Hi!

I think what you need is ICA (Independent component analysis ). Look at Bishop's book (Bishop C. Pattern Recognition and Machine Learning, Springer, 2006, chapter 12 (more specific 12.4.1) about this topic.

Best regards,

Hello!

Have you already tried KNN classification algorithm?

See you...

Hi!

I think what you need is ICA (Independent component analysis ). Look at Bishop's book (Bishop C. Pattern Recognition and Machine Learning, Springer, 2006, chapter 12 (more specific 12.4.1) about this topic.

Best regards,

I am a fan of bayesian networks because of their flexibility and transparency. With ANN however I would be very careful with overfitting. This generally happens if you test on a limited set of data. Sometimes best is not the most accurate. Obviously it's difficult to generalise solutions unless we know more about the problem.

Competitive ANN, like Self Organizing Maps, also can manage partial input classification.

I recomend you to use somtoolbox for Matlab (http://www.cis.hut.fi/somtoolbox/about.shtml). With a simple definiton of a mask with 0-1 values for the input data process (the input data present are asigned 1 in the mask, and the input data missed are asigned with 0 in the mask), the SOM model can manage any combination of inputs.

Decission trees like Random Forest or C4.5 Algorithms, Lazy Classifiers like KNN nearest neighbor, have good performance. If you want to use selection features to reduce and remove the redundant features you can use it, like bassed on filter are faster and have good results.

I don't really think you have a classification problem. If I did understand you already have the features of the individual patterns. And the problem is to find the patterns when you have mixtures of the features. The primary question here is how is the mixing procedure. Does it respond to a linear combination or is it non-linear? does patterns appear in the mixture in a crisp way? ...

Actually this depends on type and nature of the data, and number of class labels. Usually you can not say in general this algorithm is better than others because this also depends on the type of application domain. The simplest methods are ANN and KNN. However, if you are concerned with classification, it means you have to use supervised approach. You can use some tools like Matlab or rapidminer (i recommend Matlab) . You may also try to use other methods such as SVM. However, hybrid systems in general proven to be good approach, which means you can use for example ANN with GA. As you mentioned in your problem, if your dataset size is comperhensive and accurate, the ANN will be mostly the best. KNN can perform better if the dataset is less accurate and less comperhensive.

Give us description n samples about your case,,, (picture, movie,etc)

What you need is a neural net with a feedback component : once the most salient pattern is recognized, the activations are fed back to suppress the pattern in the original image, leaving the next salient pattern to be recognized ... I think Grossberg's Adaptive Resonance Theory can do the trick, but I didn't experiment it myself ... you may look into it

Thank you to everybody for the comments and below is a wider description of the problem. I would appreciate if you can provide a literature example on which the algorithm you propos has been actually used in a mixed-label problem, or of course, explain why you think it would be suitable for this problem. This is mainly because I can’t try them all (time constraint) and I would like to focus on algorithms that have an indication to be good in this specific problem, “mixed-lables / simultaneous patec”.

@ Leonardo. No, I haven’t but I can’t think on why it would do a better job than the MLP for mixed-labels. Any hint?

@ Costin, I believe ICA has been used but seemed like MLP had better performance in the individual patterns. I don’t think it has been used for mixed-label so I’ll take a look myself anyway. Thanks!

@ Josh, Julio and Hind any hint on why your algorithms would be better than the MLP for mixed-labels? Overfitting doesn’t seem to be an issue so far.

@ Miguel, the way in which the signals mix is quite tricky, and as I mentioned, probably non-linear, please see below.

@ Nabil and Muhammad, please see below.

Even more specific:

I’m doing patrec of myoelectric signals for the control of prosthetic devices. The typical scenario is placing some bipolar electrodes in the remaining muscles of the stump, each of these electrodes give information of the overall myoelectric activity of its local area. I can easily predict over 10 different hand and wrist movements using this technique, e.g. hand open/close, wrist extension/flexion, etc. Now, what if I need to predict when the user wants to close the hand and flex the wrist simultaneously (mixed-labels). This simultaneous movement generates new myoelectric signal that are obviously composed with those of individual movements in some way (very tricky way). I normally ask the patient to execute only the individual movements, asking him to do all possible combinations is unpractical, so I’m want to avoid this.

On the signals: Myoelectric signal are non-stationary, stochastic with close zero mean. You normally don’t feed directly the signal to the classifier but you do a snapshot or time window, from which you can extract different features such as the RMS value, standard deviation, and so on. These are the features you feed the classifier with.

If anyone would like to try some data, let me know ;)

Would this be a multi-task learing problem? I'm not sure I understand your description, for instance, what do you mean by "signal"? A more formal specification and/or a more complte example would help.

I dont think you will we able to handle this without either (a) having a theoretical model of how the signals are mixed (linear combination or whatever function) or (b) having data of the individual and the mixed signals, such that a learning algorithm can extract the relations.

If a multi-task learning problem is the same as a multi-class problem, yes BUT more complicated. Multi taks/label/classes prediction is probable the most common patrec problem, and all the algorithms that have been suggested (and more) are good for that, BUT they normally only predict one class at the time. In the problem I described there are some task/labels/classes which are the limb movements. I want to predict when these movements are done simultaneously by only training in the individual movements, so simultaneous prediction if you may.

By signal, I mean the myoelectric signal. As I mentioned before, you don't normally use the raw signal, you extract features from time windows (snapshots).

If i understand your problem, you are looking for something that does http://en.wikipedia.org/wiki/Blind_signal_separation.

The problem is not about selecting proper algorithm. The problem is in the formulation. Namely, you want to have the answer to question "Is activity of type 1 present in my mix signal? Y/N", then second question "Is activity of type 2 present in my mix signal?" and so forth. The complication of formulation is due to the fact, that mix samples WITH presence of type 1 can be collected. But opposite samples constitute "the open world" (anything without the activity of type 1).

One potential approach is to train so called "one-class classifier", which actually estimates the support (the region) of signals, containing type 1 activity. These signals within the support will vote for "Y", all the rest is "N". One-class classifiers are implemented e.g. in LibSVM package.

You may also try regular classifiers, but use actual mixtures as training samples (e.g. random combinations with and without activity of type 1 - if you cam mix them artificially, or use real records, with labels "1 present" vs "1 is not present"). So, build as many binary classifiers, as the total number of activities to be detected (e.g. 6 in your original posting).

i think a fuzzy classification rules with GMM or any classifier that gives probability for the presence of each observation (or signal) will do and you can adjust the rules to select outputs with probs more than a certain value. just an idea, i'm not that expert.

You could try to do a regression, fitting a weighted sum of the known patterns to the mixed signal, with the weights as the free variables. Ideally this would result in a vector w where, for each pattern i, w[i] is 1.0 if i is present in the signal, and 0.0 otherwise.

Perhaps this example is like the one you have: you take images of your head in several fixed orientations: frontal, lateral, upper and low, and label them...but we do not label head in intermediate positions, or even we never take photos of intermediate head positions. I think it is similar, because you train with "simple" movements, but all their possible combinations are not generated (or not labeled), because you want the system to "interpolate" these complex signals...

Again the SOM model makes this kind of processing possible. As the SOM neural map is trained with the "simple" movements (or head positions), the neurons "recognizing" labeled data are pulling their neighbour neurons, so after the training process, these interneurons are interpolating the data space between clusters of data labeled (simple movements).

But it can be better, as the SOM model is unsupervised, you can generate multiple data "free" of labeling in order to generate additional data, and select some data (simple movements or head positions) to be labeled. After training a SOM, the set of weigths of the map is a quantified representation (codebook) of the data distributions. With the SOM you generate a codebook like a look-up table to identify the simple movements and extra codes to "identify" never-labeled movements.

I explained the head-photo example because one of my students solved it with a SOM. The map recognized intermediate head positions with interpolating neurons, even when the fixed head positions were the only data during training process.

@ Paul, A model would be very useful but not necessary. After all, most of the patrec algorithms don’t need information about the system. Thanks for the link, I'll check on that.

@ Serge, It seems to me that what you described are binary classifiers and then major voting. I'm already trying that and as I said, a single classifier (MLP) is doing better so far. The binary classifiers I'm trying are One-Vs-One and One-Vs-All. With both MLP and LDA.

@ Ahmed, Thanks and you are right. MLP and others do provide you with some sort of probability of presence. That’s one of the reasons MLP is doing well, however, I’m sure someone has created a better algorithm for this specific problem.

@ Helio, the issue is that there is no "known-patterns". I can't have stored patterns, It is pattern/label/class recognition problem. I have to predict the pattern/label/class every time, either individually or mixed, depending on what the user is doing.

@ Julio, In a way MLP does the same from the point of view that it can make predictions out of the available knowledge from the individual movements. Currently the MLP gives me over 95% accuracy for the individual movements and around 70% when 2 out of 6 movements are combined. I'm sure it can be better, that's why I posted the question. I haven’t tried an unsupervised approached so I guess it worths testing. Thanks.

Hi Max,

You mentioned you had some data you could share? I was wondering if I could take a look. It sounds interesting and I'd like tos ee the data to get a better understanding of the problem. I'm wondering if something like Minimum Message Length classification could be of use. It's lengthy to implement, and I think more related to the LDA side than the MLP, but I think it might have an edge over LDA. That could just be my bias against NNs and implicit models coming out, though.

Cheers.

Just use the strength of the output from each predicted movement to control each movement. If the estimated strength of hand closure is .3 and the estimated strength of wrist turn is .6 then simply apply those strengths (at the same time) to your hand closure and wrist turn motors.

There is no need to quantize the result to one or the other of the actions. Do both or many.

With a Hidden Markov Model each action is trained separately. Each action has its own mathematical model whose transitions probabilities are adjusted by training. When given unknown input all of the models compete against each other and produce a probability that the input was generated by that model. Instead of choosing the 'best' model choose all of them each with its own strength (probability). In that way you get a 'mixed' model representation.

Hi Max,

Your input signal is a convolution of many (8) signals, is not going to be easy to raise a flag when "individual signal 3" is present in the input. Just as an example, adding two signals can cancel them both (like noise reduction signal generation).

Looks like your problem is reduced by the "Cocktail Party Problem", i.e. de-multiplexing a signal which contains the sum of several ones. Since the addition of one signal to another makes the process irreversible, techniques such as (previously suggested) ICA would be your best shot. I suggest to use several (>8) sensors on the patient, to measure the signal >8 times, then trying to obtain its individual components.

If you used Matlab for simulating MLP, the use of SOM with the same data is very simple. SOM is very quick in training process and has another interesting property: you can define a maximun distance between weights of winner neuron and input data (a threshold of Quantization error, Tqerr) to decide if the data received belongs to the data distribution modelled. Lets suposse that the signals are corrupted by, for example, electrostatic discharges from an object touched by the user of the artificial limb. This strange variation of the signal would be classified as a movement and the user would see how the limb moves without his/her will. However if the data vector gives a value of quantization error over Tqerr, it can be considered as an outlier of the data distribution and will not be processed by the mechanical part, so avoiding undesired movements due to strange signals. This property also has been used for novelty detecction in many works.

@ Max: if you dont have/make a model and just take an algorithm and hope it solves your problem you implicitly take the model the algorithm was built from. e.g. if you try to solve it with non-negative matrix factorization you implicitly assume the components are added linearly, ...

The problem is that you don't have all the combinations of patterns in your learning database, so the classifier must be able to generalize about the way patterns are nonlinearly combining, and we can assume there are some invariants across all different combinations ... I suggest trying a partially shared-weight neural network, with weights from the hidden layer to the output layer identical for all output nodes with a translation in respect of hidden neurons (ex: if you have 6 output and 60 hidden heurons, weight to output node 1 from hidden neuron n equals weight to output node 2 from hidden neuron n+10, to 3 from n+20, etc) hence the ANN could learn the general nonlinearities of combining several patterns, independently of the patterns themselves.

@Max Reply to @ Serge, It seems to me that what you described are binary classifiers and then major voting. I'm already trying that and as I said, a single classifier (MLP) is doing better so far. The binary classifiers I'm trying are One-Vs-One and One-Vs-All. With both MLP and LDA.

I mean binary classifiers trained on sample mixtures, but not on pure signals. The difference is big :) . As another option - train 1-class classifiers to estimate supports of mixtures with presense of each pure signal in the mix. In all cases, trainnig from PURE signals is not exactly relevent to your needs.

Douglas, I don't believe he can calculate those strengths easily, given that he mentioned the signals are combined non-linearly.

Marx, without knowing the underlying non-linear function, it would be very hard to determine what the underlying signals are. This is simply not a problem that can be accurately solved by blindly applying signal processing algorithms. I believe you first need to look into how exactly the signals are being combined. If they are neural signals, chances are good that the signals are being combined with a sigmoid-type function. This would explain why MLP does so well.

@Alireza

Well it will be necessary to construct an adaptive system that can return a vector of strengths, one strength for each controller desired.

Since I work with speech recognition I have used Hidden Markov Models (HMM) which are trained via the 'data', the signals recorded from the probes on the human.

I also use Sparse Distributed Memories (SDM) which allow one to associate input vectors with output vectors. Given an input the memory can generate a vector of 'sums' (strengths) which are usually quantized to a single bit for each component but in this case that would not be necessary and the strength would be applied directly to control the motors (with a possible non-linear scaling).

The SDM and the HMM work on arbitrary on-linear relationships between input and output.

Hence Max Ortiz Catalan needs to use a system (such as SDM or HMM) that will give him such control.

@ Douglas:

> Well it will be necessary to construct an adaptive system that can return a vector of strengths, one strength for each controller desired.

Yes, that is the output he wants to obtain, the question here is how to obtain such an output.

Systems like sparse distributed memory could be used, however I'm not sure if such systems are well-suited to this problem. I have worked with neural signals and they have a tendency to add up in strange ways. For sake of example consider two input neurons. The signal for one action (like, say, 'open hand') could be (0.5, 0.8), and the signal for another could be (0.9,0.2). If the two signals occurred simultaneously, the result wouldn't look like (1.4,1.0), rather it would look more like (0.95, 0.9), for example. SDM's are poorly suited to this kind of superposition. On the other hand, MLPs were designed for this kind of superposition, which is why they work so well for this problem.

As for HMM's, they might be useful in order to infer the kind of non-linearity present in the data. However, HMM's require one to know the relationship between the hidden states and the observed states. For speech, this is known (the sound a word makes). However, in this case, this is precisely what is unknown (the relationship between the desired action and the neural signals, when multiple actions are present). Even if they were known, the number of observed states is exponential and so it seems HMMs would be of little utility here.

This is an area in which a lot of research has been done. Most approaches generally try to get at least a rough picture of how the inputs from the neurons are being superposed. Even a small amount of information on that can greatly help. Again, I think it's best to focus on that aspect of the problem first.

@Alireza,

"The signal for one action (like, say, 'open hand') could be (0.5, 0.8), and the signal for another could be (0.9,0.2). "

If you mean for the input x=(0.5,0.8) and another x'=(0.9,0.2) AND they are to be associated with the same output y then there is absolutely not problem using an SDM to produce the desired output y. If, however, you mean that x is associated with y and x' is associated with y' then again there is no problem since x will activate its own internal local hard locations and x' will activate a different set of internal hard locations to give x->y and x'->y'.

Now what one really should do is have the input be x associated with (y,0) and x' associated with (0,y') for the case that x is supposed to activate device d and x' is supposed to activate device d' such that when a new signal is input say, x", then one will get a mixture over the output (y, y') set of values which can be used to drive device d and d' (hand close, wrist turn).

Your comment about HMMs is not quite correct. One does not know in general the relationship between the hidden states and the observed states. It is up to the Viterbi algorithm to deduce that relationship using the forward backward training.

Look, there really is very little difference between MLP, SDM, HMM. It all simply boils down to the efficiency of training. The SDM does not use backpropogation. HMM does not use backpropogation. So the training of those two systems is more efficient than MLP. One simply needs to select the correct way of representing the data for the problem at hand.

For this problem we have as input the nerve signals from the limb. For output we have a vector of values for the actions desired (hand, wrist, ...). One trains the system by saying "Think of closing your hand" and then using the signals x as measured one associates that with (1,0) or some other strength value. One then says "Think of rotating your wrist" and then associates the measured x values with (0, 1). That's for training.

For control (recognition) one then lets the person think of whatever action they desire and the measured x is the input to you trained system and the output (y, y') is the used to drive the devices (d, d').

That's it.

@ Douglas,

Your comment reflects a misunderstanding of the Viterbi algorithm. The forward-backward algorithm can only find the emission probabilities if the precise transition probabilities are known, which is obviously impossible to determine in this case (if it even made theoretical sense).

Unfortunately, there is a great deal of difference between SDMs and MLPs.

> Now what one really should do is have the input be x associated with (y,0) and x' associated with (0,y') for the case that x is supposed to activate device d and x' is supposed to activate device d' such that when a new signal is input say, x", then one will get a mixture over the output (y, y') set of values which can be used to drive device d and d' (hand close, wrist turn).

Yes, that is what needs to be done. However, you have not shown any way of doing this, and that is what Marx asked for.

I could suggest the „Potential Function Method”, the so called kernel trick developed by Aizerman and Bravernan. The method is very efficacy.

Bi-directional associativ memory is a good algorithm for pattern recognition

@Alireza Nejati

The full HMM algorithm finds both the state to state transition probabilities, alpha, and the emission probabilities, beta.

I did show you how to do this.

(1) x = (x1, x2, ...,xm) input neurological signals

(2) y = (y1, y2, ..., yn) output (clamped) desired driving signals for device motors.

Examples of clamped y vectors are

y = (1, 0, ..., 0) for first motor

y = (0, 1, 0, .., 0) for second motor

y = (0, 0, ..., 0, 1) for last motor

The motor signals form a basis set for your output space.

Training: M.Write(x, y) where many (x,y) pairs are presented to M.

Recognition: M.Read(x, y) for a given x one obtains an output y vector.

where M is a model (either an SDM, or MLP, for some other adaptive model).

The read y vector will be a mixture (a point, a vector in the output space) what can be used to drive

simultaneously all of the device motors.

@Tim, and to who is interested. I made available some data sets from Matlab. There are three data sets: training, validation and testing. The information is not normalized. For each set there are two variables, e.g. for training there are trSets and trOuts. The trSets is a mxn matrix were m are the set index and n are the actual features forming the m set. For each m set there is a corresponding output in the trOuts. So for the set trSets(5,:), the corresponding output is in tOut(5,:). You will notice that the xOuts variables are mxn matrices where n = 7. This is because 6 movements were recorded plus resting, so there are 7 possible outputs for each set. If the correct movement for trSets(1,:) is number 1, then trOuts(1,:) = [1 0 0 0 0 0 0 0]. So “m” for both trSets and trOuts is the same, since for each feature sets, there is a corresponding output.

These are data sets for 6 individual movements, and also combination by 2. If the testing set x is a combination of movement 2 and 4, then tOut(x,:) = [0 1 0 1 0 0 0]. You will notice that the training (tr) and validation (v) sets only have feature sets of individual movements. The testing (t) set, contains features sets of individual movements AND simultaneous movements. This is because the algorithm must only be trained with individual movements.

Finally, these are 2 features extracted in time windows of 200 ms, from 8 channels, so 16 features in total. These signal were recorded from surface electrode in the forearm of a subject performing the movements. From this particular set, I get an average accuracy over 80% including simultaneous movements. Let me know how it goes for you :)

Hi Max,

Thanks to the data set, I will see if I can try my hand at it, too.

However, I don't agree to the validation data being devoid of simultaneous movement cases: if the idea is that the classifier must be able to recognize these, then surely there must be a way to assess its performance at it, even if they're not present in the training set. And of course we cannot do that with the test set, it would be peeking.

Have you considered taking Fourier Transforms of your time signals, as a means for separating the signals? Or maybe taking a series of transforms using a sliding time window in order to follow the time variations in the frequency domain?

@Max

Thanks for the data. I don't have access to Matlab so I have thrown together a .mat file reader so I can see what you have.

Just a question: I have never been able to afford or even down load Matlab for use. Is your use of it due

to access via a university or did you buy a copy for your own use?

-Doug

@ Douglas:

The data can be opened in GNU Octave without any problems.

@ Helio, the thing is that you wont have the data sets for the simultaneous movements at all in the practical aplication. I recorded the simultaneous movements only for testing the algorithms. Every electrode placement and every patient would give you different signals. Anyhow, you are welcome to mix them as you like :)

@ Antonio, all the signals peek in the same frequencies, so it's hard to split them just with a FT. I have actually tried several signal features from the frequency domain but they give pretty much the same results, same with wavelets. Thanks.

@ Douglas, Yes, I have a matlab license through the university, but as Alireza said, octave is a good option.

@ Yuri and Thaker, any hint or better yet, a paper that explains why your algorithms would do good with mixed-labels? I'm not only looking for a general patrec algorithm, but one considerable better with mixed-labels :)

@Alirez and @Max,

I just looked at Octave. Yuk! Why in heavens name should I have to download all of those files and muck around with their installation? What happened to a single .exe file installation process??? Unix based systems force this nonsense on everyone.

Do I really want all of these files?

The included packages are:

actuarial-1.1.0

ad-1.0.6_patched

audio-1.1.4

benchmark-1.1.1

bim-1.0.2

bioinfo-0.1.2

civil-engineering-1.0.7

combinatorics-1.0.9

communications-1.1.0_svn20120127_patched

control-2.2.4

data-smoothing-1.2.3

dataframe-0.8.2

econometrics-1.0.8

fenv-0.1.0

financial-0.3.2

fpl-1.2.0

fuzzy-logic-toolkit-0.3.0

ga-0.9.8

general-1.2.2

generate_html-0.1.3

geometry-1.4.0

gnuplot-1.0.1

gpc-0.1.7

gsl-1.0.8

ident-1.0.7

image-1.0.15

informationtheory-0.1.8

integration-1.0.7_svn20120128

io-1.0.16

irsa-1.0.7

java-1.2.8_patched

linear-algebra-2.1.0_svn20120127

mapping-1.0.7

mechanics-1.2.0

miscellaneous-1.0.11_svn20120127

missing-functions-1.0.2

msh-1.0.2

multicore-0.2.15

nlwing2-1.2.0

nnet-0.1.13

nurbs-1.3.5

ocs-0.1.3_svn20120128_patched

octclip-1.0.0

octgpr-1.2.0

odebvp-1.0.6

odepkg-0.8.0_svn20120127

optim-1.0.17_patched

optiminterp-0.3.4_svn20120128_patched

outliers-0.13.9

physicalconstants-0.1.7

plot-1.1.0

quaternion-1.0.0

secs1d-0.0.8

secs2d-0.0.8

secs3d-0.0.1

signal-1.1.2

simp-1.1.0

sockets-1.0.7_svn20120128_patched

specfun-1.1.0

special-matrix-1.0.7

spline-gcvspl-1.0.8

splines-1.0.7

statistics-1.1.0_svn20120128

strings-1.0.7

struct-1.0.9

symband-1.0.10

symbolic-1.1.0

tcl-octave-0.1.8

time-1.0.9

tsa-4.1.1

video-1.0.2_patched

vrml-1.0.12_svn20111014_patched

windows-1.1.0

xraylib-1.0.8

zenity-0.5.7

Maintainer: Nitzan Arazi

Latest update: 2012-01-28

Hi Douglas,

You may want to take a look into Python with NumPy and SciPy; if anything it's relatively easier to install (three click-through Windows installers). Meanwhile, see the attached file for a plain-text version of the data sets.

Hi Max,

Yes, I understand that the algorithm will be trained on individual inputs only. But for development purposes, your validation set must have some composed cases: otherwise, how are you to gauge its performance dealing with them, and whether improvements are needed? You cannot use the test data for that; it would invalidate your development process.

@ Douglas,

No, just install the base package. Octave-forge (those packages you listed) are simply a set of packages that include functions that are in MATLAB but not used often enough to justify their inclusion in the base package. Whenever you need a function that Octave doesn't appear to have, simply locate it's corresponding package in Octave-forge and install it. You can think of them as plugins (MATLAB has a large library of plugins as well).

If all of them were installed together, it would just be a huge set of functions 99% of which you'd never use. I've been using octave for years and have only ever used just 2 octave-forge packages (specfun and communications).

Good luck.

@ Helio, it depends on your algorithm and your approach. It "must" not have for MLP, since you use the validation set to know when to stop the training, therefore the validation set is part of the training process. The testing set in the other hand, is never used during the training for any purpose. As I said, you can do as you want as long as you are aware that you won't have the simultaneous information during your training process in the real aplication.

A context-sensitive grammatical analysis approach may be of use - could you send me example data? [email protected]

http://www.ee.columbia.edu/~sfchang/course/svia/papers/jain99statistical-review.pdf

https://books.google.com.tr/books?hl=tr&lr=&id=3-ubOzjjneUC&oi=fnd&pg=PR11&dq=good+algorithm+for+pattern+recognition+of+mixed+labels&ots=MBDdOdgJob&sig=XFHzuGBj1fKFRY_sOABqJMJUqCc&redir_esc=y#v=onepage&q=good%20algorithm%20for%20pattern%20recognition%20of%20mixed%20labels&f=false