Posted by
Kanav
Saturday, October 3, 2009
Some Feedback
Content:So it seems like math may start getting to most of you... especially in pattern recognition. what is the thought of the class on a single day bootcamp for math...
Posted by
Posted by
Content:
Posted by
In Mr. Ji’s second lecture, we again move into an area that is not completely transparent to the uninitiated. In the second iteration, the quality of his handout from initial lecture to this lecture has more sense. I suspect that Tan et al’s book on data mining would be instructive. Given the shear amount of data in biomedical informatics and bioinformatics, the ability to use machine learning and data mining to interpret and find relationships where they are not immediately identifiable (unsupervised learning) versus asking discrete questions with large data sets, i.e. supervised learning is important. Training sets become ways to evaluate models and with further testing with test sets. Classification is one approach for machine learning and data mining, but decision trees, complex mathematical support vectors, regression, and neural networks seem most applicable to clinical and bioinformatics questions. The nearest neighbor analysis that is part of classification can become part of clustering methods; it is unclear to me how attributes in this situation distinguish supervised versus unsupervised learning. Part of the k means clustering involves the use of centroids, which is the mean of points in the cluster. I need or maybe I should say I would like to know what Euclidean and cosine similarity is. I suspect that hierarchical clustering with dendrograms, including agglomerative and divisive types are more useful in this data mining approach. Intercluster similarity is an important piece of this approach. Further cluster validity can be determined. Clustering and classification protocols can be brought together into supervised and unsupervised application with regression, another data mining technique. A combination of supervised and unsupervised machine learning may allow bringing both approaches together for solution of a variety of BMI and bioinformatics problems. What the differences between labeled and unlabeled for supervised and unsupervised learning needs to be clarified.
Stuart
Posted by
When I was young, I grew up on the South side of Chicago, on the campus of the University of Chicago. This was an interesting neighborhood, as the site of the Manhattan Project was 1 block away from my home and many very literal individuals arguing about every possible concept walked the streets. When I saw the slide of the computer guys discussing and setting up guidelines, theories, and new languages and programs, I thought maybe I had returned home. The clear limitations of this singular approach and focus must be balanced against the need for interoperability if we are going to advance clinical care. In sharp contrast to these left brained folks, Dr. Fridsma provides us with both right and left brain excellence and is in sharp contrast with balance; we certainly can see the conflict. These blogs are useful because they provide a means to set down our thoughts and understanding of topics that are not necessarily immediately clear. As such, I will lay out what my concept is wrong or right on his latest 502 lecture. It is extremely important to move to machine computable applications so that information and meaning can be shared. This should not only be syntactical, but also semantic. Ontologies that are domain specific serve as the underpinning of semantics and are purposeful as part of an engineered artifact. They serve as hard specifications for conceptualizations as quoted. Ontologies can be applied to syntactic as well as semantic realms and provide interoperability, an essential feature for data exchange and exchange of meaning, but with limitations. Interoperability is defined as exchanging and using information between systems. These ontologies, which have names for concepts and also domain specific knowledge, depend on controlled vocabularies that may have fundamental issues related to enumeration. Controlled vocabularies provide a means to resolve semiotic triangles that involve different representations of the same item or phenomena with a single word. Controlled vocabularies along with ontology languages specify ontologies and the degree of formality increases the ability for machines to appropriately communicate. RDFS or resource description framework with schema vocabulary and expanded URIs is one language attempt, which has limitations. It is not clear to me how RDFS or other ontology languages provide a means to utilize XML to foster interoperability. RDFS has limitations in attaching meaning to information or reasoning support. This is improvement over syntactical web with information that is hyperlinked. Unfortunately, as illustrated by the opposing ideas and groups that are trying to achieve interoperability, there is neither uniformity, unanimity present or even a definitive model on how to handle these issues of ontologies and interoperability and the functionality of the semantic web. OWL language has the potential for meeting some of these objectives, but still only takes up several levels on the “layer cake.” Comments on these interpretations would be useful.
In Mr. Ji’s second lecture, we again move into an area that is not completely transparent to the uninitiated. In the second iteration, the quality of his handout from initial lecture to this lecture has more sense. I suspect that Tan et al’s book on data mining would be instructive. Given the shear amount of data in biomedical informatics and bioinformatics, the ability to use machine learning and data mining to interpret and find relationships where they are not immediately identifiable (unsupervised learning) versus asking discrete questions with large data sets, i.e. supervised learning is important. Training sets become ways to evaluate models and with further testing with test sets. Classification is one approach for machine learning and data mining, but decision trees, complex mathematical support vectors, regression, and neural networks seem most applicable to clinical and bioinformatics questions. The nearest neighbor analysis that is part of classification can become part of clustering methods; it is unclear to me how attributes in this situation distinguish supervised versus unsupervised learning. Part of the k means clustering involves the use of centroids, which is the mean of points in the cluster. I need or maybe I should say I would like to know what Euclidean and cosine similarity is. I suspect that hierarchical clustering with dendrograms, including agglomerative and divisive types are more useful in this data mining approach. Intercluster similarity is an important piece of this approach. Further cluster validity can be determined. Clustering and classification protocols can be brought together into supervised and unsupervised application with regression, another data mining technique. A combination of supervised and unsupervised machine learning may allow bringing both approaches together for solution of a variety of BMI and bioinformatics problems. What the differences between labeled and unlabeled for supervised and unsupervised learning needs to be clarified.
Stuart
Week Six: Ontologies/Semantics & Machine Learning
Content:
Overall, as Annie mentions, our colleagues have extensively covered much of the topics that were cover this week by Dr. Fridsma and Mr. Ji. I will just sum up some quick important points I got from the lectures.
Ontologies/Semantic Web-
Dr. Fridsma's presentation was informational and understandable, although it is hard to distinguish lecture from lecture since he taught 3 of the 4 lectures this week. He spoke of the difference of semantics and syntax:
He then spoke of how the exchange of information is syntactic interoperability while the use of the information is semantic interoperability. As well he defines for us that ontology is an engineering construct rather than an underlying truth. This is reasonable since if it was an underlying truth than it cannot be rebutted or tested as noted by Dr. Fridsma. The lecture then went on to all the three letter acronyms such as OWL, OIL, DAML-ONT, and etc. He also distinguised that there were different levels (versions) of OWL implementation:
- Semantics - meaning and understanding
- syntax - structure
The talk also included information of RDFs and its importance in helping to build vocabularies.
- OWL Full - Fully expressive but there are serious problems
- OWL dl - full power of descriptive objects, it is deliverable
- OWL lite - easy to implement
Machine Learning: unsupervised, supervised, and semisupervised-
Mr. Ji covered the topics of machine learning but focused primarily on k-nearest neighbor and k-means since these are the simpler of the topics and were appropriate as an introduction into the topic in the short period of the class.
In summary, supervised learning is when we want to determine the label of an attribute, but unsupervised learning is when we want to group data points not apply labels.
- Supervised Learning: This involves the topic of k-nearest neighbor to produce a comparative results based on an existing training set. The value of k in this case is very important since if k = all data, then the training set value that makes up the majority of set is returned, if k is too small (ie 1) then the results may not be comparative since it was compared to a small subset.
- Unsupervised Learning: This involved the idea of grouping similar objects into the same group, and different objects into different groups. We then covered k-means as a method that is used for unsupervised machine learning. Typically this method results in grouping were intradata point distances are minimized and intercluster distances are maximized.
- Semisupervised Learning: This is a topic where people believe that it is an ideal situation, however, despite looking like they work ideally this situation actually does not work.
Other Stuff
I found this interesting article in the news about how social networks that are being used by students are causing them to get into trouble with their schools and others that have resulted in disciplinary action that included expulsion from medical school. This could possible be applied to the other health professions as well.
http://news.bbc.co.uk/2/hi/health/8266546.stm
As well, a slightly off topic article on who is responsible for private information that stems from a bank employee sending customer information to the wrong person that included information about hundred of other customers. As one of our lecturers stated, the finacial system is very similar to how the healthcare system: finacial data versus PHI.
http://blogs.techrepublic.com.com/itdojo/?p=1031&tag=nl.e099.dl090930&tag=nl.e099
Posted by Eric
Friday, October 2, 2009
Ontology and datamining
Content:
I have to say the two classes in this week is tough, but I got tons of information from the Dr. Fridsma's introduction on ontology and Mr.Ji's introduction on datamining. Although I feel a little lost during the class because of to many new concepts and new terms flow into my mind, I feel more clear now after I review the information in the class.
In Dr.Fridsma's class about ontology, he talked from semantic web and compare the difference between semantic and syntatic, which give me a clear conception about the core value of the semantic and the significance of semantic language, especially in health care field. Actually, compared with the tradionally used the syntatic language, the semantic language more focus on the underlying understandable meaning of language, and concentrate on how to compose and structure the sentence or syntax to deliver the language to computer, and make computer can extract the meaningful information like human being understand the meaning of received information. I think this is the one important obstacle in the Aritificial Intellegence to prevent the direct communication between human beings and machine. When I first understand what is the semantic in the class, the first example that I thought about is one example given in the Dr. Shortillif's class. The example is that input the below sentence into computer, " The distance between city A and city B is 360 mile, one person drives a car from city A to B taking 6 hours. What is the average speed of the car?" After input these two sentences into computer, the computer can prompt up a dialog box showing "The average speed is 60 miles/hour". Althoug it is kind of a too stupid question to human being, how it can be understood by computer when computer gets this input and how can computer know what it should do to respond to this question are two real key questions. Actually, the one important step in this example is to set up a semantic language that can first transfer the human being readable language to the the computer readable language, and then the further processing will be conducted by computer to choose the right response to the question and calculate the result. For the current technology and protocol used for semantic representation of medical language, Dr. Fridsma also introduced many examples, such as XML format, ICD10, RDF and so on, the principle of each technology and standard were discussed briefly and advantage and disadvantage were also compared.
In Mr. Ji's second class on introduction of Data mining, the unsupervised data mining method-cluster, was mainly introduced. Compared to supervised method-classification, the cluster is also based on the predicators to set up the model. But the unsupervised cluster is not aim to assign a predicated label to the case with unknown label, but focus on how to group the similar cases together. Corresponding the supervised and unsupervised data mining methods, the corresponding two algorithms, k-nearest neighbor and k-means are introduced deeper, and simple demos were also played in class for understanding. In bioinformatics study, the cluster is one of the most common method for genetics and protein study. For example, the cluster method can be used to sample the the representative protein structure from molecular simulation result, and can be used to group the similar genomes into sub catagories. Although in the class there is no introduction on how to implement the algorithm into practice, there is one very easily use and powerful software-WEKA, which has most of the data mining algorithm and have been used in some studies in bioinformatics datamining.
The link of WEKA is as below. It is also free. You can try to play it.
http://www.cs.waikato.ac.nz/ml/weka/
Post by Di Pan
I have to say the two classes in this week is tough, but I got tons of information from the Dr. Fridsma's introduction on ontology and Mr.Ji's introduction on datamining. Although I feel a little lost during the class because of to many new concepts and new terms flow into my mind, I feel more clear now after I review the information in the class.
In Dr.Fridsma's class about ontology, he talked from semantic web and compare the difference between semantic and syntatic, which give me a clear conception about the core value of the semantic and the significance of semantic language, especially in health care field. Actually, compared with the tradionally used the syntatic language, the semantic language more focus on the underlying understandable meaning of language, and concentrate on how to compose and structure the sentence or syntax to deliver the language to computer, and make computer can extract the meaningful information like human being understand the meaning of received information. I think this is the one important obstacle in the Aritificial Intellegence to prevent the direct communication between human beings and machine. When I first understand what is the semantic in the class, the first example that I thought about is one example given in the Dr. Shortillif's class. The example is that input the below sentence into computer, " The distance between city A and city B is 360 mile, one person drives a car from city A to B taking 6 hours. What is the average speed of the car?" After input these two sentences into computer, the computer can prompt up a dialog box showing "The average speed is 60 miles/hour". Althoug it is kind of a too stupid question to human being, how it can be understood by computer when computer gets this input and how can computer know what it should do to respond to this question are two real key questions. Actually, the one important step in this example is to set up a semantic language that can first transfer the human being readable language to the the computer readable language, and then the further processing will be conducted by computer to choose the right response to the question and calculate the result. For the current technology and protocol used for semantic representation of medical language, Dr. Fridsma also introduced many examples, such as XML format, ICD10, RDF and so on, the principle of each technology and standard were discussed briefly and advantage and disadvantage were also compared.
In Mr. Ji's second class on introduction of Data mining, the unsupervised data mining method-cluster, was mainly introduced. Compared to supervised method-classification, the cluster is also based on the predicators to set up the model. But the unsupervised cluster is not aim to assign a predicated label to the case with unknown label, but focus on how to group the similar cases together. Corresponding the supervised and unsupervised data mining methods, the corresponding two algorithms, k-nearest neighbor and k-means are introduced deeper, and simple demos were also played in class for understanding. In bioinformatics study, the cluster is one of the most common method for genetics and protein study. For example, the cluster method can be used to sample the the representative protein structure from molecular simulation result, and can be used to group the similar genomes into sub catagories. Although in the class there is no introduction on how to implement the algorithm into practice, there is one very easily use and powerful software-WEKA, which has most of the data mining algorithm and have been used in some studies in bioinformatics datamining.
The link of WEKA is as below. It is also free. You can try to play it.
http://www.cs.waikato.ac.nz/ml/weka/
Post by Di Pan
Ontologies and Machine Learning
The first lecture's emphasis was placed on the importance of semantics-with examples of the semiotic triange, the semantic web, and various web ontology languages. The semiotic triange reference struck me as familiar because I could remember work instances in which the particular example that Dr. Fridsma gave us of a protocol represented different meanings to different people. Even in the same working environment, different people place meaning of an object in different ways, due to past/present experiences. The current state of things on the web places more importance on syntax (the exchange of information) rather on semantics (the use of that information in a meaningful way). The ideal situation is to implement more semantics into the way the web operates so that it more accurately represents what people are trying to achieve.
The continutation of the machine learning lecture focused on two main types of clustering-k means and hierarchical. Clustering has implications in many types of information processing, including neural networking and bioinformatics. As I can see, my classmates have done a great job of defining all the clustering methods (way to go everyone!), so I will just list a site where you can see clustering applied to gene expression data. It is from the Center for Information Technology-National Institutes of Health and is a good summary of their research and goals for the future pertaining to clustering on larger datasets (which is much needed!).
http://cit.nih.gov/NR/exeres/DBC46094-ECE8-4419-AE65-48E041CECCB1,frameless.htm
Posted by Annie
The continutation of the machine learning lecture focused on two main types of clustering-k means and hierarchical. Clustering has implications in many types of information processing, including neural networking and bioinformatics. As I can see, my classmates have done a great job of defining all the clustering methods (way to go everyone!), so I will just list a site where you can see clustering applied to gene expression data. It is from the Center for Information Technology-National Institutes of Health and is a good summary of their research and goals for the future pertaining to clustering on larger datasets (which is much needed!).
http://cit.nih.gov/NR/exeres/DBC46094-ECE8-4419-AE65-48E041CECCB1,frameless.htm
Posted by Annie
Week of 09/28/09
Content:
The first lecture was on Ontologies, Information models, and the Semantic Web by Dr. Fridsma. The lecture included a description of Tim Berners-Lee. Mr. Berners-Lee among other people invented the World Wide Web (web). Mr. Berners-Lee's vision of the web included an advanced semantic web. That vision of a semantic web is probably too hard to complete now. As an alternative to that semantic web, the web currently extensively uses syntactic representations of language. One method that can help with the addition of semantics to the web is the use of controlled vocabularies. However, enumerations in controlled vocabularies may be highly confusing with large vocabularies. Another way to help add semantics to the web is through the use of a resource description framework (RDF). RDF data models allow for subject, predicate, object connections which add semantic meaning to web pages. The Web Ontology Language (OWL) was created to go beyond RDF in depth of semantic meaning. OWL is available as a full union of OWL syntax and RDF (OWL Full), restricted to first order logic fragment (OWL DL), or an “easier to implement” subset of OWL DL (OWL Lite). I am impressed by the advanced forms of semantic technology that have been designed. I can imagine a future where programs on the web can provide useful responses to questions involving complex semantics. It would be nice to be able to get a computer to effectively interpret a quesiton like “What recent popular books contain science fiction without any metaphysics but with some adventure and current science?” Currently, I can find an answer to such a question as that but an advanced semantic system on the web could make my search for that answer much more efficient.
The second lecture was on machine learning with Shuiwang Ji. Supervised was a type of machine learning that was taught. K-nearest-neighbor (KNN) is a kind of supervised learning. KNN learns by comparing attributes to labels. KNN can use what it has learned from those comparisons to make predictions about labels for new sets of attributes. Similarity measures between attributes can be used to make those predictions. If KNN is being used to analyze attributes with continuous values than euclidean distances can be used to make comparisons of similarity measures of attributes. Other problem specific measures can be used for non-continuous measures. The production of similarity measures by scaling different measures of numbers seems interesting to me. In the lecture, three large attribute numbers and one small attribute number seemed to produce a similarity measure that was the most similar in size to the large numbers. I'm curious to know what mathematical technique was used to scale the attribute measures to create the similarity measurement. I am also interested in knowing how large past training sets of numbers have been that have produced accurate predictions. Perhaps how large training sets should be to be useful for predictions is related to the number of attributes that are compared.
Unsupervised, regression, and semi-supervised were other forms of machine learning that were taught. Two kinds of unsupervised learning methods are flat and hierarchical clustering. Flat clustering includes K-means, spectral, and graph-based clustering. K-means clustering uses centriods to designate clusters. K-means clustering uses an iterative process of reorganizing points to clusters. The centriods can be repositioned for improved clustering in each iteration of a k-means clustering process. Hierarchical clustering includes agglomerative and divisive clustering. Those two clustering techniques are different in the way they start. Agglomerative clustering starts with points as individual clusters. Divisive clustering starts with one all inclusive cluster of points. Machine learning with regression can identify linear or nonlinear forms of dependency in variables with continuous values. That regression has been used in statistics and neural network fields. Semi-supervised clustering can use a mixture of labeled and unlabeled data. Semi-supervised learning has been used for gene/protein function classifications as well as other applications. Additionally, the validities of clustering techniques have been tested with a variety of measures.
Putting semi-supervised learning in the context of genetics research and specifically gene/protein function predictions was interesting to me. I can imagine how attributes of genetic mutations could be clustered into groups of protein functions. It would be interesting to predict what gene mutations would lead to different protein functions. The connections between protein functions and genes could be further understood by examining what changes in gene sequences lead to different protein functions. For example, if a genetic mutation caused a large difference in protein function then that genetic mutation could be crucial to the protein's function. I found an article that explored some genetics research with a semi-supervised learning method. Here is a link to the article on using a semi-supervised learning method to predict transcription factor-gene interactions in Escherichia coli:
http://www.pubmedcentral.nih.gov.ezproxy1.lib.asu.edu/articlerender.fcgi?tool=pubmed&pubmedid=18369434
Posted by:
Nate
The first lecture was on Ontologies, Information models, and the Semantic Web by Dr. Fridsma. The lecture included a description of Tim Berners-Lee. Mr. Berners-Lee among other people invented the World Wide Web (web). Mr. Berners-Lee's vision of the web included an advanced semantic web. That vision of a semantic web is probably too hard to complete now. As an alternative to that semantic web, the web currently extensively uses syntactic representations of language. One method that can help with the addition of semantics to the web is the use of controlled vocabularies. However, enumerations in controlled vocabularies may be highly confusing with large vocabularies. Another way to help add semantics to the web is through the use of a resource description framework (RDF). RDF data models allow for subject, predicate, object connections which add semantic meaning to web pages. The Web Ontology Language (OWL) was created to go beyond RDF in depth of semantic meaning. OWL is available as a full union of OWL syntax and RDF (OWL Full), restricted to first order logic fragment (OWL DL), or an “easier to implement” subset of OWL DL (OWL Lite). I am impressed by the advanced forms of semantic technology that have been designed. I can imagine a future where programs on the web can provide useful responses to questions involving complex semantics. It would be nice to be able to get a computer to effectively interpret a quesiton like “What recent popular books contain science fiction without any metaphysics but with some adventure and current science?” Currently, I can find an answer to such a question as that but an advanced semantic system on the web could make my search for that answer much more efficient.
The second lecture was on machine learning with Shuiwang Ji. Supervised was a type of machine learning that was taught. K-nearest-neighbor (KNN) is a kind of supervised learning. KNN learns by comparing attributes to labels. KNN can use what it has learned from those comparisons to make predictions about labels for new sets of attributes. Similarity measures between attributes can be used to make those predictions. If KNN is being used to analyze attributes with continuous values than euclidean distances can be used to make comparisons of similarity measures of attributes. Other problem specific measures can be used for non-continuous measures. The production of similarity measures by scaling different measures of numbers seems interesting to me. In the lecture, three large attribute numbers and one small attribute number seemed to produce a similarity measure that was the most similar in size to the large numbers. I'm curious to know what mathematical technique was used to scale the attribute measures to create the similarity measurement. I am also interested in knowing how large past training sets of numbers have been that have produced accurate predictions. Perhaps how large training sets should be to be useful for predictions is related to the number of attributes that are compared.
Unsupervised, regression, and semi-supervised were other forms of machine learning that were taught. Two kinds of unsupervised learning methods are flat and hierarchical clustering. Flat clustering includes K-means, spectral, and graph-based clustering. K-means clustering uses centriods to designate clusters. K-means clustering uses an iterative process of reorganizing points to clusters. The centriods can be repositioned for improved clustering in each iteration of a k-means clustering process. Hierarchical clustering includes agglomerative and divisive clustering. Those two clustering techniques are different in the way they start. Agglomerative clustering starts with points as individual clusters. Divisive clustering starts with one all inclusive cluster of points. Machine learning with regression can identify linear or nonlinear forms of dependency in variables with continuous values. That regression has been used in statistics and neural network fields. Semi-supervised clustering can use a mixture of labeled and unlabeled data. Semi-supervised learning has been used for gene/protein function classifications as well as other applications. Additionally, the validities of clustering techniques have been tested with a variety of measures.
Putting semi-supervised learning in the context of genetics research and specifically gene/protein function predictions was interesting to me. I can imagine how attributes of genetic mutations could be clustered into groups of protein functions. It would be interesting to predict what gene mutations would lead to different protein functions. The connections between protein functions and genes could be further understood by examining what changes in gene sequences lead to different protein functions. For example, if a genetic mutation caused a large difference in protein function then that genetic mutation could be crucial to the protein's function. I found an article that explored some genetics research with a semi-supervised learning method. Here is a link to the article on using a semi-supervised learning method to predict transcription factor-gene interactions in Escherichia coli:
http://www.pubmedcentral.nih.gov.ezproxy1.lib.asu.edu/articlerender.fcgi?tool=pubmed&pubmedid=18369434
Posted by:
Nate
Ontology and Machine Learning revisited!!
Content:
This week we had 2 interesting sessions, one each from Dr. Fridsma and Dr. Ji. The first lecture was given by Dr. Fridsma about Ontologies and Symantic Webs. As Dr. Fridsma said, ontology is just a representation of a set of concepts within a domain and how those concepts are related to each other (Wikipedia). The best way to measure whether the ontology is correct or not is to compare the output based on the used ontology to the desired output. Based on the definition of ontology, we can imagine of using it in a semantic web, where everything is done based on what-it-tries-to-infer. A simple example of semantic web; I want to search for something that is burning. The meaning of burning is something that is on fire. So I would also like to see the pages that includes "on fire". But unfortunately, when I typed "burning" in the google search page, instead of providing me anything based on the actual meaning of "burning", it compared the word, it stemmed (process of extracting the base or root word) the word "burning", compared the huge library of web pages that included "burn", used some ranking algorithm to cluster the related web sites and presented that to me, and that was not what I was looking for. Similarly, when I typed "on fire" and clicked on 'Search', the first item on the list was a song by 50 Cent featuring Lloyd Banks :). There were times when I had to be embarassed because of google's stupid matches. But still I think that it's the best among the rest (bing lovers might not like the statement, try the same thing with bing and compare the results). So, to get the real thing that I wanted, I can create a ontology of "burning", and then categorize it according to different meanings (and meaning of the meaning upto certain depth), and then try to find different pages based on the new words, which carry the meanings of "burning". It's just a concept, and we all know this. If anyone have used Prolog before, then try this once, it's interesting. I love using Prolog to do these kinds of funny stuffs.
There was another interesting slide (I am saying that because that is the most premitive concept on how to make machines intelligent as humans) on first order predicate logic. The notations are easy but to re-write a sentence in FOPL is challanging. FoPL are mainly used to infer something from the given sets of knowledge (or simply knowledge base). We, humans, can easily understand the meaning of "child" and "parent", but we cannot simply add this concept to the machine's knowledge base. We have to translate the meaning of parent and child into machine "understandable" form first. And FOPL provides a very simple notations to convert a human readable sentence to a machine readable one. If anybody wants to go into the detail of FOPL, I'll suggest to go through the books "Artificial Intelligence" by Rich and Knight, and "Artificial Intelligence" by Russell and Norvig.
There was a brief talk on XML as well. According to w3cschools.com, "XML was designed to transport and store data, with focus on what data is; and HTML was designed to display data, with focus on how data looks". So, from the concept of ongology and the objective of XML, we can deal with the limitations of syntax based approach. We have to take a detailed course on XML, ontologies and semantic webs to learn how to implement these concepts together. I recommend the link: http://www.w3schools.com/xmL/xml_whatis.asp if you want to know (more) about XML.
The second lecture by Dr. Ji was the second part of machine learning series. He started from the review of
classification. He then explained the other important part of machine learning which is clustering. The importance of clustering is that it does not require a supervisor. Data sets are grouped together according to their features or values. The main disadvantage of clustering is that we don't know whether the results after clustering is good or bad (google's search result is an example). However, if we don't know the decision boundary and we don't know any guidance to create a boundary, then clustering is the best techniue to train a machine. He went further into the basic classes of clustering: fast clustering and hierarchical clustering. Regression and semi-supervised learning concepts were covered after that. The examples provided in the slides were helpful, but to understand each topic will be very difficult. For now, I'm sticking with K-nearest neighbour and K-means types as per his suggestion. :)
Posted by
Prabal Khanal
This week we had 2 interesting sessions, one each from Dr. Fridsma and Dr. Ji. The first lecture was given by Dr. Fridsma about Ontologies and Symantic Webs. As Dr. Fridsma said, ontology is just a representation of a set of concepts within a domain and how those concepts are related to each other (Wikipedia). The best way to measure whether the ontology is correct or not is to compare the output based on the used ontology to the desired output. Based on the definition of ontology, we can imagine of using it in a semantic web, where everything is done based on what-it-tries-to-infer. A simple example of semantic web; I want to search for something that is burning. The meaning of burning is something that is on fire. So I would also like to see the pages that includes "on fire". But unfortunately, when I typed "burning" in the google search page, instead of providing me anything based on the actual meaning of "burning", it compared the word, it stemmed (process of extracting the base or root word) the word "burning", compared the huge library of web pages that included "burn", used some ranking algorithm to cluster the related web sites and presented that to me, and that was not what I was looking for. Similarly, when I typed "on fire" and clicked on 'Search', the first item on the list was a song by 50 Cent featuring Lloyd Banks :). There were times when I had to be embarassed because of google's stupid matches. But still I think that it's the best among the rest (bing lovers might not like the statement, try the same thing with bing and compare the results). So, to get the real thing that I wanted, I can create a ontology of "burning", and then categorize it according to different meanings (and meaning of the meaning upto certain depth), and then try to find different pages based on the new words, which carry the meanings of "burning". It's just a concept, and we all know this. If anyone have used Prolog before, then try this once, it's interesting. I love using Prolog to do these kinds of funny stuffs.
There was another interesting slide (I am saying that because that is the most premitive concept on how to make machines intelligent as humans) on first order predicate logic. The notations are easy but to re-write a sentence in FOPL is challanging. FoPL are mainly used to infer something from the given sets of knowledge (or simply knowledge base). We, humans, can easily understand the meaning of "child" and "parent", but we cannot simply add this concept to the machine's knowledge base. We have to translate the meaning of parent and child into machine "understandable" form first. And FOPL provides a very simple notations to convert a human readable sentence to a machine readable one. If anybody wants to go into the detail of FOPL, I'll suggest to go through the books "Artificial Intelligence" by Rich and Knight, and "Artificial Intelligence" by Russell and Norvig.
There was a brief talk on XML as well. According to w3cschools.com, "XML was designed to transport and store data, with focus on what data is; and HTML was designed to display data, with focus on how data looks". So, from the concept of ongology and the objective of XML, we can deal with the limitations of syntax based approach. We have to take a detailed course on XML, ontologies and semantic webs to learn how to implement these concepts together. I recommend the link: http://www.w3schools.com/xmL/xml_whatis.asp if you want to know (more) about XML.
The second lecture by Dr. Ji was the second part of machine learning series. He started from the review of
classification. He then explained the other important part of machine learning which is clustering. The importance of clustering is that it does not require a supervisor. Data sets are grouped together according to their features or values. The main disadvantage of clustering is that we don't know whether the results after clustering is good or bad (google's search result is an example). However, if we don't know the decision boundary and we don't know any guidance to create a boundary, then clustering is the best techniue to train a machine. He went further into the basic classes of clustering: fast clustering and hierarchical clustering. Regression and semi-supervised learning concepts were covered after that. The examples provided in the slides were helpful, but to understand each topic will be very difficult. For now, I'm sticking with K-nearest neighbour and K-means types as per his suggestion. :)
Posted by
Prabal Khanal
Lectures this week
Content:
The Lecture by Dr.Fridsma gave a good understanding on the differences between semantic web and syntactic web, importance of semantic web and applicability of ontologies in the development of semantic web.
Semantic Web is an evolving development of the World Wide Web in which the meaning (semantics) of information and services on the web is defined. The semantic web comprises a set of design principles, collaborative working groups, and a variety of enabling technologies. Some elements of the semantic web include Resource Description Framework (RDF), a variety of data interchange formats (e.g. RDF/XML, N3, Turtle, N-Triples), and notations such as RDF Schema (RDFS) and the Web Ontology Language (OWL). These elements provide a formal description of concepts, terms, and relationships within a given knowledge domain. On the other hand, Syntactic Web is a phrase that describes the current, mostly HTML-based World Wide Web. In a syntactic web, if the creators of the site ever decide to change around the layout or HTML of the site, the computer program would most likely need to be rewritten in some way. In contrast, if the data is presented semantically, the program could retrieve that semantic data, and the site's creators could change the look and feel of the site without affecting that retrieval ability.
Shuiwang’s lecture mainly focused on various types of clustering. Clustering is the method by which like records are grouped together. Usually this is done to give the end user a high level view of what is going on in the database.
A simple example of clustering would be the clustering that most people perform when they do the laundry - grouping the permanent press, dry cleaning, whites and brightly colored clothes is important because they have similar characteristics. And it turns out they have important attributes in common about the way they behave (and can be ruined) in the wash. To “cluster” your laundry most of your decisions are relatively straightforward. There are of course difficult decisions to be made about which cluster your white shirt with red stripes goes into (since it is mostly white but has some color and is permanent press). When clustering is used in business the clusters are often much more dynamic - even changing weekly to monthly and many more of the decisions concerning which cluster a record falls into can be difficult.
There are two main types of clustering techniques, those that create a hierarchy of clusters and those that do not. The hierarchical clustering techniques create a hierarchy of clusters from small to big. This hierarchy of clusters is created through the algorithm that builds the clusters. There are two main types of hierarchical clustering algorithms:
• Agglomerative - Agglomerative clustering techniques start with as many clusters as there are records where each cluster contains just one record. The clusters that are nearest each other are merged together to form the next largest cluster. This merging is continued until a hierarchy of clusters is built with just a single cluster containing all the records at the top of the hierarchy.
• Divisive - Divisive clustering techniques take the opposite approach from agglomerative techniques. These techniques start with all the records in one cluster and then try to split that cluster into smaller pieces and then in turn to try to split those smaller pieces.
Posted by
Harsha Undapalli
The Lecture by Dr.Fridsma gave a good understanding on the differences between semantic web and syntactic web, importance of semantic web and applicability of ontologies in the development of semantic web.
Semantic Web is an evolving development of the World Wide Web in which the meaning (semantics) of information and services on the web is defined. The semantic web comprises a set of design principles, collaborative working groups, and a variety of enabling technologies. Some elements of the semantic web include Resource Description Framework (RDF), a variety of data interchange formats (e.g. RDF/XML, N3, Turtle, N-Triples), and notations such as RDF Schema (RDFS) and the Web Ontology Language (OWL). These elements provide a formal description of concepts, terms, and relationships within a given knowledge domain. On the other hand, Syntactic Web is a phrase that describes the current, mostly HTML-based World Wide Web. In a syntactic web, if the creators of the site ever decide to change around the layout or HTML of the site, the computer program would most likely need to be rewritten in some way. In contrast, if the data is presented semantically, the program could retrieve that semantic data, and the site's creators could change the look and feel of the site without affecting that retrieval ability.
Shuiwang’s lecture mainly focused on various types of clustering. Clustering is the method by which like records are grouped together. Usually this is done to give the end user a high level view of what is going on in the database.
A simple example of clustering would be the clustering that most people perform when they do the laundry - grouping the permanent press, dry cleaning, whites and brightly colored clothes is important because they have similar characteristics. And it turns out they have important attributes in common about the way they behave (and can be ruined) in the wash. To “cluster” your laundry most of your decisions are relatively straightforward. There are of course difficult decisions to be made about which cluster your white shirt with red stripes goes into (since it is mostly white but has some color and is permanent press). When clustering is used in business the clusters are often much more dynamic - even changing weekly to monthly and many more of the decisions concerning which cluster a record falls into can be difficult.
There are two main types of clustering techniques, those that create a hierarchy of clusters and those that do not. The hierarchical clustering techniques create a hierarchy of clusters from small to big. This hierarchy of clusters is created through the algorithm that builds the clusters. There are two main types of hierarchical clustering algorithms:
• Agglomerative - Agglomerative clustering techniques start with as many clusters as there are records where each cluster contains just one record. The clusters that are nearest each other are merged together to form the next largest cluster. This merging is continued until a hierarchy of clusters is built with just a single cluster containing all the records at the top of the hierarchy.
• Divisive - Divisive clustering techniques take the opposite approach from agglomerative techniques. These techniques start with all the records in one cluster and then try to split that cluster into smaller pieces and then in turn to try to split those smaller pieces.
Posted by
Harsha Undapalli
Week 9/28-30
Content:
This week was very informative. One one hand we learned about the approach from syntax level to semeantics level and the and the need of definite vocabulary(ontology) for indexing, interworking and sharing of complex information in a well organised manner and on the other hand we learned more about clustering and regression, the fundamental domains of data mining.Dr. Fridsma lecture started with talk on transition from syntactic web to semantic web, the make up and applications and also its limitations.The usage of ontology in semantics will not only have better interoperability and correct functionallity but would also be more robust and less ambiguious.He later discussed the usage of ontologies in information system and the advantages from it.On a whole i got a clear picture that why do we need standards in clinical science but also computers in order to have a better interoperability among the organizations.
I agree with all that Mr.Shuiwang lecture was much clearer and more understandable.The lecture was mainly focussed on supervised and unsupervised learning and also k nearest neighbour and k means. He told aproper definition and also mentioned difference between them and also used general examples(like that of students in the class).The demos that he presented helped to understand the concept in a better way. I think that the refernces that he provided would be really helpful espicially the book on introduction to Data Mining.
Posted by
Ashutosh Singraur
This week was very informative. One one hand we learned about the approach from syntax level to semeantics level and the and the need of definite vocabulary(ontology) for indexing, interworking and sharing of complex information in a well organised manner and on the other hand we learned more about clustering and regression, the fundamental domains of data mining.Dr. Fridsma lecture started with talk on transition from syntactic web to semantic web, the make up and applications and also its limitations.The usage of ontology in semantics will not only have better interoperability and correct functionallity but would also be more robust and less ambiguious.He later discussed the usage of ontologies in information system and the advantages from it.On a whole i got a clear picture that why do we need standards in clinical science but also computers in order to have a better interoperability among the organizations.
I agree with all that Mr.Shuiwang lecture was much clearer and more understandable.The lecture was mainly focussed on supervised and unsupervised learning and also k nearest neighbour and k means. He told aproper definition and also mentioned difference between them and also used general examples(like that of students in the class).The demos that he presented helped to understand the concept in a better way. I think that the refernces that he provided would be really helpful espicially the book on introduction to Data Mining.
Posted by
Ashutosh Singraur
Week of 9/28
Content: I have to admit that both lectures given this week were much clearer than the first time. On Monday, Dr. Fridsma have a lecture on the big O, ontologies and the semantic web. Dr. Fridsma gave an overview of the hype of semantic web and he told us why enumeration doesn't work. I really like how he explained that ontologies are engineering artifacts.
On Wednesday, Mr. Shuiwang lectured on machine learning. Again, the second time was much clearer than the first time. Shuiwang really emphasised that we know the difference between k-nearest neighbor and k-means. Where k-means is clustering. K-means is usually used in image processing. I really liked the online demos that he showed us. I found an interesting paper that talked about how k-means is used in gene expression data. http://www.biomedcentral.com.ezproxy1.lib.asu.edu/1471-2105/5/172
Posted by
Posted by
Thursday, October 1, 2009
Week 9-28 - Ontologies and Machine Learning
These two lectures this week were packed full of information. Both had subjects that could be taught easily over several class periods but I understand this course is about getting exposure to the various methods and we can't go into depth. I think Dr. Fridsma stated it best in the town hall meeting --- in 502, you are exposed to a lot of new information but to fully understand it you have to spend hours researching to grasp the understanding. The content being taught is synonymous to learning a new language.
The syntactic versus semantic discussion made sense and the lecture for ontologies made a little more sense than the first round. Probably the best slide illustrating the differences in semantic and syntactic was slide # 30 with the following statement: "The ability of multiple systems to exchange information and to be able to use the information that has been exchanged." From a healthcare perspective, this is what decision support is all about. We need to pull information from disparate system together use the information to evaluate trends and suggest alternative paths if necessary.
Machine learning is interesting for data mining. This is another which appears to be very complex but hugely important in data analysis. This lecture, in my opinion, was a little better to understand. There were concrete examples and nice tutorials to show how certain models worked and the instructor helped by restating the concepts several times. I understand the supervised versus unsupervised and the K nearest versus the K means. Other than that, the other models made sense but I'm having a hard time trying to understand how these are applied in real situations. It would be nice to take concrete data sets and run them through a few of these models in class. The student grouping for the clustering example is a good example of how to illustrate the different models but still very hard to keep it all straight.
These as well as other lectures have been full of information but it would be nice to apply some of what we learn into some real applications during the class instead of having just lectures without any hands on or real-world application.
Posted by : Debbie Carter
The syntactic versus semantic discussion made sense and the lecture for ontologies made a little more sense than the first round. Probably the best slide illustrating the differences in semantic and syntactic was slide # 30 with the following statement: "The ability of multiple systems to exchange information and to be able to use the information that has been exchanged." From a healthcare perspective, this is what decision support is all about. We need to pull information from disparate system together use the information to evaluate trends and suggest alternative paths if necessary.
Machine learning is interesting for data mining. This is another which appears to be very complex but hugely important in data analysis. This lecture, in my opinion, was a little better to understand. There were concrete examples and nice tutorials to show how certain models worked and the instructor helped by restating the concepts several times. I understand the supervised versus unsupervised and the K nearest versus the K means. Other than that, the other models made sense but I'm having a hard time trying to understand how these are applied in real situations. It would be nice to take concrete data sets and run them through a few of these models in class. The student grouping for the clustering example is a good example of how to illustrate the different models but still very hard to keep it all straight.
These as well as other lectures have been full of information but it would be nice to apply some of what we learn into some real applications during the class instead of having just lectures without any hands on or real-world application.
Posted by : Debbie Carter
9/28/09-9/30/09
First lecture of the week by Dr. Fridsma. I'm finally getting a sense that this is all interconnected. His discussion had much focus on the syntactic web (the web as we know it; HTML based) and the semantic web (machine language-where information can be systematically read and understood by machines). We again heard mention of SNOMED, and its likely inevitability for medical ontology.
http://linkinghub.elsevier.com.ezproxy1.lib.asu.edu/retrieve/pii/S1386-5056(08)00091-9
Mr. Shuiwang's presentation was a bit clearer this week. Perhaps repeated exposure and linkage of material is finally working on me! His focus was primarily on supervised and unsupervised learning. Both are algorithm types used in machine learning. Unsupervised learning doesn't use labels and seeks to explain how data is organized. Clustering is a type of unsupervised learning, and assigns observations into clusters, all of which have something in common. K-means is a type of flat clustering. Supervised learning considers model relationships and is a technique for creating a function from training data. K-closest neighbor is of this type.
Back to reading! Lee B.
First lecture of the week by Dr. Fridsma. I'm finally getting a sense that this is all interconnected. His discussion had much focus on the syntactic web (the web as we know it; HTML based) and the semantic web (machine language-where information can be systematically read and understood by machines). We again heard mention of SNOMED, and its likely inevitability for medical ontology.
http://linkinghub.elsevier.com.ezproxy1.lib.asu.edu/retrieve/pii/S1386-5056(08)00091-9
Mr. Shuiwang's presentation was a bit clearer this week. Perhaps repeated exposure and linkage of material is finally working on me! His focus was primarily on supervised and unsupervised learning. Both are algorithm types used in machine learning. Unsupervised learning doesn't use labels and seeks to explain how data is organized. Clustering is a type of unsupervised learning, and assigns observations into clusters, all of which have something in common. K-means is a type of flat clustering. Supervised learning considers model relationships and is a technique for creating a function from training data. K-closest neighbor is of this type.
Back to reading! Lee B.
Monday, September 28, 2009
Classifying Study Designs
Content: Dr. Petitti gave us a "compendium" of study designs in the field of epidemiology and medicine. It was interesting to learn about the different types and recognize the difference between them. At first glance I would have put them into one of two groups, observational or experimental, and I would have left it at that. In my past those are the only two groups that I have been taught. It was interesting to see the breakdown of these two groups into more distinct and specific groups. Again I liked the way I was able to see connections with past lectures, specifically Dr. Greenes lecture on August 31st on research methods. He concentrated more on what is and what is not research, but he also classified research designs and introduced us to case control studies, and randomized controlled trials.
Posted by Laura Wojtulewicz
Posted by Laura Wojtulewicz
Saturday, September 26, 2009
Data mining in BMI
Content:
The two classes of BMI 502 in this week cover the two topics in biomedical informatics study from quantitative perspectives, one is the data mining technology used in computer science, and the other is the technology used in statistics study.
The lecturer from computer science gave us a general introduction to the data mining, and introduce the usually used the data mining methods, which include the classification, cluster, association rule discovery, and regression. From this class, I understand that the data mining is a really inter-disciplinary subject, which covers the statistics and computer science. Actually, the principle of data mining is rooted in the statistics, and implemetation and developing of data mining rely on the computer science study. The principle and basic algorithm of the above methods were introduced briefly, and the practical applications of these method in our daily life are also discussed in some simple examples. Further more, the more detailed discussion in this class focused on the method on classification. The decision tree classification, rule based classification, nearest-neighbor classifiers, and artificial neural network methods are mentioned in this class. Although no further detialed discussion was around each methods in classification, the introduction also initiates my interesting to dig in deeper to learn about how these methods were realized and the particular advantages and disadvantages of these methods.
Because the classification has broad applications in artificial intelligence, and especially can be used in decision making procedure, so I think the classification is close related to the clinical decision making support study in biomedical informatics field. For example, for the traditional clinical diagnostic procedure conducted by clinical professionals, the decision making procedure is actually very like a decision tree classifing procedure. The clinical professional classify the disease of a patient from the symptoms of the patient, and think about if... then the patient probably is ..., and further from another symptom, the if... then... procedure is performed again, until the last conclusion can be obtained. This procedure is actually the same procedure of the decision tree classification. Therefore, the computer can also be trained to learn the rules to build the classification tree model, and based on this model to provide decisioin support in clinical practice.
There is one book about data mining is very classic for get a further idea about data mining method.
Introduction to data mining. Author: Pang-Ning Tan, Michael Steinbach, Vipin Kumar.
Hope it can be useful
Posted by Di Pan
The two classes of BMI 502 in this week cover the two topics in biomedical informatics study from quantitative perspectives, one is the data mining technology used in computer science, and the other is the technology used in statistics study.
The lecturer from computer science gave us a general introduction to the data mining, and introduce the usually used the data mining methods, which include the classification, cluster, association rule discovery, and regression. From this class, I understand that the data mining is a really inter-disciplinary subject, which covers the statistics and computer science. Actually, the principle of data mining is rooted in the statistics, and implemetation and developing of data mining rely on the computer science study. The principle and basic algorithm of the above methods were introduced briefly, and the practical applications of these method in our daily life are also discussed in some simple examples. Further more, the more detailed discussion in this class focused on the method on classification. The decision tree classification, rule based classification, nearest-neighbor classifiers, and artificial neural network methods are mentioned in this class. Although no further detialed discussion was around each methods in classification, the introduction also initiates my interesting to dig in deeper to learn about how these methods were realized and the particular advantages and disadvantages of these methods.
Because the classification has broad applications in artificial intelligence, and especially can be used in decision making procedure, so I think the classification is close related to the clinical decision making support study in biomedical informatics field. For example, for the traditional clinical diagnostic procedure conducted by clinical professionals, the decision making procedure is actually very like a decision tree classifing procedure. The clinical professional classify the disease of a patient from the symptoms of the patient, and think about if... then the patient probably is ..., and further from another symptom, the if... then... procedure is performed again, until the last conclusion can be obtained. This procedure is actually the same procedure of the decision tree classification. Therefore, the computer can also be trained to learn the rules to build the classification tree model, and based on this model to provide decisioin support in clinical practice.
There is one book about data mining is very classic for get a further idea about data mining method.
Introduction to data mining. Author: Pang-Ning Tan, Michael Steinbach, Vipin Kumar.
Hope it can be useful
Posted by Di Pan
Week of 09/21/09
Content:
The first lecture was on machine learning. The difference between supervised and unsupervised learning by machines was covered. Another topic that was taught was a few tasks that machines use for learning. Those tasks are classification, clustering, regression, and semi-supervised learning. Classification can be performed by using a model for a class attribute. That model is created by using a record training set to find a class attribute as a function of the values of other attributes. Some applications and algorithms for classifications were explained. A definition of clustering and some applications of clustering were instructed. Machine learning seems like a useful but complicated topic. At what point to set k values in the knn classification method seems complicated. That complication is from finding a balance between avoiding having that k value be sensitive to noise points and including points from other classes in its neighborhood. I think the presenter did a nice job at including a diverse number of applications of machine learning in the lecture.
The second lecture covered a variety of study designs. The importance of knowing about an assortment of study designs was taught. Knowledge of study designs can help with interpreting studies and the statistics they contain. That knowledge can also help a researcher to select a study and understand its practicality. Some key study design dimensions were instructed. Those dimensions are purpose, a questionnaire's effect on subjects, the subjects' and researchers' views of time, and feasibility. Classifications of study designs were also explained in the second lecture. In epidemiology and medicine study designs can be classified as descriptive, observational, or experimental. In behavioral science and evaluation research study designs can be classified as experimental or quasi-experimental. Additionally, subtypes and examples of the studies were taught. The study design lecture was interesting to me. Using interrupted time series in quasi-experimental design seems like a particularly helpful type of study design to know about. Researchers using a interrupted time series timeline can avoid spending their resources to collect middle data in a timeline if that data is not useful to them. That could be a valuable way to save resources in a study.
The Journal of Machine Learning Research offers a large variety of articles on machine learning. That journal's site can help anyone who is trying to gain an indepth understanding of machine learning. The site provides an article search engine for anyone who is interested in specific machine learning topics. The link for the journal's site is: http://jmlr.csail.mit.edu/ .
Posted by:
Nate
The first lecture was on machine learning. The difference between supervised and unsupervised learning by machines was covered. Another topic that was taught was a few tasks that machines use for learning. Those tasks are classification, clustering, regression, and semi-supervised learning. Classification can be performed by using a model for a class attribute. That model is created by using a record training set to find a class attribute as a function of the values of other attributes. Some applications and algorithms for classifications were explained. A definition of clustering and some applications of clustering were instructed. Machine learning seems like a useful but complicated topic. At what point to set k values in the knn classification method seems complicated. That complication is from finding a balance between avoiding having that k value be sensitive to noise points and including points from other classes in its neighborhood. I think the presenter did a nice job at including a diverse number of applications of machine learning in the lecture.
The second lecture covered a variety of study designs. The importance of knowing about an assortment of study designs was taught. Knowledge of study designs can help with interpreting studies and the statistics they contain. That knowledge can also help a researcher to select a study and understand its practicality. Some key study design dimensions were instructed. Those dimensions are purpose, a questionnaire's effect on subjects, the subjects' and researchers' views of time, and feasibility. Classifications of study designs were also explained in the second lecture. In epidemiology and medicine study designs can be classified as descriptive, observational, or experimental. In behavioral science and evaluation research study designs can be classified as experimental or quasi-experimental. Additionally, subtypes and examples of the studies were taught. The study design lecture was interesting to me. Using interrupted time series in quasi-experimental design seems like a particularly helpful type of study design to know about. Researchers using a interrupted time series timeline can avoid spending their resources to collect middle data in a timeline if that data is not useful to them. That could be a valuable way to save resources in a study.
The Journal of Machine Learning Research offers a large variety of articles on machine learning. That journal's site can help anyone who is trying to gain an indepth understanding of machine learning. The site provides an article search engine for anyone who is interested in specific machine learning topics. The link for the journal's site is: http://jmlr.csail.mit.edu/ .
Posted by:
Nate
Friday, September 25, 2009
Research design
Content:
This Mednesday, Dr. Petitti introduced us classification of research design in epidemiology research. There are three types of epidemiology research--descriptive studies, observational studies and experimental studies. Descriptive studies aim at description, the goal of experimental studies is causual inference, while observational study can study causual relationship or not.
Observation studies can be devided in to cross-sectional studies, cohort studies and case-control studies. The difference between cross-sectional studies and cohort studies is that the former one measures prevalence while the latter one measures incidence. In my opinion, cross-sectional studies are more close to descriptive studies. Usually, cross-sectional studies just state a situation rather than analysis the reason of the situation. The difference between cohort studies and case-control studies is that cohort studies need researchers to on-going monitor the subject which expose to risk factor to ascertain the outcomes, while conducting case-control studies, researchers do not need to wait, what they need to do is to compare cases with controls to find out the causaul relationship between risk factor and outcome.
Experimental studies are very similar to cohort studies, the only difference is that in a experimental study, researchers manipulate variables, while in cohort studies, researchers just wait. Experimental studies can be devided into Post only design and Pre and Post design. The difference between them is that the latter add information of baseline to assure adequacy of randomization and comparability at baseline.
I like the way Dr. Petitti giving her lecture that she compared different types of studies to help us understand the trait of each research design.
Posted by
Jing Lu
This Mednesday, Dr. Petitti introduced us classification of research design in epidemiology research. There are three types of epidemiology research--descriptive studies, observational studies and experimental studies. Descriptive studies aim at description, the goal of experimental studies is causual inference, while observational study can study causual relationship or not.
Observation studies can be devided in to cross-sectional studies, cohort studies and case-control studies. The difference between cross-sectional studies and cohort studies is that the former one measures prevalence while the latter one measures incidence. In my opinion, cross-sectional studies are more close to descriptive studies. Usually, cross-sectional studies just state a situation rather than analysis the reason of the situation. The difference between cohort studies and case-control studies is that cohort studies need researchers to on-going monitor the subject which expose to risk factor to ascertain the outcomes, while conducting case-control studies, researchers do not need to wait, what they need to do is to compare cases with controls to find out the causaul relationship between risk factor and outcome.
Experimental studies are very similar to cohort studies, the only difference is that in a experimental study, researchers manipulate variables, while in cohort studies, researchers just wait. Experimental studies can be devided into Post only design and Pre and Post design. The difference between them is that the latter add information of baseline to assure adequacy of randomization and comparability at baseline.
I like the way Dr. Petitti giving her lecture that she compared different types of studies to help us understand the trait of each research design.
Posted by
Jing Lu
Week Five: Machine Learning & Study Design
Content:
Posted by Eric
Machine Learning-
At first some of the machine learning lecture just flew over my head but after reviewing the lectures and the comments of colleagues, it looks like the lecture finally arrived. To summarize quickly since there is extensive commenting on Machine Learning in the previous posts, machine learning is a field where we make or give a machine the ability to learn. And this is accomplished through several methods that either classify or cluster. Classification methods includes k-nearest neighbor, decision trees, and support vectors. As Xiaoxiao mentions, Mr. Ji mentioned google as a cluster example and I specifically remember stumbling across an article/webpage that actually describes the algorithms of google as well as talks about how google clusters its data. I will have to review my BME Capstone documentation for the link and I will post it in the comments when I find it.
Study Design-
I found this lecture on study design very interesting since she covered the different types of studies that the different fields conduct (even though the studies may be similar they are called different things in different fields). I especially liked the classic studies that she presented since like she mentioned, these studies are older and are much simpler. They provide a very good basis for understanding the different types of studies. What I liked most was the study on salt (intersalt i believe). It was interesting how the study when it looked at individuals, did not find a correlation between sodium intake and hypertension but when it was performed like it was then the points all lined up and provided a correlation that indicated that salt correlated with blood pressure.
Posted by Eric
Study Design Overview
Dr. Petitti introduced the topic of study designs in such a way clearly outlined the purpose of each type. As with programming, the hardest part of research is actually evaluating the problem and creating a design. The hard part is never the implementation and testing. Creating the idea and coming up with the appropriate steps to answer the problem in a consistent way always seems to take more effort. This lecture defined the characteristics of each type of study design, including causality, time frame, and point of view. I really liked the fact that specific study examples were provided with each type of design. This really caused me to take a close look at what separates each type. Some studies seemed similar in nature, but once we broke down the pieces of the design , they followed the rules that we were given and fit in the appropriate category. I found myself thinking of my days involved in clinical trials versus my work in post-marketing surveillance and how these two environments contained similar types of data but were structured so differently based on the nature of each environment. The causality and time frame really define the type of study in this case. However, not all designs are this clear cut. I will be taking a closer look at studies in articles and why the researchers chose that particular design.
Posted by Annie
Compendium of Study Designs Review
Dr. Pettiti lectures on Compendiums of Study Designs was helpful in understanding the different purposes of studies. She was a good presenter making the information on different study designs pertitnent and easily understood. Dr. Pettiti included some well known “gold standard” studies to help us understand the differences. As a way to help me understand and remember the difference, I have tried to write my understanding of each of the study types:
• Descriptive studies – majority of medical studies fall in this category.
• Quasi-Experimental – used when randomization of the group is not possible.
Posted by : Debbie Carter
• Descriptive studies – majority of medical studies fall in this category.
- Case – discussing a new or unusual finding about one thing
- Case Series – discussing something unusual or common several cases
- Surveillance – on going observation and monitoring over a period of time
- Cross-sectional – studying two different findings at one point in time or across the same population (determines prevalence)
- Cohort – studying findings from now into a time of the future (determines incidence)
- Case-control – comparing two groups of population – one with a particular condition versus a group without the condition
- Ecologic – study of populations and epidemiology and the effects of environmental exposures.
• Quasi-Experimental – used when randomization of the group is not possible.
I attended a conference on Thursday where the presenter was giving data about risk factors for DVT, prevalence with different risk factors and new requirements coming from the government and found myself thinking of these different study types and how they came to these conclusions, what types of studies were performed, how they determined prevalence, probabilities, etc. It's a lot of information to process but at least some of this is getting applied in my day to day work.
Machine Learning and Study Design
One of my teachers once said that computers can solve problems which humans think difficult, but computers are not able to do things that humans can do easily. For example, if we try to calculate 459287 x 979435 by hand, it might take a long time, whereas computers can solve it in less than a second. What if we want a computer to do an easy task, let's say recognize a person? It will take a long time (if it can give solution), but humans can do those recognition thing comfortably. The basic difference between man and machine is the same thing: humans can recognize something (that's need parallel processing) easily but can't perform serial computation where as machine is better at performing serial computations like arithmatic calculations, but they are very bad at parallel processing (must say the designers of computers aren't smart enough to design something like that). Human can classify and infer something, machines cannot. So, the basic difference lies on learning. Humans can learn and store those things in memory for a long time and are able to infer new things. Computers need some technique to "learn" and the procedure that we apply to train a computer in norder to perform some tasks in the future is called "Machine Learning".
There are mainly two types of learning: supervised and unsupervised. In supervised learning, there is something that will "guide" a computer to give the best result. In unsupervised learning, there is no such thing like "guiding". According to the lecture, there are mainly four major topics in machine learning: classification, clusturing, regression, and semi-supervised learning.In classification, we would want to classify things (anything: yes/no, good/bad, healthy/unhealthy and so on). We use several technique to train a computer to classify those things. We first create a model using training data. The model, which is designed based on the training data, is then used to classify new data (test data). In case of surgical training simulator, if a surgeon performs a task, we would want to classify his performance as good or bad (let's not consider fuzzy answers). To evaluate the result, we need to check the results of previous similar cases and notice some key parameters like how it's done, how long did it take etc. Based on the majority of the results, we can classify the performance of the new surgeon. The results obtained from previous surgery cases are referred to as "training set", and the one that we wanted to classify is "test set". There are so many tools that can be used for classification. Few of them, which are mentioned in the class are: k-nearest neighbor, neural networks (artificial), naive bayes classifier, svm etc. Naive bayes classifiers are used in spam filtering purposes (like in spam-assasin). SVM can be used to classify non linear classification problems (like XOR-gates). Neural Networks are mainly used in computer vision to train a model to recognize some parts in images. Classification falls under supervised learning category.
In unsupervised learning, clustering is one of the popular techniques.In this technique, we take observations and put into subsets(clusters) in a way that the observations are similar in some sense (wiki). We didn't go much in detail during the lecture. It would be interesting to know in detail about clustering and semi-supervised learning in his next class.
The next lecture by Dr. Petiti was on study design, which I found very informative. She talked about various design techniques and how we can manipulate the information in an experiment. We went in detail of descriptive, observational, experimental and quasi-experimental studies. The main thing that I understood from the lecture was that experiments are randomized. In fact, experiments mean randomization. If experiments are not randomized, we cannot trust the results from those experiments. But, randomized experiments are difficult to do for a few reasons. The examples given in the lecture slides were practical reason ("we can't randomize smoking") and ethical reason("can't randomize cocaiine use). The classic studies presented during the lecture were interesting. For each of design study, there was a classic study, and they really made things easier to understand.
Posted by
Prabal Khanal
There are mainly two types of learning: supervised and unsupervised. In supervised learning, there is something that will "guide" a computer to give the best result. In unsupervised learning, there is no such thing like "guiding". According to the lecture, there are mainly four major topics in machine learning: classification, clusturing, regression, and semi-supervised learning.In classification, we would want to classify things (anything: yes/no, good/bad, healthy/unhealthy and so on). We use several technique to train a computer to classify those things. We first create a model using training data. The model, which is designed based on the training data, is then used to classify new data (test data). In case of surgical training simulator, if a surgeon performs a task, we would want to classify his performance as good or bad (let's not consider fuzzy answers). To evaluate the result, we need to check the results of previous similar cases and notice some key parameters like how it's done, how long did it take etc. Based on the majority of the results, we can classify the performance of the new surgeon. The results obtained from previous surgery cases are referred to as "training set", and the one that we wanted to classify is "test set". There are so many tools that can be used for classification. Few of them, which are mentioned in the class are: k-nearest neighbor, neural networks (artificial), naive bayes classifier, svm etc. Naive bayes classifiers are used in spam filtering purposes (like in spam-assasin). SVM can be used to classify non linear classification problems (like XOR-gates). Neural Networks are mainly used in computer vision to train a model to recognize some parts in images. Classification falls under supervised learning category.
In unsupervised learning, clustering is one of the popular techniques.In this technique, we take observations and put into subsets(clusters) in a way that the observations are similar in some sense (wiki). We didn't go much in detail during the lecture. It would be interesting to know in detail about clustering and semi-supervised learning in his next class.
The next lecture by Dr. Petiti was on study design, which I found very informative. She talked about various design techniques and how we can manipulate the information in an experiment. We went in detail of descriptive, observational, experimental and quasi-experimental studies. The main thing that I understood from the lecture was that experiments are randomized. In fact, experiments mean randomization. If experiments are not randomized, we cannot trust the results from those experiments. But, randomized experiments are difficult to do for a few reasons. The examples given in the lecture slides were practical reason ("we can't randomize smoking") and ethical reason("can't randomize cocaiine use). The classic studies presented during the lecture were interesting. For each of design study, there was a classic study, and they really made things easier to understand.
Posted by
Prabal Khanal
Subscribe to:
Posts (Atom)