I have been reading papers about the Markov model, suddenly a great extension like TML(Tractable Markov Logic) coming out.
It is a subset of Markov logic, and uses probabilistic class and part hierarchies to control complexity.
This model has both complex logical structure and uncertainty.
It can represent objects, classes, and relations between objects, subject to certain restrictions which ensure that inference in any model built in TML can be queried efficiently.
I am just wondering why such a good idea not widely spreading around the area of application scenarios like activity analysis?
More info
My understanding is that TML is polynomial in the size of the model, but the size of the model needs to be compiled to a given problem and may become exponentially large. So, at the end, it's still not really tractable.
However, it may be advantageous to use it in the case that the compiled form will be used multiple times, because then the compilation is done only once for multiple queries. Also, once you obtain the compiled form, you know what to expect in terms of run-time.
However, I think the main reason you don't see TML being used more broadly is that it is just an academic idea. There is no robust, general-purpose system based on it. If you try to work on a real problem with it, you will probably find out that it lacks certain practical features. For example, there is no way to represent a normal distribution with it, and lots of problems involve normal distributions. In such cases, one may still use the ideas behind the TML paper but would have to create their own implementation that includes further features needed for the problem at hand. This is a general problem that applies to lots and lots of academic ideas. Only a few become really useful and the basis of practical systems. Most of them exert influence at the level of ideas only.
Related
Suppose you're trying to use machine learning for a classification task like, let's say, looking at photographs of animals and distinguishing horses from zebras. This task would seem to be within the state of the art.
But if you take a bunch of labelled photographs and throw them at something like a neural network or support vector machine, what happens in practice is that zebras are so much rarer than horses that the system just ends up learning to say 'always a horse' because this is actually the way to minimize its error.
Minimal error that may be but it's also not a very useful result. What is the recommended way to tell the system 'I want the best guess at which photographs are zebras, even if this does create some false positives'? There doesn't seem to be a lot of discussion of this problem.
One of the things I usually do with imbalanced classes (or skewed data sets) is simply generate more data. I think this is the best approach. You could go out in the real world and gather more data of the imbalanced class (e.g. find more pictures of zebras). You could also generate more data by simply making copies or duplicating it with transformations (e.g. flip horizontally).
You could also pick a classifier that uses an alternate evaluation (performance) metric over the one usually used - accuracy. Look at precision/recall/F1 score.
Week 6 of Andrew Ng's ML course talks about this topic: link
Here is another good web page I found on handling imbalanced classes: link
With this type of unbalanced data problem, it is a good approach to learn patterns associated with each class as opposed to simply comparing classes - this can be done via unsupervised learning learning first (such as with autoencoders). A good article with this available at https://www.r-bloggers.com/autoencoders-and-anomaly-detection-with-machine-learning-in-fraud-analytics/amp/. Another suggestion - after running the classifier, the confusion matrix can be used to determine where additional data should be pursued (I.e. many zebra errors)
[I'm approaching this as an outsider to machine learning. It just seems like a classification problem which I should be able to solve with fairly good accuracy with Machine Larning.]
Training Dataset:
I have millions of URLs, each tagged with a particular category. There are limited number of categories (50-100).
Now given a fresh URL, I want to categorize it into one of those categories. The category can be determined from the URL using conventional methods, but would require a huge unmanageable mess of pattern matching.
So I want to build a box where INPUT is URL, OUTPUT is Category. How do I build this box driven by ML?
As much as I would love to understand the basic fundamentals of how this would work out mathematically, right now much much more focussed on getting it done, so a conceptual understanding of the systems and processes involved is what I'm looking to get. I suppose machine learning is at a point where you can approach reasonably straight forward problems in that manner.
If you feel I'm wrong and I need to understand the foundations deeply in order to get value out of ML, do let me know.
I'm building this inside an AWS ecosystem so I'm open to using Amazon ML if it makes things quicker and simpler.
I suppose machine learning is at a point where you can approach reasonably straight forward problems in that manner.
It is not. Building an effective ML solution requires both an understanding of problem scope/constraints (in your case, new categories over time? Runtime requirements? Execution frequency? Latency requirements? Cost of errors? and more!). These constraints will then impact what types of feature engineering / processing you may look at, and what types of models you will look at. Your particular problem may also have issues with non I.I.D. data, which is an assumption of most ML methods. This would impact how you evaluate the accuracy of your model.
If you want to learn enough ML to do this problem, you might want to start looking at work done in Malicious URL classification. An example of which can be found here. While you could "hack" your way to something without learning more about ML, I would not personally trust any solution built in that manner.
If you feel I'm wrong and I need to understand the foundations deeply in order to get value out of ML, do let me know.
Okay, I'll bite.
There are really two schools of thought currently related to prediction: "machine learners" versus statisticians. The former group focuses almost entirely on practical and applied prediction, using techniques like k-fold cross-validation, bagging, etc., while the latter group is focused more on statistical theory and research methods. You seem to fall into the machine-learning camp, which is fine, but then you say this:
As much as I would love to understand the basic fundamentals of how this would work out mathematically, right now much much more focussed on getting it done, so a conceptual understanding of the systems and processes involved is what I'm looking to get.
While a "conceptual understanding of the systems and processes involved" is a prerequisite for doing advanced analytics, it isn't sufficient if you're the one conducting the analysis (it would be sufficient for a manager, who's not as close to the modeling).
With just a general idea of what's going on, say, in a logistic regression model, you would likely throw all statistical assumptions (which are important) to the wind. Do you know whether certain features or groups shouldn't be included because there aren't enough observations in that group for the test statistic to be valid? What can happen to your predictions and hypotheses when you have high variance-inflation factors?
These are important considerations when doing statistics, and oftentimes people see how easy it is to do from sklearn.svm import SVC or somthing like that and run wild. That's how you get caught with your pants around your ankles.
How do I build this box driven by ML?
You don't seem to have even a rudimentary understanding of how to approach machine/statistical learning problems. I would highly recommend that you take an "Introduction to Statistical Learning"- or "Intro to Regression Modeling"-type course in order to think about how you translate the URLs you have into meaningful features that have significant power predicting URL class. Think about how you can decompose a URL into individual pieces that might give some information as to which class a certain URL pertains. If you're classifying espn.com domains by sport, it'd be pretty important to parse nba out of http://www.espn.com/nba/team/roster/_/name/cle, don't you think?
Good luck with your project.
Edit:
To nudge you along, though: every ML problem boils down to some function mapping input to output. Your outputs are URL classes. Your inputs are URLs. However, machines only understand numbers, right? URLs aren't numbers (AFAIK). So you'll need to find a way to translate information contained in the URLs to what we call "features" or "variables." One place to start, there, would be one-hot encoding different parts of each URL. Think of why I mentioned the ESPN example above, and why I extracted info like nba from the URL. I did that because, if I'm trying to predict to which sport a given URL pertains, nba is a dead giveaway (i.e. it would very likely be highly predictive of sport).
Sorry, I have just started machine learning and am not by any means an expert in it. So, most likely this question will sound ignorant which I am afraid that I cannot avoid. Also, I searched to the best of my ability and was incapable of finding similar questions or answers that may address my question.
I learned that a model cannot learn if it was not from a dataset that has a normal distribution. Also, the only way I use to find out that a data set is normally distributed is the graphical method described here for each parameter. Which may be unadvisable, and if so I am always subject to change, so please correct me if that is the case.
To get to my question, if I see a normal distribution for certain parameters
yet not for a few others, does that mean the dataset is flawed? Or does it mean that I should not use those parameters for the model?
Thanks in advance, and sorry if there are any fundamental errors in my understanding of the concepts.
As cel said, every model has its own assumptions and limitations.
While there might be a model that can only learn on completely normally distributed data - there are plenty of models which don't, such as SVMs or Random Forests.
In practice if you know that your data does not conform to the assumptions of your model you could consider using a different model or to manipulate your data to fit your assumption.
The latter option is something that you should consider carefully to make sure your manipulation does not render your model useless when used in real-life scenarios.
My question is related to the project I've just started working on, and it's a ChatBot.
The bot I want to build has a pretty simple task. It has to automatize the process of purchasing movie tickets. This is pretty close domain and the bot has all the required access to the cinema database. Of course it is okay for the bot to answer like “I don’t know” if user message is not related to the process of ordering movie tickets.
I already created a simple demo just to show it to a few people and see if they are interested in such a product. The demo uses simple DFA approach and some easy text matching with stemming. I hacked it in a day and it turned out that users were impressed that they are able to successfully order tickets they want. (The demo uses a connection to the cinema database to provide users all needed information to order tickets they desire).
My current goal is to create the next version, a more advanced one, especially in terms of Natural Language Understanding. For example, the demo version asks users to provide only one information in a single message, and doesn’t recognize if they provided more relevant information (movie title and time for example). I read that an useful technique here is called "Frame and slot semantics", and it seems to be promising, but I haven’t found any details about how to use this approach.
Moreover, I don’t know which approach is the best for improving Natural Language Understanding. For the most part, I consider:
Using “standard” NLP techniques in order to understand user messages better. For example, synonym databases, spelling correction, part of speech tags, train some statistical based classifiers to capture similarities and other relations between words (or between the whole sentences if it’s possible?) etc.
Use AIML to model the conversation flow. I’m not sure if it’s a good idea to use AIML in such a closed domain. I’ve never used it, so that’s the reason I’m asking.
Use a more “modern” approach and use neural networks to train a classifier for user messages classification. It might, however, require a lot of labeled data
Any other method I don’t know about?
Which approach is the most suitable for my goal?
Do you know where I can find more resources about how does “Frame and slot semantics” work in details? I'm referring to this PDF from Stanford when talking about frame and slot approach.
The question is pretty broad, but here are some thoughts and practical advice, based on experience with NLP and text-based machine learning in similar problem domains.
I'm assuming that although this is a "more advanced" version of your chatbot, the scope of work which can feasibly go into it is quite limited. In my opinion this is a very important factor as different methods widely differ in the amount and type of manual effort needed to make them work, and state-of-the-art techniques might be largely out of reach here.
Generally the two main approaches to consider would be rule-based and statistical. The first is traditionally more focused around pattern matching, and in the setting you describe (limited effort can be invested), would involve manually dealing with rules and/or patterns. An example for this approach would be using a closed- (but large) set of templates to match against user input (e.g. using regular expressions). This approach often has a "glass ceiling" in terms of performance, but can lead to pretty good results relatively quickly.
The statistical approach is more about giving some ML algorithm a bunch of data and letting it extract regularities from it, focusing the manual effort in collecting and labeling a good training set. In my opinion, in order to get "good enough" results the amount of data you'll need might be prohibitively large, unless you can come up with a way to easily collect large amounts of at least partially labeled data.
Practically I would suggest considering a hybrid approach here. Use some ML-based statistical general tools to extract information from user input, then apply manually built rules/ templates. For instance, you could use Google's Parsey McParseface to do syntactic parsing, then apply some rule engine on the results, e.g. match the verb against a list of possible actions like "buy", use the extracted grammatical relationships to find candidates for movie names, etc. This should get you to pretty good results quickly, as the strength of the syntactic parser would allow "understanding" even elaborate and potentially confusing sentences.
I would also suggest postponing some of the elements you think about doing, like spell-correction, and even stemming and synonyms DB - since the problem is relatively closed, you'll probably have better ROI from investing in a rule/template-framework and manual rule creation. This advice also applies to explicit modeling of conversation flow.
I am asking this question because I know there are a lot of well-read CS types on here who can give a clear answer.
I am wondering if such an AI exists (or is being researched/developed) that it writes programs by generating and compiling code all on it's own and then progresses by learning from former iterations. I am talking about working to make us, programmers, obsolete. I'm imagining something that learns what works and what doesn't in a programming languages by trial and error.
I know this sounds pie-in-the-sky so I'm asking to find out what's been done, if anything.
Of course even a human programmer needs inputs and specifications, so such an experiment has to have carefully defined parameters. Like if the AI was going to explore different timing functions, that aspect has to be clearly defined.
But with a sophisticated learning AI I'd be curious to see what it might generate.
I know there are a lot of human qualities computers can't replicate like our judgement, tastes and prejudices. But my imagination likes the idea of a program that spits out a web site after a day of thinking and lets me see what it came up with, and even still I would often expect it to be garbage; but maybe once a day I maybe give it feedback and help it learn.
Another avenue of this thought is it would be nice to give a high-level description like "menued website" or "image tools" and it generates code with enough depth that would be useful as a code completion module for me to then code in the details. But I suppose that could be envisioned as a non-intelligent static hierarchical code completion scheme.
How about it?
Such tools exist. They are the subject of a discipline called Genetic Programming. How you evaluate their success depends on the scope of their application.
They have been extremely successful (orders of magnitude more efficient than humans) to design optimal programs for the management of industrial process, automated medical diagnosis, or integrated circuit design. Those processes are well constrained, with an explicit and immutable success measure, and a great amount of "universe knowledge", that is a large set of rules on what is a valid, working, program and what is not.
They have been totally useless in trying to build mainstream programs, that require user interaction, because the main item a system that learns needs is an explicit "fitness function", or evaluation of the quality of the current solution it has come up with.
Another domain that can be seen in dealing with "program learning" is Inductive Logic Programming, although it is more used to provide automatic demonstration or language / taxonomy learning.
Disclaimer: I am not a native English speaker nor an expert in the field, I am an amateur - expect imprecisions and/or errors in what follow. So, in the spirit of stackoverflow, don't be afraid to correct and improve my prose and/or my content. Note also that this is not a complete survey of automatic programming techniques (code generation (CG) from Model-Driven Architectures (MDAs) merits at least a passing mention).
I want to add more to what Varkhan answered (which is essentially correct).
The Genetic Programming (GP) approach to Automatic Programming conflates, with its fitness functions, two different problems ("self-compilation" is conceptually a no-brainer):
self-improvement/adaptation - of the synthesized program and, if so desired, of the synthesizer itself; and
program synthesis.
w.r.t. self-improvement/adaptation refer to Jürgen Schmidhuber's Goedel machines: self-referential universal problem solvers making provably optimal self-improvements. (As a side note: interesting is his work on artificial curiosity.) Also relevant for this discussion are Autonomic Systems.
w.r.t. program synthesis, I think is possible to classify 3 main branches: stochastic (probabilistic - like above mentioned GP), inductive and deductive.
GP is essentially stochastic because it produces the space of likely programs with heuristics such as crossover, random mutation, gene duplication, gene deletion, etc... (than it tests programs with the fitness function and let the fittest survive and reproduce).
Inductive program synthesis is usually known as Inductive Programming (IP), of which Inductive Logic Programming (ILP) is a sub-field. That is, in general the technique is not limited to logic program synthesis or to synthesizers written in a logic programming language (nor both are limited to "..automatic demonstration or language/taxonomy learning").
IP is often deterministic (but there are exceptions): starts from an incomplete specification (such as example input/output pairs) and use that to constraint the search space of likely programs satisfying such specification and then to test it (generate-and-test approach) or to directly synthesize a program detecting recurrences in the given examples, which are then generalized (data-driven or analytical approach). The process as a whole is essentially statistical induction/inference - i.e. considering what to include into the incomplete specification is akin to random sampling.
Generate-and-test and data-driven/analytical§ approaches can be quite fast, so both are promising (even if only little synthesized programs are demonstrated in public until now), but generate-and-test (like GP) is embarrassingly parallel and then notable improvements (scaling to realistic program sizes) can be expected. But note that Incremental Inductive Programming (IIP)§, which is inherently sequential, has demonstrated to be orders of magnitude more effective of non-incremental approaches.
§ These links are directly to PDF files: sorry, I am unable to find an abstract.
Programming by Demonstration (PbD) and Programming by Example (PbE) are end-user development techniques known to leverage inductive program synthesis practically.
Deductive program synthesis start with a (presumed) complete (formal) specification (logic conditions) instead. One of the techniques leverage automated theorem provers: to synthesize a program, it constructs a proof of the existence of an object meeting the specification; hence, via Curry-Howard-de Bruijn isomorphism (proofs-as-programs correspondence and formulae-as-types correspondence), it extracts a program from the proof. Other variants include the use of constraint solving and deductive composition of subroutine libraries.
In my opinion inductive and deductive synthesis in practice are attacking the same problem by two somewhat different angles, because what constitute a complete specification is debatable (besides, a complete specification today can become incomplete tomorrow - the world is not static).
When (if) these techniques (self-improvement/adaptation and program synthesis) will mature, they promise to rise the amount of automation provided by declarative programming (that such setting is to be considered "programming" is sometimes debated): we will concentrate more on Domain Engineering and Requirements Analysis and Engineering than on software manual design and development, manual debugging, manual system performance tuning and so on (possibly with less accidental complexity compared to that introduced with current manual, not self-improving/adapting techniques). This will also promote a level of agility yet to be demonstrated by current techniques.