pass is then computed as b_i = a_i + scale_i ^ -beta.
To backward propagate the error, let's say that the error coming from the next layer is be_i, and the error that we need to compute is ae_i. Then ae_i is computed as:
ae_i = scale_i ^ -b * be_i - (2 * alpha * beta / n) * a_i * sum(be_j * b_j / scale_j)
Since you are planning to implement it manually, I will also share two tricks that Caffe uses in their code that makes the implementation simpler:
When you compute the addends for the sum, allocate an array of size N + n - 1, and pad it with n/2 zeros on each end. This way you can compute the sum from i - n/2 to i + n/2, without caring about going below zero and beyond N.
You don't need to recompute the sum on each iteration, instead compute the the addends in advance (a_j^2 for the front pass, be_j * b_j / scale_j for the backward pass), then compute the sum for i = 0, and then for each consecutive i just add addend[i + n/2] and subtract addend[i - n/2 - 1], it will give you the value of the sum for the new value of i in constant time.
Of cause,you can either print the variables to observe the changes with them or use the debug model to see how errors change during passing the net.
I have an alternative formulation of the backward and I don't know if it is equivalent to caffe's:
So caffe's is :
ae_i = scale_i ^ -b * be_i - (2 * alpha * beta / n) * a_i * sum(be_j * b_j / scale_j)
by differentiating the original expression
b_i = a_i/(scale_i^-b)
I get
ae_i = scale_i ^ -b * be_i - (2 * alpha * beta / n) * a_i * be_i*sum(ae_j)/scale_i^(-b-1)
If I have three points that create an angle, what would be the best way to determine if a fourth point resides within the angle created by the previous three?
Currently, I determine the angle of the line to all three points from the origin point, and then check to see if the test angle is in between the two other angles but I'm trying to figure out if there's a better way to do it. The function is run tens of thousands of times an update and I'm hoping that there's a better way to achieve what I'm trying to do.
Let's say you have angle DEF (E is the "pointy" part), ED is the left ray and EF is the right ray.
* D (Dx, Dy)
/
/ * P (Px, Py)
/
/
*---------------*
E (Ex, Ey) F (Fx, Fy)
Step 1. Build line equation for line ED in the classic Al * x + Bl * y + Cl = 0 form, i.e. simply calculate
Al = Dy - Ey // l - for "left"
Bl = -(Dx - Ex)
Cl = -(Al * Ex + Bl * Ey)
(Pay attention to the subtraction order.)
Step 2. Build line equation for line FE (reversed direction) in the classic Ar * x + Br * y + Cr = 0 form, i.e. simply calculate
Ar = Ey - Fy // r - for "right"
Br = -(Ex - Fx)
Cr = -(Ar * Ex + Br * Ey)
(Pay attention to the subtraction order.)
Step 3. For your test point P calculate the expressions
Sl = Al * Px + Bl * Py + Cl
Sr = Ar * Px + Br * Py + Cr
Your point lies inside the angle if and only if both Sl and Sr are positive. If one of them is positive and other is zero, your point lies on the corresponding side ray.
That's it.
Note 1: For this method to work correctly, it is important to make sure that the left and right rays of the angle are indeed left and right rays. I.e. if you think about ED and EF as clock hands, the direction from D to F should be clockwise. If it is not guaranteed to be the case for your input, then some adjustments are necessary. For example, it can be done as an additional step of the algorithm, inserted between steps 2 and 3
Step 2.5. Calculate the value of Al * Fx + Bl * Fy + Cl. If this value is negative, invert signs of all ABC coefficients:
Al = -Al, Bl = -Bl, Cl = -Cl
Ar = -Ar, Br = -Br, Cr = -Cr
Note 2: The above calculations are made under assumption that we are working in a coordinate system with X axis pointing to the right and Y axis pointing to the top. If one of your coordinate axes is flipped, you have to invert the signs of all six ABC coefficients. Note, BTW, that if you perform the test described in step 2.5 above, it will take care of everything automatically. If you are not performing step 2.5 then you have to take the axis direction into account from the very beginning.
As you can see, this a precise integer method (no floating point calculations, no divisions). The price of that is danger of overflows. Use appropriately sized types for multiplications.
This method has no special cases with regard to line orientations or the value of the actual non-reflex angle: it work immediately for acute, obtuse, zero and straight angle. It can be easily used with reflex angles (just perform a complementary test).
P.S. The four possible combinations of +/- signs for Sl and Sr correspond to four sectors, into which the plane is divided by lines ED and EF.
* D
/
(-,+) / (+,+)
/
-------*------------* F
/ E
(-,-) / (+,-)
/
By using this method you can perform the full "which sector the point falls into" test. For an angle smaller than 180 you just happen to be interested in only one of those sectors: (+, +). If at some point you'll need to adapt this method for reflex angles as well (angles greater than 180), you will have to test for three sectors instead of one: (+,+), (-,+), (+,-).
Describe your origin point O, and the other 2 points A and B then your angle is given as AOB. Now consider your test point and call that C as in the diagram.
Now consider that we can get a vector equation of C by taking some multiple of vector OA and some multiple of OB. Explicitly
C = K1 x OA + K2 OB
for some K1,K2 that we need to calculate. Set O to the origin by subtracting it (vectorially) from all other points. If coordinates of A are (a1,a2), B = (b1,b2) and C = (c1,c2) we have in matrix terms
[ a1 b1 ] [ K1 ] = [ c1 ]
[ a2 b2 ] [ K2 ] = [ c2 ]
So we can solve for K1 and K2 using the inverse of the matrix to give
1 / (a1b2 - b1a2) [ b2 -b1 ] [ c1 ] = [ K1 ]
[ -a2 a1 ] [ c2 ] = [ K2 ]
which reduces to
K1 = (b2c1 - b1c2)/(a1b2 - b1a2)
K2 = (-a2c1 + a1c2)/(a1b2 - b1a2)
Now IF the point C lies within your angle, the multiples of the vectors OA and OB will BOTH be positive. If C lies 'under' OB, then we need a negative amount of OA to get to it similarly for the other direction. So your condition is satisfied when both K1 and K2 are greater than (or equal to) zero. You must take care in the case where a1b2 = b1a2 as this corresponds to a singular matrix and division by zero. Geometrically it means that OA and OB are parallel and hence there is no solution. The algebra above probably needs verifying for any slight typo mistake but the methodology is correct. Maybe long winded but you can get it all simply from point coordinates and saves you calculating inverse trig functions to get angles.
The above applies to angles < 180 degrees, so if the your angle is greater than 180 degrees, you should check instead for
!(K1 >= 0 && K2 >= 0)
as this is exterior to the segment less than 180 degree. Remember that for 0 and 180 degrees you will have a divide by zero error which must be checked for (ensure a1b2 - b1a2 != 0 )
Yes, I meant the smallest angle in my comment above. Look at this thread for an extensive discussion on cheap ways to find the measure of the angle between two vectors. I have used the lookup-table approach on many occasions with great success.
Triangle O B C has to be positive oriented and also triangle O C A. To calaculate orientation, just use Shoelace formula. Both values has to be positive.