Interesting. Just for future reference, do you have any papers specifically in mind when you say "cool techniques"?
Interesting. Just for future reference, do you have any papers specifically in mind when you say "cool techniques"?
O_oIs this going to bite me in the ass? I wonder...
Why? A statement of intent isn't really bragging.
*shrug*
I also can't, despite effort, read the paste as anything similar to a thesis.
We know what you want can be done because the individual components all exist.
o_O
Or is the goal simply to show that you can build such a thing?
Soma
“Salem Was Wrong!” -- Pedant Necromancer
“Four isn't random!” -- Gibbering Mouther
Oh yeah. Quite a few.
One of the recent successes of the deep convolutional neural network technique was on ImageNet - a machine learning competition where people train systems to classify each of 1.2 million images into 1000 different classes (classes can be for example, wolf, cars, guitar, owl, etc). I believe the current state of the art on that competition is held by a CNN system.
http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf
People have also achieved very good results in things like hand-writing digit recognition (there is a famous MNIST dataset for that). The current state of the art for that is also CNN, and it's comparable to human performance (0.23% classification error).
MNIST handwritten digit database, Yann LeCun, Corinna Cortes and Chris Burges
There are many other similar datasets for people to experiment on. Traffic signs, different types of birds, etc. The one I am using is 200 MRI scans of volunteers' brains, but I don't think it's publicly available.
Biomedical imaging is something a lot of machine learning researchers are working on because it has immediate applications. For example, I remember reading about a CNN system used to classify tumours as benign or malignant (cancer), and it has already surpassed human expert performance for certain types of tumours. I can't find the paper anymore unfortunately.
Beyond image processing and computer vision, similar techniques (non-convolutional artificial neural networks) are also used in things like finding patterns in human genes to better predict serious illnesses or response to different drugs.
There is also a lot of work done on natural language processing for things like sentiment analysis - estimating how people feel about a product by what they are posting on Twitter or Facebook. They are able to predict the number of stars on an Amazon review to something like 0.3 star error on average.
Really exciting time for machine learning.
My master's thesis (not the MRI one - that's just what I'm working on on the side) is using deep artificial neural networks to have a program learn to play chess by itself (figure out winning strategies, etc), given almost no hardcoded knowledge, besides rules of the game, and the basic idea of tree search.
O_oThis is not research; this is building a product.
I'm sure you'd need to build a model in support of your thesis, but I would have expected the actual thesis to be an exploration of effective measures in multiplexing/duplicating/whatever `N' sources over `M' signals or some such statement.
The point I had offered wasn't that your thesis was inappropriate; the point was that the spiel didn't read, to me, like a thesis.
*shrug*
I'm sure your target is fine. I really just hadn't considered, as cyberfish explains, that different places would have a different standard for papers. If the paper fits the requirements, my expectation aren't relevant so it really doesn't matter.
Soma
“Salem Was Wrong!” -- Pedant Necromancer
“Four isn't random!” -- Gibbering Mouther
Yeah, it's not a requirement that it needs to be research here. It needs a certain scientific depth, sure, but it does not need to be research.
Sure, this is what I consider that I'm "working on." Part of it is also researching what you say, things like "effective measures in multiplexing/duplicating/whatever `N' sources over `M' signals" and figuring out how to put everything together. Part of it is also looking at related works and comparable that is to my own product. How applicable is their methods to mine? Advantages? Disadvantages? Etc.
So yeah, it's not "just" about building a product - the research is also part of it and all of it, including building the model and describing how it was done, will be part of the paper.
It's like half the work is building the product and half the work is the research, writing the paper, etc.
Enormous shout. Also it is wonderful to see the juxtaposition of programming levels/interests/projects that people on this forum are involved in, it is just such a wealthy resource for anybody to have access to that.I'm currently working with convolutional neural networks, to automatically segment MRI images of brains - labeling each voxel (3D pixel) with an anatomical region of the brain. This has applications in things like Alzheimer's diagnosis, where the physician needs to be able to see shape and size of specific parts of the brain, that may be buried deep inside the scan.
Thought for the day:FLTK: "The most fun you can have with your clothes on.""Are you sure your sanity chip is fully screwed in sir?" (Kryten)
Stroustrup:
"If I had thought of it and had some marketing sense every computer and just about any gadget would have had a little 'C++ Inside' sticker on it'"
I'm a muppet.
Not the Jim Henson kind, the other kind.
The code I'm doing at work would require a fair bit of explanation and isn't that interesting, so it's not really worth explaining.
I currently don't have time to work on any of the significant personal projects I want to do, but when chances arise, I do some of the smallish "fun" programs on my list. I just wrapped up a console version of a "sudoku solver" that appears to be working well. Soon I will create a GUI front end for this program, and cross that one off my list.
I was wondering if that required an understanding of combinatorics? I've tried learning combinatorics several times and have always been thrown off by the amount of mathematical notation used with no explanation. It's hard to tell if I'm interpreting that sort of thing correctly with no feedback lol (actually very frustrating).
WndProc = (2[b] || !(2[b])) ? SufferNobly : TakeArms;