The following images are often represented to describe the word2vec model with skip-gram:
However, after reading this discussion on stackoverflow, it seems that word2vec actually take 1 word and input and 1 word as output. The output word is randomly samples from the window. (And this is performed X number of times to generate X input/output pairs.)
It seems to me then that the above image is not correctly describing the network. My question is: is the 1 input/1 output standard (the Tensorflow word2vec tutorial takes this approach and calls it skip-gram) or do some networks actually take the structure of the above image?
It's not a great diagram.
In CBOW, those converging arrows are an averaging that happens all-at-once, to create one single 'training example' (desired prediction) that is (average(context1, context2, ..., contextN) -> target-word). (In practice averaging is more common than the 'SUM' shown in the diagram.)
In Skip-Gram, those diverging arrows are multiple training examples (desired predictions) made one-after-the-other.
And in both diagrams, while they look a bit like neural-net node-architectures, the actual hidden-layer and internal-connection weights are just implied inside the middle-column-to-right-column arrows.
Skip-gram is always 1 "input" context word used to predict 1 nearby (within the effective 'window') "output" target word.
Implementations tend to iterate through the whole effective window, so every (context -> target) pair gets used as a training-example. And in practice, it doesn't matter if you consider the central word the target-word and each word around it to be context-words, or the central word the context-word and each word around it to be target-words – both methods result in the exact same set of (word -> word) pairs being trained, just in a slightly different iteration order. (I believe the original Word2Vec paper described it one way, but then Google's released code did it the other way for reasons of slightly-better cache efficiency.)
In fact the effective window, for each central word considered, is chosen to be some random number from 1 to the configured maximum window value. This turns out to be a cheap way of essentially weighting nearer-words more: the immediate neighbors are always part of training-pairs, further words only sometimes. That is, pairs are not randomly sampled from the whole window - it's just a random window size. (There's another down-sampling where the most-frequent words will be randomly dropped so as not to overtrain them at the expense of less-frequent words, but that's a totally separate process not reflected in the above.)
In CBOW, instead of up-to 2*window input-output pairs of the (context-word -> target-word) form, there's a single input-output pair of (context-words-average -> target-word). (In CBOW, a loop creates the average value for a single N:1 training-example for one central word, and then splits the backpropagated error across all contributing words. In skip-gram, a loop creates multiple alternate 1:1 training-examples for one central word.)
Related
I have a question about the word representation algorithms:
Which one of the algorithms word2Vec, doc2Vec and Tf-IDF is more suitable for handling text classification tasks ?
The corpus used in my supervised learning classification is composed of a list of multiple sentences, with both short length sentences and long length ones. As discussed in this thread, doc2vec vs word2vec choice is a matter of document length. As for Tf-Idf vs. word embedding, it's more a matter of text representation.
My other question is, what if for the same corpus I had more than one label to link to the sentences in it ? If I create multiple entries/labels for the same sentence, it affects the decision of the final classification algorithm. How can I tell the model that every label counts equal for every sentence of the document ?
Thank you in advance,
You should try multiple methods of turning your sentences into 'feature vectors'. There are no hard-and-fast rules; what works best for your project will depend a lot on your specific data, problem-domains, & classification goals.
(Don't extrapolate guidelines from other answers – such as the one you've linked that's about document-similarity rather than classification – as best practices for your project.)
To get initially underway, you may want to focus on some simple 'binary classification' aspect of your data, first. For example, pick a single label. Train on all the texts, merely trying to predict if that one label applies or not.
When you have that working, so you have a understanding of each step – corpus prep, text processing, feature-vectorization, classification-training, classification-evaluation – then you can try extending/adapting those steps to either single-label classification (where each text should have exactly one unique label) or multi-label classification (where each text might have any number of combined labels).
I am working user behavior project. Based on user interaction I have got some data. There is nice sequence which smoothly increases and decreases over the time. But there are little discrepancies, which are very bad. Please refer to graph below:
You can also find data here:
2.0789 2.09604 2.11472 2.13414 2.15609 2.17776 2.2021 2.22722 2.25019 2.27304 2.29724 2.31991 2.34285 2.36569 2.38682 2.40634 2.42068 2.43947 2.45099 2.46564 2.48385 2.49747 2.49031 2.51458 2.5149 2.52632 2.54689 2.56077 2.57821 2.57877 2.59104 2.57625 2.55987 2.5694 2.56244 2.56599 2.54696 2.52479 2.50345 2.48306 2.50934 2.4512 2.43586 2.40664 2.38721 2.3816 2.36415 2.33408 2.31225 2.28801 2.26583 2.24054 2.2135 2.19678 2.16366 2.13945 2.11102 2.08389 2.05533 2.02899 2.00373 1.9752 1.94862 1.91982 1.89125 1.86307 1.83539 1.80641 1.77946 1.75333 1.72765 1.70417 1.68106 1.65971 1.64032 1.62386 1.6034 1.5829 1.56022 1.54167 1.53141 1.52329 1.51128 1.52125 1.51127 1.50753 1.51494 1.51777 1.55563 1.56948 1.57866 1.60095 1.61939 1.64399 1.67643 1.70784 1.74259 1.7815 1.81939 1.84942 1.87731
1.89895 1.91676 1.92987
I would want to smooth out this sequence. The technique should be able to eliminate numbers with characteristic of X and Y, i.e. error in mono-increasing or mono-decreasing.
If not eliminate, technique should be able to shift them so that series is not affected by errors.
What I have tried and failed:
I tried to test difference between values. In some special cases it works, but for sequence as presented in this the distance between numbers is not such that I can cut out errors
I tried applying a counter, which is some X, then only change is accepted otherwise point is mapped to previous point only. Here I have great trouble deciding on value of X, because this is based on user-interaction, I am not really controller of it. If user interaction is such that its plot would be a zigzag pattern, I am ending up with 'no user movement data detected at all' situation.
Please share the techniques that you are aware of.
PS: Data made available in this example is a particular case. There is no typical pattern in which numbers are going to occure, but we expect some range to be continuous with all the examples. Solution I am seeking is generic.
I do not know how much effort you want to involve in this problem but if you want theoretical guaranties,
topological persistence seems well adapted to your problem imho.
Basically with that method, you can filtrate local maximum/minimum by fixing a scale
and there are theoritical proofs that says that if you sampling is
close from your function, then you extracts correct number of maximums with persistence.
You can see these slides (mainly pages 7-9 to get the idea) to get an idea of the method.
Basically, if you take your points as a landscape and imagine a watershed starting from maximum height and decreasing, you have some picks.
Every pick has a time where it is born which is the time where it becomes emerged and a time where it dies which is when it merges with an higher pick. Now a persistence diagram pictures a point for every pick where its x/y coordinates are its time of birth/death (by assumption the first pick does not die and is not shown).
If a pick is a global maximal, then it will be further from the diagonal in the persistence diagram than a local maximum pick. To remove local maximums you have to remove picks close to the diagonal. There are fours local maximums in your example as you can see with the persistence diagram of your data (thanks for providing the data btw) and two global ones (the first pick is not pictured in a persistence diagram):
If you noise your data like that :
You will still get a very decent persistence diagram that will allow you to filter local maximum as you want :
Please ask if you want more details or references.
Since you can not decide on a cut off frequency, and not even on the filter you want to use, I would implement several, and let the user set the parameters.
The first thing that I thought of is running average, and you can see that there are so many things to set, to get different outputs.
I’m not specialist in signal processing. I’m doing simple processing on 1D signal using c++. I want really to know how I can determine the part that have the highest zero cross rate (highest frequency!). Is there a simple way or method to tell the beginning and the end of this part.
This image illustrate the form of my signal, and this image is what I need to do (two indexes of beginning and end)
Edited:
Actually I have no prior idea about the width of the beginning and the end, it's so variable.
I could calculate the number of zero crossing, but I have no idea how to define it's range
double calculateZC(vector<double> signals){
int ZC_counter=0;
int size=signals.size();
for (int i=0; i<size-1; i++){
if((signals[i]>=0 && signals[i+1]<0) || (signals[i]<0 && signals[i+1]>=0)){
ZC_counter++;
}
}
return ZC_counter;
}
Here is a fairly simple strategy which might give you some point to start. The outline of the algorithm is as follows
Input: Vector of your data points {y0,y1,...}
Parameters:
Window size sigma.
A threshold 0<p<1 defining when to start looking for a region.
Output: The start- and endpoint {t0,t1} of the region with the most zero-crossings
I won't give any C++ code, but the method should be easy to implement. As example let us use the following function
What we desire is the region between about 480 and 600 where the zero density higher than in the front. First step in the algorithm is to calculate the positions of zeros. You can do this by what you already have but instead of counting, you store the values for i where you met a zero.
This will give you a list of zero positions
From this list (you can do this directly in the above for-loop!) you create a list having the same size as your input data which looks like {0,0,0,...,1,0,..,1,0,..}. Every zero-crossing position in your input data is marked with a 1.
The next step is to smooth this list with a smoothing filter of size sigma. Here, you can use what you like; in the simplest case a moving average or a Gaussian filter. The higher you choose sigma the bigger becomes your look around window which measures how many zero-crossings are around a certain point. Let me give the output of this filter together with the original zero positions. Note that I used a Gaussian filter of size 10 here
In a next step, you go through the filtered data find the maximum value. In this case it is about 0.15. Now you choose your second parameter which is some percentage of this maximum. Lets say p=0.6.
The final step is to go through the filtered data and when the value is greater than p you start to remember a new region. As soon as the value drops below p, you end this region and remember start and endpoint. Once you are finished walking through the data, you are left with a list of regions, each defined by a start and an endpoint. Now you choose the region with the biggest extend and you are done.
(Optionally, you could add the filter size to each end of the final region)
For the above example, I get 11 regions as follows
{{164,173},{196,205},{220,230},{241,252},{259,271},{278,290},
{297,309},{318,327},{341,350},{458,468},{476,590}}
where the one with the biggest extend is the last one {476,590}. The final result looks (with 1/2 filter region padding)
Conclusion
Please don't be discouraged by the length of my answer. I tried to explain everything in detail. The implementation is really just some loops:
one loop to create the zero-crossings list {0,0,..,1,0,...}
one nested loop for the moving average filter (or you use some library Gaussian filter). Here you can at the same time extract the maximum value
one loop to extract all regions
one loop to extract the largest region if you haven't already extracted it in the above step
I'm very new in image processing and my first assignment is to make a working program which can recognize faces and their names.
Until now, I successfully make a project to detect, crop the detected image, make it to sobel and translate it to array of float.
But, I'm very confused how to implement the Backpropagation MLP to learn the image so it can recognize the correct name for the detected face.
It's a great honor for all experts in stackoverflow to give me some examples how to implement the Image array to be learned with the backpropagation.
It is standard machine learning algorithm. You have a number of arrays of floats (instances in ML or observations in statistics terms) and corresponding names (labels, class tags), one per array. This is enough for use in most ML algorithms. Specifically in ANN, elements of your array (i.e. features) are inputs of the network and labels (names) are its outputs.
If you are looking for theoretical description of backpropagation, take a look at Stanford's ml-class lectures (ANN section). If you need ready implementation, read this question.
You haven't specified what are elements of your arrays. If you use just pixels of original image, this should work, but not very well. If you need production level system (though still with the use of ANN), try to extract more high level features (e.g. Haar-like features, that OpenCV uses itself).
Have you tried writing your feature vectors to an arff file and to feed them to weka, just to see if your approach might work at all?
Weka has a lot of classifiers integrated, including MLP.
As I understood so far, I suspect the features and the classifier you have chosen not to work.
To your original question: Have you made any attempts to implement a neural network on your own? If so, where you got stuck? Note, that this is not the place to request a complete working implementation from the audience.
To provide a general answer on a general question:
Usually you have nodes in an MLP. Specifically input nodes, output nodes, and hidden nodes. These nodes are strictly organized in layers. The input layer at the bottom, the output layer on the top, hidden layers in between. The nodes are connected in a simple feed-forward fashion (output connections are allowed to the next higher layer only).
Then you go and connect each of your float to a single input node and feed the feature vectors to your network. For your backpropagation you need to supply an error signal that you specify for the output nodes. So if you have n names to distinguish, you may use n output nodes (i.e. one for each name). Make them for example return 1 in case of a match and 0 else. You could very well use one output node and let it return n different values for the names. Probably it would even be best to use n completely different perceptrons, i.e. one for each name, to avoid some side-effects (catastrophic interference).
Note, that the output of each node is a number, not a name. Therefore you need to use some sort of thresholds, to get a number-name relation.
Also note, that you need a lot of training data to train a large network (i.e. to obey the curse of dimensionality). It would be interesting to know the size of your float array.
Indeed, for a complex decision you may need a larger number of hidden nodes or even hidden layers.
Further note, that you may need to do a lot of evaluation (i.e. cross validation) to find the optimal configuration (number of layers, number of nodes per layer), or to find even any working configuration.
Good luck, any way!
This is to be done in C++ or C....
I know we can read the MP3s' meta data, but that information can be changed by anyone, can't it?
So is there a way to analyze a file's contents and compare it against another file and determine if it is in fact the same song?
edit
Lots of interesting things coming out that I hadn't thought of. Not at all a good idea to attempt this.
It's possible, but very hard.
Even the same original recording may well be encoded differently by different MP3 encoders or the same encoder with different settings... leading to different results when the MP3 is then decoded. You'd need to work out an aural model to "understand" how big the differences are, and make a judgement.
Then there's the matter of different recordings. If I sing "Once in Royal David's City" and Aled Jones sings it, are those the same song? What if there are two different versions of a song where one has slightly modified lyrics? The key could be different, it could be in a different vocal range - all kinds of things.
How different can two songs be but still count as "the same song"? Once you've decided that, then there's the small matter of implementing it ;)
If I really had to do this, my first attempt would be to take a Fourier transform of both songs and compare the histograms. You can use FFTW (http://www.fftw.org/) to take the Fourier transform, and then compare the histograms by summing the squares of the differences at each frequency. If the resultant sum is greater than some threshold (which you must determine by experimentation) then the songs are deemed to be different, otherwise they are the same.
No. Not SO simple.
You can check they contain the same encoded data, BUT:
Could be a different bitrate
Could be the same song, just a 1/100ths of a second off
In both cases the bytes would not match.
Basically, if a solution looks too simple to be true, it often is.
If you mean "same song" in the iTunes sense of "same recording", it would be possible to compares two audio files, but not by byte-by-byte comparison of an encoded file since even for the same format there are variables such as data rate and compression that are selected at time of encoding.
Also each encoding of the same recording may include different lead-in/lead-out timings, different amplitude and equalisation, and may have come from differing original sources (vinyl, CD, original master etc.). So you need a comparison method that takes all these variables into account, and even then you will end up with a 'likelihood' of a match rather than a definitive match.
If you genuinely mean "same song", i.e. any recording by any artist of the same composition and lyrics, then you are unlikely to get a high statistical correlation in most cases since pitch, tempo, range, instrumental arrangement will be very different.
In the "same recording" scenario, relatively simple signal processing and statistical techniques could be applied, in the "same song" scenario, AI techniques would need to be deployed, and even then the results I suspect would be poor.
If you want to compare MP3 files that originated from the same MP3, but have tagged with metadata differently, it would be straight forward to just compare the actual audio data. Since it originated from the same MP3 encoding, you should be able to do a byte by byte comparison. You would have to compare all byte. It should be sufficient to sample just a few to get a unique key that would be statistically almost impossible to find in another song.
If the files have been produced by different encoders, you would have to extract some "fuzzy" feature keys from the data and compare those keys. In a hurry I would probably construct an algorithm like this:
Decode audio to pulse-code modulation (wave) in a standard bit rate.
Find a fixed number of feature starting points using some dynamic location algorithm. For example find top 10 highest wave peaks ordered from beginning of wave or simply spread evenly across the wave (it would be a good idea to fix the first and last position dynamically though, since different encodings might not start and end at exactly the same point). An improvement would be to select feature points at positions in the wave that are not likely to be too repetitive.
Extract a set of one-dimensional feature key scalars from the feature points. For example, for each feature normalize the following n-sample values and count the number of zero-crossings, peak to average ratio, mean zero-crossing distance, signal-energy. The goal is to extract robust features that are relatively unique, while still characteristic even if some noise and distortion is added to the signal. This can obviously be improved almost infinitely.
Compare the extracted feature keys of the two files using some accuracy measurement (f.eks. 9 out of 10 feature extractions must match at least 99% on 4 out of 5 of their extracted feature keys).
The benefit of a feature extraction approach is that you can build a database of features for all your mp3-files and for a single file ask the question: What other media files have exactly or almost exactly the same feature as this one. The feature lookup could be implemented very efficiently with R*-trees or similar, which could be used to give you a fast distance measurement between the n-dimensional feature sets.
The above technique is essentially a variant of what is used in image search algorithms such as SIFT, which is probably the base of such application as Photosynth and Google Goggles. In image searching you filter the image for good candidate points for relatively unique features (such as corners of shapes), then you normalize the area around that feature to get normalized color, intensity, scale and direction of features. Finally you extract the features and search an n-dimensional database of features of other images and verify that found features in other images are geometrically positioned in the same pattern as in your search image. The technique for searching audio would be the same, only simpler, since audio is one dimensional.
Use the open source EchoPrint library to create a signature of the two audio files, and compare them with each other.
The library is very easy to use, and has clear examples on how to create the signatures.
http://echoprint.me/
You can even query their database with the signature and find matching song metadata (such as title, artist, etc).
I think the Fast Fourier-Transform (FFT) approach hinted by jstanley is pretty good for most use cases; in particular, it works for verifying that the two are the same release/ same recording by the same artist/ same bitrate / audio quality.
To be more explicit, sox and spek (via command line and GUI, respectively) can do this pretty painlessly.
Spek is pretty foolproof -- just open the software and point it to the two audio files in question.
sox can generate spectograms (FFTs) from the command line line so:
sox "$file" -n spectrogram -o "$outfile".
The result from either are two images; if they look basically identical, then for almost all intents and purposes, the two songs will be equivalent.
For example, I wanted to test if these two files:
Soundtrack to an imaginary film mixtape 2011.mp3
DJRUM - Sountrack to an imaginary film mixtape 2011 (for mary-anne hobbs).mp3
were the same. diff reported a difference in the binary files (perhaps due to metadata differences or minor encoding differences), but a quick glance at their spectrograms resolved it: