Video tutorial on reinforcement learning

This video tutorial is based on a talk I gave at my Comp 650 class and hence assumes that the audience knows a little about Markov Decision Processes and Partially Observable Markov Decision Processes. If you are unfamiliar with them, I recommend that you at least read through their wikipedia pages before going through this tutorial. Also, I suggest that you view this video on youtube, because the embedded video given below is cropped.


If you have any suggestions, comments or questions, please feel free to email me :)

PDF slides

Cross-validation on SVM-Light

Before reading this blogpost you should read: Practical guide to libSVM, Wikipedia entry on SVM, Wikipedia entry on cross-validation, and PRML - Bishop's chapter on SVM.  In this post, I will not explain how to use open source SVM tools (like libsvm or svmlight), but would rather assume that you have used them and want some clever script to automate them. Also note that, the below script uses default parameters. So if you get poor results, you can add an outer for loop that iterates through your choice of parameters and supplies them to the program ./svm_learn.

To perform cross-validation on libsvm:

1. Download and extract it
2. You need gnuplot installed on the machine you want to run libsvm. If you want to run it on the server, do 'ssh -X myuserName@myMachineName'
3. make
4. Go to Tools folder and run 'python easy.py my_training_file my_test_file'

To perform cross-validation on svmlight:

1. Download and extract it. There you will find two important binaries: svm_learn and svm_classify
2. Unfortunately the data gathering using svmlight is not easy, so I have coded following script. 

#!/bin/bash
# Usage: svm_light_cross_validation full_data num_train_items num_test_items k_cycles results_file
RESULT_FILE=$5

echo "Running SVM-Light via cross validation on" $1 "by using" $2 "training items and" $3 "test items (Total number of cross-validation cycles:" $4 > $RESULT_FILE

MODEL_FILE="model."$RANDOM".txt"
TEMP_FILE="tempFile."$RANDOM".txt"
PRED_FILE="prediction."$RANDOM".txt"
DATA_FILE=$1
NUM_TRAIN=$2
NUM_TEST=$3
NUM_CYCLES=$4

TEMP_DATA_FILE=$DATA_FILE"."$RANDOM".temp"
TRAIN_FILE=$TEMP_DATA_FILE".train"
TEST_FILE=$TEMP_DATA_FILE".test"

TEMP_RESULT=$RESULT_FILE".temp"
SVM_PATTERN='s/Accuracy on test set: \([0-9]*.[0-9]*\)% ([0-9]* correct, [0-9]* incorrect, [0-9]* total)\.*/'
for k in `seq 1 $NUM_CYCLES`
do
 sort -R $DATA_FILE > $TEMP_DATA_FILE
 head -n $NUM_TRAIN $TEMP_DATA_FILE > $TRAIN_FILE
 tail -n $NUM_TEST $TEMP_DATA_FILE > $TEST_FILE
 
 echo "------------------------------------------"  >> $RESULT_FILE
 echo "Cross-validation cycle:" $k >> $RESULT_FILE
 
 # first run svm with default parameters
 echo ""  >> $RESULT_FILE
 echo "Polynomial SVM with default parameters" >> $RESULT_FILE
 for i in 1 2 3 4 5 6 7 8 9 10
 do
  echo "order:" $i >> $RESULT_FILE
  ./svm_learn -t 1 -d $i $TRAIN_FILE $MODEL_FILE > $TEMP_FILE
  ./svm_classify -v 1 $TEST_FILE $MODEL_FILE $PRED_FILE > $TEMP_RESULT
  cat $TEMP_RESULT >> $RESULT_FILE
  sed '/^Reading model/d' $TEMP_RESULT > $TEMP_RESULT"1"
  sed '/^Precision/d' $TEMP_RESULT"1" > $TEMP_RESULT
  sed "$SVM_PATTERN$k poly $i \1/g" $TEMP_RESULT >> "better"$RESULT_FILE
 done

 echo ""  >> $RESULT_FILE
 echo "RBF SVM with default parameters" >> $RESULT_FILE
 for g in 0.00001 0.0001 0.001 0.1 1 2 3 5 10 20 50 100 200 500 1000
 do
  echo "gamma:" $g >> $RESULT_FILE
  ./svm_learn -t 2 -g $g $TRAIN_FILE $MODEL_FILE > $TEMP_FILE
  ./svm_classify -v 1 $TEST_FILE $MODEL_FILE $PRED_FILE >> $TEMP_FILE
  cat $TEMP_RESULT >> $RESULT_FILE
  sed '/^Reading model/d' $TEMP_RESULT > $TEMP_RESULT"1"
  sed '/^Precision/d' $TEMP_RESULT"1" > $TEMP_RESULT
  sed "$SVM_PATTERN$k rbf $g \1/g" $TEMP_RESULT >> "better"$RESULT_FILE
 done

done

rm $MODEL_FILE $TEMP_FILE $PRED_FILE $TEMP_DATA_FILE $TEMP_RESULT $TEMP_RESULT"1"
echo "Done." >> $RESULT_FILE

Here is how you run this script:

./svm_light_cross_validation myDataInSVMLightFormat.txt 5000 2000 10 MyResults.txt

myDataInSVMLightFormat.txt looks like "binaryClassifier(-1 or 1) featureNum:value ....". Example:

1 1:1.75814e-14 2:1.74821e-05 3:1.37931e-08 4:1.25827e-14 5:4.09717e-05 6:1.28084e-09 7:2.80137e-22 8:2.17821e-24 9:0.00600121 10:0.002669 11:1.19775e-05 12:8.24607e-15 13:8.36358e-10 14:0.0532899 15:3.25028e-06 16:0.148439 17:9.76215e-08 18:0.00148927 19:3.69801e-16 20:4.20283e-16 21:7.9941e-05 22:5.7593e-09 23:0.000251052 24:0.000184218 25:7.07359e-06 

After running this script you will get two important files: MyResults.txt and betterMyResults.txt (which is a space separated file with format: 'validation_cycle_num [poly or rbf] parameterValue accuracy'). You can then import betterMyResults.txt in R or Matlab and get some pretty graphs :)

How to analyze algorithms ?

Important assumptions:

•  This process of analysis should be independent of programming language and machine architecture. So you can't just write your algorithm in C++ and compare it using running time of some other algorithm written in Java. 
• Also, since each algorithm can take as an input arbitrary data structure (eg: graph, array, binary object, tree, ...), there has to be some standard way to specify it. For sake of our analysis, we use input size (n) as the measure for our analysis. For eg: length of array, number of vertices or edges of the graph, number of nodes of a tree etc.

The detailed explanations of these two points is given in Introduction to Algorithms by Cormen, Leiserson, Rivest, .. (CLR)

Setup:

Input of size n ==> Algorithm that takes T(n) steps ==> some output

The number of steps in the algorithm often depends on factors other than the input size. For eg: the number of steps for quicksort for the input [1, 2, 3, 4, 5] is different than as compared to the input [5, 4, 3, 2, 1]; even though both have same input size (and in this case same elements).
Hence, we will try to find 3 functions (upper bound, lower bound and Theta bound) that can capture the effects of these factors: $$ g_1(n), g_2(n) \text{ and } g_3(n) \text{ such that } $$ $$ T(n) \in O(g_1(n)), T(n) \in \Omega(g_2(n)) \text{ and } T(n) \in \Theta(g_3(n))$$ Many authors also write above equations as: $$ T(n) = O(g_1(n)), T(n) = \Omega(g_2(n)) \text{ and } T(n) = \Theta(g_3(n))$$

Big-O notations:

Now, let's understand definition of above notations: \begin{align} \text{To prove that } & g_1(n) \text{ is upper bound of T(n), i.e.} T(n) \in O(g_1(n)) \\ \text{ find two constants } & c, n_0 \text{ such that } \\ & \forall n \geq n_0, T(n) \leq c g_1(n) \end{align} For lower bound, change ≤ sign to ≥ sign and for theta bound find both lower and upper bound: $$ T(n) \in \Theta(g_3(n)) \text{ iff } T(n) \in O(g_3(n)) \text{ and } T(n) \in \Omega(g_3(n))$$
To understand these in more details, I suggest you solve problems from CLR chapter 3. Here are some example you can start with: \begin{align} 2n^2 \in O(n^3) & \text{ with } c=1 \text{ and } n_0 = 2 \\ \sqrt{n} \in \Omega(lg n) & \text{ with } c=1 \text{ and } n_0 = 16 \\ \frac{n^2}{2} - 2n \in \Theta(n^2) & \text{ with } c_1=\frac{1}{4}, c_2=\frac{1}{2} \text{ and } n_0 = 8 \\ \end{align} (Note: theoretically very large functions such as infinity^n or n^infinity can always be lower bounds. So, it is necessary to supply as tight bound as possible for any given algorithm. Mathematically, the notations for tight upper bound is small o and for tight lower bound is small omega, but we will ignore them for sake of this discussion)

Common measures for algorithms:

Above three notations are often replaced by the terms 'worst case', 'best case' and 'average case' complexity. But, more precise definitions are as follows: \begin{align} \text{Worst case}& T(n) = max \{ T(x) | \text{ x is an instance of size n} \} \\ \text{Best case}& T(n) = min \{ T(x) | \text{x is an instance of size n} \} \\ \text{Average case}& T(n) = \sum_{\text{input }x} Prob(x) T(x) \end{align} This means, to find average case complexity of algorithm, you need to find probability of a given input and use it to find expectation of T(n).

Pseudo code to Big-O:

Now, let's address the key question: given a pseudo code, how do we find its lower bound (or upper bound or ...). Here is one possible way to find those bounds:

• First, express the algorithm in recursive form
• Determine appropriate metric for input size 'n'
• Then, find recurrence relation for the algorithm
• Solve the recurrence relation using either Substitution method, recursion tree method or the master theorem

For example, the recurrence relations for popular algorithms are: \begin{align} \text{Binary search: } & T(n) & = T(\frac{n}{2}) + O(1) \\ \text{Sequential search: } & T(n) & = T(n-1) + O(1) \\ \text{Tree traversal: } & T(n) & = 2T(\frac{n}{2}) + O(1) \\ \text{Selection sort: } & T(n) & = T(n-1) + O(n) \\ \text{Merge sort: } & T(n) & = 2T(\frac{n}{2}) + O(n) \end{align} Try writing the pseudo code for any of the above algorithm and verify the recurrence relations.

Master theorem:

Master theorem can only solve the recurrence relation of the form: T(n) = a T(n/b) + f(n). Here is my modus operandi for master theorem: \begin{align} \text{If } f(n)=n^{log_{b}a}lg^{k+1}n, & \text{ then } T(n) = \Theta(n^{log_{b}a} lg^{k+1}n) \\ \text{Else if } n^{log_{b}a} \text{ is polynomially greater than } f(n), & \text{ then } T(n) = \Theta(n^{log_{b}a}) \\ \text{Else if } f(n) \text{ is polynomially greater than } n^{log_{b}a}, & \text{ then } T(n) = \Theta(f(n)) \\ \text{Else use Substitution method}& \end{align}
Here are solutions to two important recurrence relations, that cannot be solved using Master theorem: \begin{align} \text{Recurrence relation} & & \text{Complexity} \\ T(n) = T(n-1) + b n^k & => & O(n^{k+1})\\ T(n) = cT(n-1) + b n^k & => & O(c^n) \end{align}
For recurrence relation of other forms, refer to CLR for substitution and recursion tree method.
(In short, for substitution method, you guess the solution and use induction to find the constants and to prove that the guessed solution is correct.
Recursion tree method are often used to find the guess for the substitution method)

Some useful approximations:

For non-negative real b,c,d:
\begin{align} \text{Polynomial growth rate:}& \sum_{i=1}^n i^c &= \Theta(n^{c+1}) \\ \text{Logarithmic growth rate:}& \sum_{i=1}^n \dfrac{1}{i} &= \Theta(log(n)) \\ \text{Exponential growth rate:}& \sum_{i=1}^n c^i &= \Theta(c^n) \\ \text{n-Log(n) growth rate:}& \sum_{i=1}^n log(i)^c &= \Theta(n log(n)^c) \\ \text{Combination of above:}& \sum_{i=1}^n log(i)^c i^{d} &= \Theta(n^{d+1} log(n)^c) \\ &\sum_{i=1}^n log(i)^c i^d b^{i} &= \Theta(n^{d+1} log(n)^c b^{n}) \\ \text{Stirling's approximation:}& n! &= \sqrt(2 \pi n) (\frac{n}{e})^n (1 + \Theta(\frac{1}{n})) \end{align}

Presentation on Online Aggregation for Large MapReduce Jobs

This post has two purposes:
1. Give information about my Comp 600 presentation at Rice University
2. Enumerate the lesson's learnt while preparing and giving the presentation.

(If the above videos does not work, click on: Dailymotion or Vimeo)

Also, if you have time, go through my slides or better still read my paper :)
- Slides
- Paper

---

FAQs during the presentation at VLDB:

1. What if the blocks are equally sized and spread across uniformly ?
- We used exactly that configuration. The input blocks of wikipedia logs are approximately of same size and HDFS does good job of spreading the data across uniformly. Still, we found biased results in the beginning.

2. Does your system replace general database ?
- Nope. As of now, it only performs single table aggregation with group-by. For a general online database, see DBO paper by our group.

3. Will you release your code ?
- Yes, we intend to release our code as open-source. For current version, see http://code.google.com/p/hyracks-online-aggregation/

4. Who do you see as potential user of the system ?
- Any company that has massive data stored in HDFS (logically as single table) and want to perform aggregation in online fashion or only till certain accuracy. Note, that the table containing the data need not be pre-randomized.

5. What does user of the system needs to know ?
- Just how to write and run a hadoop job ... All he/she has to do is write a hadoop mapper and set appropriate parameters in job configuration. For details see appendix of our paper.

6. You didn't explain too much statistics in the talk ... I want to know more
- Yes, you are absolutely right. Infact, that was intentional :) ... If I have to give one-line summary of statistics here is it: Sample from a Multi-variate normal on value, processing time and scheduling time (which have correlation factor associated with between each of them) which takes in account the 'lower bound' information of processing time of the blocks that are currently been processed ... rather than simply sampling from the distribution over data.

7. You don't talk about performance ... Isn't randomized scheduling policy bad for performance ?
- We didnt talk about the performance, mainly because we didnot provide any comparative study for performance in our paper. Note that, our policy is not same as a strict randomized policy. We expect the master to maintain a randomize queue and start the timer when a worker asks for the block. If for some reason (say locality), worker doesnot want that block or master has some internal (locality-based) scheduling algorithm, he can leave the timer ON and go for next block in the randomized queue. This is necessary machinery for sampling theory. This will yield performance that is similar to locality-based scheduling. However, in our current implementation, we have not used this mechanism. We intend to do the performance study in our future work.

---

Presentation tips:

Clearly my presentation has lot of room for improvement, for example: need to talk slower, more uniform layout of the slides, more interaction with audience etc. However, I have learnt a great deal while preparing and giving this presentation. I will summarize few points that I (and probably lot of graduate students) need to be careful about. For more detailed analysis of these points, I suggest you read Presentation Zen, Slide:ology and When the scientist present.

Step 1: Planning your presentation
1. Create a relaxing environment
- Go to a library, coffee place, garden (place where no one will disturb you)
- Take time to relax and forget about daily chores.
- Don't bring your laptop. In this step, you will create rough sketch of the presentation on paper (not slideware)
- Lot of the time, changing locations helps you to be more creative.

2. Make sure you understand the big picture of your presentation.
Can you find 30-45 seconds answer to following questions:
- What is the single most important "key" point that I want my audience to remember
- Why should that point matter to the audience

Also, make sure that your presentation title has keywords that highlights the big picture (as many people (if not all) will come to your talk either reading just the title (and abstract) or your name). This does not mean the title has to be too long; in fact shorter and thought-provoking title are preferred.

3. Now, create a sketch of the presentation. This involves enumerating key ideas you will be presenting and also organising/ordering them. You can use either or all of the following tools:
- Mindmap on whiteboard or paper
- Post-it (use sharpie to make sure you don't misuse post-it)
- Brainstorm the list of ideas + Grouping them
- Sketching or drawing 
Nancy Duarte says most effective presentations involves telling a story that audience can connect and relate to, rather than just present the information/data. Also, best professors are good story-tellers. They just don't go through the material in the book, but put their own personality, character and experience into the material in form of a narrative, which is illuminating, engaging and memorable.

The most useful point I have learnt from Presentation Zen is that the presentation consists of three parts:
- Slides the audience will see (which you will create in Step 2)

- Notes only you will see (which you have created in this Step)

- Handouts that audience can take away (so that they don't have to jot down key definition or formula). This in worst case could simply be a copy of your paper.

This essentially means that the slides are not telepromptor, that you read aloud throughout your presentation.

Step 2: Design the slides
To succeed as a presenter, you need to think as a designer (from slide:logy).

1. First few slides are very very important !!!
- I like to give use them to tell the audience the big picture of the presentation.
- It is also necessary that this has to be catch their attention, as we all know the attention-span of audience decreases exponentially after first few minutes. Don't use the first few minutes reading through the 'Outline' slide.
- Use following principle: Tell them what you are going to present, present it, then tell them again what you presented :)

2. Slides should be as simple as possible.
- Don't use fancy, distracting background images. I prefer sticking to simple white background slide.
- Good design has plenty of empty spaces.
- Avoid typical bullet point presentations. They are boring and they do more harm than good.
- Don't use 3D figures when 2D figures can suffice
- Never ever clutter the slides with useless data, words, figures, images, etc

3. 1 message per slide ... not more !!
- Do you understand that message ?
- Does your slide reflect that message ? You don't have to write down the message into your slide
- Can I make this slide simpler and still make sure that the message is conveyed. In most cases, this is done by replace bunch of text either with an appropriate image or animation.
- A lot of the time, entire message can be conveyed just by a quote from a reliable or important person.

4. Try to find a good figure or animation to convey a complex idea, algorithm or system. Here are some tips on creating animation:
- Using same slide background, create many slides (by copy operation). Change only 1 detail per slide. When you keep moving from one slide to next, it would give an effect of animation, something like flip-book. However, this process creates a lot of slides (which is problematic in many cases). So, if you want to fit everything into 1 slide, you can create a flash object, gif file or avi file and embed it into the slide. Most people use Adobe tools to create professional looking flash files, but here is poor man's solution:
- Use flip-book style slide transition technique and desktop recorder (eg: recordmydesktop) to create a movie. It usually creates movie in ogv file format.
- Also, lot of time you need to create animation using statistical distribution. In that case, use animation library of R and create ogv movie using desktop recorder. (This library also allows you store the animation in gif, but I am having some problems with it).

# R code for animation
library(animation)
ani.options(interval = 1, nmax = 20)
sd1 <- 41
for (i in 1:ani.options("nmax")) {
    x=seq(900, 1100,length=200)
    y=dnorm(x,mean=1000,sd=sd1)
    plot(x,y,type="l",lwd=2,col="red")
    ani.pause() ## pause for a while ('interval')
    sd1 <- sd1 - 2
}

a. Convert OGV to AVI:
mencoder input.ogv -ovc xvid -oac mp3lame -xvidencopts pass=1 -speed 0.5 -o output.avi
b. Convert OGV to GIF (or swf):
ffmpeg -i input.ogv -loop_output 0 -pix_fmt rgb24 -r 10 -s 320x240 output.gif
"-loop_output 0" means loop forever.

5. Is the content tailored to meet audience' expectation:
- Employers/future collaborators will try to evaluate you rather than the topic.
- Students (undergrad/first year grads/ ..) or people from different area want to understand the problem and intuition behind the solution. Hence, they are more interested in motivation section.
- Co-author will look for recognition.
- Sponsor will look for acknowledgement.
- People who work in same area will try to see if how your system/solution compares to theirs (i.e. related work) or may be are there to look for a new problem (i.e. your future work).

6. Other design tips:
- Use contrast (using font type, font color, font size, shape, color, images) accentuate the difference. (If it is different, make it very different - Garr Reynolds). It helps audience to identify the message of the slide quickly.
- Repetition: Use same image (or anchor) to represent the same concept throughout the presentation.
- Alignment: Use grids, invisible lines. Audience needs to know the order in which to process the information.
- Proximity
- Hierarchy, Unity
- Use high quality graphics. There are websites that create and sell them like www.istockphoto.com; or even give them for free like http://www.flickr.com/creativecommons, etc.
- Nothing distracts the audience more than a cluttered slide.
- Make sure your font size is greater than 24 point (so that audience can read it from the projector)
- Use comics ... it not only conveys the information efficiently, but also makes the presentation playful. Though many of my colleagues disagree with this, I believe most audience wants to have fun even if it is a research presentation .. so lighten the mood, be creative and "artsy" with your slides :)

"The more strikingly visual your presentation is, the more people will remember it. And more importantly they will remember you" - Paul Arden.
Here are some slides that are well designed:
http://www.slideshare.net/jbrenman
http://www.slideshare.net/chrislandry
http://www.slideshare.net/gkawasaki
http://www.slideshare.net/garr

Step 3: Rehearse and re-rehearse
- Use both notes and slides.
- Rehearse until you can remember and are comfortable with them. Imagine the worst case scenario that you lost your laptop, slides and all your notes ... and you have to give your presentation !!! ... If you can do that, you are ready :)
- Here you also do the job of editing, if you feel certain slides are just fillers and can be removed ... be brutal and delete them.
- Listen to http://www.ted.com/talks for motivation on how to present.

Step 4: Delivery
- Go to the presentation room 15 min before the scheduled time and make sure your slides are compatible, projector works, ... Also, bring a bottle of water, just in case.
- Leave the lights on and face the audience. People are there to see you, not read through the slides. Use wireless presenter so that you don't need to hide behind the podium/monitor.
- Slides are not telepromptor ... so don't keep staring at the monitor ... connect with the audience.

Remember these two quotes:
- Every presentation has potential to be great; every presentation is high stakes; and every audience deserves the absolute best - Nancy Duarte.
- You are presenting yourself rather than the topic. So, negative impact of poor presentation far exceeds failure of getting your point across ... Jean-Luc Lebrun

An Attempt to understand Machine Learning (Part 1)

I still remember the excitement when I took "Theory of Automata" course in my undergrad and was introduced to Turing machine. Few days after digging a little, I read Church-Turing thesis which stated "Every algorithms can be expressed in terms of Turing machine". This was like saying ... well you have framework for understand everything you will eventually learn in your Masters or PhD. However, mapping everything to Turing machine kept on becoming difficult as I started doing research and particularly when I took Machine Learning course. Most people come up with different perception when they take Machine Learning or Artificial Intelligence course.

Before I begin let me make you aware that Machine "Learning" is not same as human "learning". Humans learn something is a way that is not fully understood by scientific community. Let's say two persons X and Y see a mango tree. X has never seen a mango tree before so he will create a mental picture of it as a medium sized tree with green leaves. Y on other hand will create a mental image of tree and associate it with a mental image of mango or may be even with other senses like "it taste like mango shake". Any further discussion between X and Y will be based on their mental images and not reality. Note: information in mental images is less than information in reality and the reasons for this information loss are:
- We cannot store or process all the information in the world (due to lack of 'infinite' focus)
- Everyone filters information from reality according to their past experiences or beliefs.

Hence, since it is virtually impossible to have two person with exactly same experiences, it is impossible to have two people with exactly same mental images or information. If we treat the mental image as an output or even intermediate variable that we use to spit out an output, there is no "common" theory/algorithm that will generate a common output for all humans for same problem. Hence, this huge diversity (which makes life interesting) makes humans difficult to understand.

Most scientists (at least the neurologist/psychiatrist/psychologist) use abstraction (i.e. weed out details and rely on central theme i.e. "tree") to specify a theory/algorithm that applies to all. Yet there is another problem with human learning, humans don't learn or process with objects but with patterns. Furthermore, humans use association (and not exhaustive search through all the patterns) to interpret a situation. This process of association is somewhat random or at least extremely diverse (i.e. every person associates object A to object B using different sets of patterns and these patterns change over a period of time). Hence, it is extremely difficult to play around with such diversity using a single unifying theory. So, machine learning is not about building a machine that will always pass the "Turing test" for all the humans and in all the situations (though that is the ultimate goal of Machine Learning). (Btw, if you don't know about Turing test, read http://en.wikipedia.org/wiki/Turing_test.)

Having said that, it is also very important not to be too humble while defining what Machine Learning is, especially if you compare every machine learning situation to Turing machine. Remember, assumptions for Turing machine are (that are not necessarily applicable to all Machine Learning problems):
- Entire input is available at the beginning of computation.
- No state is maintained between execution of two (same or different) programs.
Especially interesting case where it is very difficult to define the problem in terms of Turing machine is Reinforcement Learning that we will study later. Also paper by Leeuwen and Wiedermann that discusses Turing machine in comparison to contemporary computing would be a nice read.

By now I made two key points (or mental images :)):
1. Machine Learning != Human Learning
2. Machine Learning != Turing machine

References:
1. The Turing Machine Paradigm in Contemporary Computing - Leeuwen and Wiedermann.

Research Notebook

After experimenting a bit with Evernote, I decided to move to OneNote for following reasons:
1. Evernote has no drawing tools. I like to draw and scribble a lot.
2. I get free subscription of OneNote as a student, whereas I need to pay for Evernote for some additional features like image recognition etc ... Though I don't need them right now, it would be handy to have them.
3. Onenote allows you to integrate audio and video notes, whereas Evernote is simply text based.
4. Onenote allows your notes to be arranged in hierachical folders, whereas Evernote is huge collections of notes with tags. Though I like the tagging feature (which Onenote also has), I don't think it is enough to organize my notes.

Though I think Synchronize feature in Evernote is awesome, I use Webexport + Publish as PDF feature in OneNote may be once a week. However, webediting feature in Evernote has no counterpart for OneNote, I will have to live with it until Office 2010. (I donot think LiveMesh solves this problem as I use Ubuntu on my lab machines)

I have used OneNote Web Exporter plugin to export my Research Notes for webview.

For now (until UF doesnot close my CISE account), I will maintain my notes at:
http://www.cise.ufl.edu/~npansare/webview.html and PDF at http://www.cise.ufl.edu/~npansare/ResearchNotebook.pdf

If you are using Firefox, you need to have Mozilla Archive Format add-on (since it is in MHT format). However, it still doesnot display properly. I recommend IE for webview or if you are a linux user, view the PDF.