Memory initialization and addressing in Neural Turing Machines - neural-network

It is not clear to me how the NTM initializes its memory to support content-based addressing with cosine distance. It's clear the initial memory cell values can not be zero but even if you initialize them to some non-zero value (say 1.0), all the cells will look the same and the memory addressing mechanism will produce a uniform distribution over the memory cells. The shift mechanism will just shift this uniform distribution, producing the same uniform distribution and the sharpening might sharpen some value, mainly due to noise.
So I don't see how the memory they described can be used beyond writing the same thing all over the place all the time (if you discount some noise).
Am I missing something and can somebody enlighten me about memory initialization?

I think I have it figured out. I looked over the interpolation and how it interacts with the shift.
The distribution over the memory cells is part of a memory head's state. The interpolation chooses between the previous distribution and the new distribution generated by the content-based cosine similarity. The shift then can decide to move the head one cell down or one cell up (or stay where it is).
Suppose the memory is empty and the initial addressing distribution is focused around the first memory cell. The content-based mechanism will produce a uniform distribution but the interpolation mechanism can decide to ignore that uniform distribution, taking the initial distribution (focusing on the first cell) and subsequently, the shift can decide to move to the next cell.
So bulk memory population (writing the whole input sequence into memory) would ignore the content-based part.

Related

Mutate weights and biases in a neural network through genetic algorithm

I have a genetic algorithm evolving a population of neural networks
Until now I make mutation on weights or biases using random.randn (Python) which is a random value from a normal distribution with mean = 0
It works "well" and I managed to achieve my project using it be wouldn't it be better to use a uniform distribution on a given interval ?
My intuition is that it would lead to more variety in my networks
I think, that this question has no simple solution. In case of normal distribution will be numbers around mean have more chances to be "selected" by your number generator, uniform distribution give almost equal chance to all numbers. That is clear but answer to question, will equal chance mean better result, lays according to me only at empirical experiments. So I suggest you to perform experiments with normal and uniform distribution a try to judge based on results.
About variety. I assume that you create some random vector which represents weights. At stage of mutation you perform addition of random number. This number will be more likely from close interval around mean, so in case 0 mutation with high probability will be change of some elements only little. So there will be only little improvements over vector and sometimes something big shows up. In case of uniform distribution will be changes more random, which leads to different individual. Question is, will be these individual better? I don't know, but I offer you another view. I look to genetic algorithms like an analogy to evolution theory. And from this point of view, cumulative little improvements of individual with little probability of some big change is more appropriate. Think about situation, used is uniform distribution, but children has low fitness due to big changes so at phase of creating new generation will be not selected. And you will wait so long for one tiny improvement which make your network works with good results.
Maybe one more thing. Your experiments maybe show that uniform/normal distribution is better. But such result may be true only for your current problem, no at general.

Dealing with a large kernel matrix in SVM

I have a matrix X, size 40-by-60000
while writing the SVM, I need to form a linear kernel: K = X'*X
And of course I would get an error
Requested 60000x60000 (26.8GB) array exceeds maximum array size preference.
How is it usually done? The data set is Mnist, so someone must have done this before. In this case rank(K) <= 40, I need a way to store K and later pass it to quadprog.
How is it usually done?
Usually kernel matrices for big datasets are not precomputed. Since optimisation methods used (like SMO or gradient descent) do only need access to a subset of samples in each iteration, you simply need a data structure which is a lazy kernel matrix, in other words - each time an optimiser requests K[i,j] you literally compute K(xi,xj) then. Often, there are also caching mechanisms to make sure that often requested kernel values are already prepared etc.
If you're willing to commit to a linear kernel (or any other kernel whose corresponding feature transformation is easily computed) you can avoid allocating O(N^2) memory by using a primal optimization method, which does not construct the full kernel matrix K.
Primal methods represent the model using a weighted sum of the training samples' features, and so will only take O(NxD) memory, where N and D are the number of training samples and their feature dimension.
You could also use liblinear (if you resolve the C++ issues).
Note this comment from their website: "Without using kernels, one can quickly train a much larger set via a linear classifier."
This problem occurs due to the large size of your data set, thus it exceeds the amount of RAM available in your system. In 64-bit systems data processing performs better than in 32-bit, so you'll want to check which of the two your system is.

Matlab - Warn me about a Large Matrix

I have been working with a few not particularly well designed or documented matlab scripts provided by my prof. I am required to modify them in order to make them do what is required by the specification, but it is not immediately clear how large certain matrices will become. I know if done properly the problems should not consume excessive resources, but more than a few times now I've had matlab hang trying to allocate huge blocks of memory. I suspect this is because I multiplied a matrix I did not fully understand in the wrong order, resulting in a larger matrix, rather than for example a scalar.
Is there a way to have matlab prompt me if it finds it will need to allocate a matrix over a certain number of bytes? That way while I'm debugging/maintaining code I'll have an opportunity to cancel, before it goes beserk.

How to find the "optimal" cut-off point (threshold)

I have a set of weighted features for machine learning. I'd like to reduce the feature set and just use those with a very large or very small weight.
So given below image of sorted weights, I'd only like to use the features that have weights above the higher or below the lower yellow line.
What I'm looking for is some kind of slope change detection so I can discard all the features until the first/last slope coefficient increase/decrease.
While I (think I) know how to code this myself (with first and second numerical derivatives), I'm interested in any established methods. Perhaps there's some statistic or index that computes something like that, or anything I can use from SciPy?
Edit:
At the moment, I'm using 1.8*positive.std() as positive and 1.8*negative.std() as negative threshold (fast and simple), but I'm not mathematician enough to determine how robust this is. I don't think it is, though. ⍨
If the data are (approximately) Gaussian distributed, then just using a multiple
of the standard deviation is sensible.
If you are worried about heavier tails, then you may want to base your analysis on order
statistics.
Since you've plotted it, I'll assume you're willing to sort all of the
data.
Let N be the number of data points in your sample.
Let x[i] be the i'th value in the sorted list of values.
Then 0.5( x[int( 0.8413*N)]-x[int(0.1587*N)]) is an estimate of the standard deviation
which is more robust against outliers. This estimate of the std can be used as you
indicated above. (The magic numbers above are the fraction of data that are
less than [mean+1sigma] and [mean-1sigma] respectively).
There are also conditions where just keeping the highest 10% and lowest 10% would be
sensible as well; and these cutoffs are easily computed if you have the sorted data
on hand.
These are somewhat ad hoc approaches based on the content of your question.
The general sense of what you're trying to do is (a form of) anomaly detection,
and you can probably do a better job of it if you're careful in defining/estimating
what the shape of the distribution is near the middle, so that you can tell when
the features are getting anomalous.

Lucas Kanade Optical Flow, Direction Vector

I am working on optical flow, and based on the lecture notes here and some samples on the Internet, I wrote this Python code.
All code and sample images are there as well. For small displacements of around 4-5 pixels, the direction of vector calculated seems to be fine, but the magnitude of the vector is too small (that's why I had to multiply u,v by 3 before plotting them).
Is this because of the limitation of the algorithm, or error in the code? The lecture note shared above also says that motion needs to be small "u, v are less than 1 pixel", maybe that's why. What is the reason for this limitation?
#belisarius says "LK uses a first order approximation, and so (u,v) should be ideally << 1, if not, higher order terms dominate the behavior and you are toast. ".
A standard conclusion from the optical flow constraint equation (OFCE, slide 5 of your reference), is that "your motion should be less than a pixel, less higher order terms kill you". While technically true, you can overcome this in practice using larger averaging windows. This requires that you do sane statistics, i.e. not pure least square means, as suggested in the slides. Equally fast computations, and far superior results can be achieved by Tikhonov regularization. This necessitates setting a tuning value(the Tikhonov constant). This can be done as a global constant, or letting it be adjusted to local information in the image (such as the Shi-Tomasi confidence, aka structure tensor determinant).
Note that this does not replace the need for multi-scale approaches in order to deal with larger motions. It may extend the range a bit for what any single scale can deal with.
Implementations, visualizations and code is available in tutorial format here, albeit in Matlab not Python.