I tried to go through this paper which describes the ECT algorithm but could not make much out of it.
I know it is different from one-against-al (oaa) and even performs better than oaa.I wanted a simple explanation about how ect works.
ECT and Filter trees are useful (only) if you have a very big number of output labels (classes), let's say N=1000. With OAA (one-against-all), it would mean to do N binary classification tasks for each example (during both training and testing). With ECT you can make the prediction much faster: log(N). You can imagine Filter trees (which are the basis of ECT) as a decision tree where in each node you ask whether the example belongs to one set of labels or another set of labels (using all the features, unlike original decision trees).
In general, ECT is worse (in terms of loss or accuracy) than OAA (but in some cases it may be almost as good as OAA). With N=10 labels, you should try OAA first. With N>1000, OAA is too slow (and even the accuracy is low), you should try ECT (or --log_multi or --csoaa_ldf in VW, if you can preselect a smaller number of labels which are relevant for each example).
See http://cilvr.cs.nyu.edu/diglib/lsml/logarithmic.pdf
Related
I have estimated a complex hierarchical model with many random effects, but don't really know what the best approach is to checking for convergend. I have complex longitudinal data from a few hundred individuals and estimate quite a few parameters for every individual. Because of that, I have way to many traceplots to inspect visually. Or should I really spend a day going through all the traceplots? What would be a better way to check for convergence? Do I have to calculate Gelman and Rubin's Rhat for every parameter on the person level? And when can I conclude that the model converged? When absolutely all of the thousends of parameters reached convergence? Is it even sensible to expect that? Or is there something like "overall convergence"? And what does it mean when some person-level parameters did not converge? Does it make sense to use autorun.jags from the R2jags package with such a model or will it just run for ever? I know, these are a lot of question, but I just don't know how to approach that.
The measure I am using for convergence is a potential scale reduction factor (psrf)* using the gelman.diag function from the R package coda.
But nevertheless, I am also quickly visually inspecting all the traceplots, even though I also have tens/hundreds of them. It can be really fast if you put them in PNG files and then quickly go through them using e.g. IrfanView (let me know if you need me to expand on this).
The reason you should inspect the traceplots is pretty well described by an example from Marc Kery (author of great Bayesian books): see "Never blindly trust Rhat for convergence in a Bayesian analysis", here I include a self explanatory image from this email:
This is related to Rhat statistics while I use psrf, but it's pretty likely that psrf suffers from this too... and better to check the chains.
*) Gelman, A. & Rubin, D. B. Inference from iterative simulation using multiple sequences. Stat. Sci. 7, 457–472 (1992).
I have a question about CBOW prediction. Suppose my job is to use 3 surrounding words w(t-3), w(t-2), w(t-1)as input to predict one target word w(t). Once the model is trained and I want to predict a missing word after a sentence. Does this model only work for a sentence with four words which the first three are known and the last is unknown? If I have a sentence in 10 words. The first nine words are known, can I use 9 words as input to predict the last missing word in that sentence?
Word2vec CBOW mode typically uses symmetric windows around a target word. But it simply averages the (current in-training) word-vectors for all words in the window to find the 'inputs' for the prediction neural-network. Thus, it is tolerant of asymmetric windows – if there are fewer words are available on either side, fewer words on that side are used (and perhaps even zero on that side, for words at the front/end of a text).
Additionally, during each training example, it doesn't always use the maximum-window specified, but some random-sized window up-to the specified size. So for window=5, it will sometimes use just 1 on either side, and other times 2, 3, 4, or 5. This is done to effectively overweight closer words.
Finally and most importantly for your question, word2vec doesn't really do a full-prediction during training of "what exact word does the model say should be heat this target location?" In either the 'hierarchical softmax' or 'negative-sampling' variants, such an exact prediction can be expensive, requiring calculations of neural-network output-node activation levels proportionate to the size of the full corpus vocabulary.
Instead, it does the much-smaller number-of-calculations required to see how strongly the neural-network is predicting the actual target word observed in the training data, perhaps in contrast to a few other words. In hierarchical-softmax, this involves calculating output nodes for a short encoding of the one target word – ignoring all other output nodes encoding other words. In negative-sampling, this involves calculating the one distinct output node for the target word, plus a few output nodes for other randomly-chosen words (the 'negative' examples).
In neither case does training know if this target word is being predicted in preference over all other words – because it's not taking the time to evaluate all others words. It just looks at the current strength-of-outputs for a real example's target word, and nudges them (via back-propagation) to be slightly stronger.
The end result of this process is the word-vectors that are usefully-arranged for other purposes, where similar words are close to each other, and even certain relative directions and magnitudes also seem to match human judgements of words' relationships.
But the final word-vectors, and model-state, might still be just mediocre at predicting missing words from texts – because it was only ever nudged to be better on individual examples. You could theoretically compare a model's predictions for every possible target word, and thus force-create a sort of ranked-list of predicted-words – but that's more expensive than anything needed for training, and prediction of words like that isn't the usual downstream application of sets of word-vectors. So indeed most word2vec libraries don't even include any interface methods for doing full target-word prediction. (For example, the original word2vec.c from Google doesn't.)
A few versions ago, the Python gensim library added an experimental method for prediction, [predict_output_word()][1]. It only works for negative-sampling mode, and it doesn't quite handle window-word-weighting the same way as is done in training. You could give it a try, but don't be surprised if the results aren't impressive. As noted above, making actual predictions of words isn't the usual real goal of word2vec-training. (Other more stateful text-analysis, even just large co-occurrence tables, might do better at that. But they might not force word-vectors into interesting constellations like word2vec.)
How does word2vec create vectors for words? I trained two word2vec models using two different files (from commoncrawl website) but I am getting same word vectors for a given word from both models.
Actually, I have created multiple word2vec models using different text files from the commoncrawl website. Now I want to check which model is better among all. How can select the best model out of all these models and why I am getting same word vectors for different models?
Sorry, If the question is not clear.
If you are getting identical word-vectors from models that you've prepared from different text corpuses, something is likely wrong in your process. You may not be performing any training at all, perhaps because of a problem in how the text iterable is provided to the Word2Vec class. (In that case, word-vectors would remain at their initial, randomly-initialized values.)
You should enable logging, and review the logs carefully to see that sensible counts of words, examples, progress, and incremental-progress are displayed during the process. You should also check that results for some superficial, ad-hoc checks look sensible after training. For example, does model.most_similar('hot') return other words/concepts somewhat like 'hot'?
Once you're sure models are being trained on varied corpuses – in which case their word-vectors should be very different from each other – deciding which model is 'best' depends on your specific goals with word-vectors.
You should devise a repeatable, quantitative way to evaluate a model against your intended end-uses. This might start crudely with a few of your own manual reviews of results, like looking over most_similar() results for important words for better/worse results – but should become more extensive. rigorous, and automated as your project progresses.
An example of such an automated scoring is the accuracy() method on gensim's word-vectors object. See:
https://github.com/RaRe-Technologies/gensim/blob/6d6f5dcfa3af4bc61c47dfdf5cdbd8e1364d0c3a/gensim/models/keyedvectors.py#L652
If supplied with a specifically-formatted file of word-analogies, it will check how well the word-vectors solve those analogies. For example, the questions-words.txt of Google's original word2vec code release includes the analogies they used to report vector quality. Note, though, that the word-vectors that are best for some purposes, like understanding text topics or sentiment, might not also be the best at solving this style of analogy, and vice-versa. If training your own word-vectors, it's best to choose your training corpus/parameters based on your own goal-specific criteria for what 'good' vectors will be.
This seems to be a fundamental question which some of you out there must have an opinion on. I have an image classifier implemented in CNTK with 48 classes. If the image does not match any of the 48 classes very well, then I'd like to be able to conclude that it was not among these 48 image types. My original idea was simply that if the highest output of the final Softmax layer was low, I would be able to conclude that the test image matched none well. While I occasionally see this occur, in most testing, Softmax still produces a very high (and mistaken) result when handed an 'unknown image type'. But maybe my network is 'over fit' and if it wasn't, my original idea would work fine. What do you think? Any way to define a 49-th class called 'none-of-the-above'?
You really have these two options indeed--thresholding the posterior probabilities (softmax values), and adding a garbage class.
In my area (speech), both approaches are their place:
If "none of the above" inputs are of the same nature as the "above" (e.g. non-grammatical inputs), thresholding works fine. Note that the posterior probability for a class is equal to one minus an estimate of the error rate for choosing this class. Rejecting anything with posterior < 50% would be rejecting all cases where you are more likely wrong than right. As long as your none-of-the-above classes are of similar nature, the estimate may be accurate enough to make this correct for them as well.
If "none of the above" inputs are of similar nature but your number of classes is very small (e.g. 10 digits), or if the inputs are of a totally different nature (e.g. a sound of a door slam or someone coughing), thresholding typically fails. Then, one would train a "garbage model." In our experience, it is OK to include the training data for the correct classes. Now the none-of-the-above class may match a correct class as well. But that's OK as long as the none-of-the-above class is not overtrained--its distribution will be much flatter, and thus even if it matches a known class, it will match it with a lower score and thus not win against the actual known class' softmax output.
In the end, I would use both. Definitely use a threshold (to catch the cases that the system can rule out) and use a garbage model, which I would just train it on whatever you have. I would expect that including the correct examples in training will not harm, even if it is the only data you have (please check the paper Anton posted for whether that applies to image as well). It may also make sense to try to synthesize data, e.g. by randomly combining patches from different images.
I agree with you that this is a key question, but I am not aware of much work in that area either.
There's one recent paper by Zhang and LeCun, that addresses the question for image classification in particular. They use large quantities of unlabelled data to create an additional "none of the above" class. The catch though is that, in some cases, their unlabelled data is not completely unlabelled, and they have means of removing "unlabelled" images that are actually in one of their labelled classes. Having said that, the authors report that apart from solving the "none of the above" problem, they even see performance gains even on their test sets.
As for fitting something post-hoc, just by looking at the outputs of the softmax, I can't provide any pointers.
I have two (or three) classes and each classes can only possess one label.
I want to optimize (automatically if possible) parameters and thresholds of classifiers in order for my first class to contain only 100 % sure data. Even if it contains a small number of instances.
I don't mind for the remaining classes to contain false alarm or correct rejection.
I don't mind to have unclassified data.
I have already been searching on stackoverflow and on the weka's wiki but maybe my lack of knowledge concerning weka made me miss some keywords.
I also tried to perform the task with the well-known "iris" database but I think that in this case, any class can be 100 % sure.
Yet, I have only succeed in testing multiple classifiers and tuning them manually but without performing 100 % correct for my first class. (I checked this result in the confusion matrix given by weka's report.)
Somehow, I know it is possible for my class to contain 100% sure data because I managed to do it in Matlab with simple threshold set manually. But I would like to try out a bigger database, to obtain better threshold and to use the power of weka.
Any suggestions would be helpful, thanks !
You probably need the "Cost Sensitive Classifier" among "meta" classifiers.
If you are working in the Explorer, here is the dialog you get.
Choose the your "classifier" (something beyond ZeroR :) ).
Set your "cost matrix". For 2-class problem this will be 2x2 matrix.
By setting one non-diagonal component very large (>>1, let us say 1000) you ensure that misclassifying one class (your "first" class) is 1000 times more expensive than misclassifying another class. This should do the job.