I'm a doctor and recently I turned down a very good job offer to become a radiologist...

There is nothing wrong with Machine Learning. But always be careful with hyped topics. Machine Learning is just a fancy new word for statistics. When you take all the buzzwords away, you turned down a job offer as a radiologist to study statistics. That's unusual, but you're probably able to earn money in both fields

>I'm a doctor and recently I turned down a very good job offer to become a radiologist. In the long term I do want to be a radiologist, but the reason why I turned the offer down was to study artificial intelligence and machine learning for a whole year.
Don't bother, you are clearly retarded.

Why would you even think about doing this? Jumping on some random technology and praying that, even if it somehow becomes extremely relevant, that you are the one to survive the market?
Why not join a research group if you are genuinely interested? Or go to an already established company and make your luck there?

Its such an enormously retarded decision to take a risk on a thing you literally know nothing about...

> So in 5 years time when i'm finally ready to start user'S RADIOLOGY MACHINE LEARNING INC will the market already be dominated by deepmind and whoever else is doing it?
Nobody knows, but what I can tell you that the chances of you actually succeeding are near zero, because if it becomes a super relevant technology some random doctor with basically zero experience in the field will be beyond irrelevant compared to the companies which actually will exist by then.

>Im not saying i've quit out on radiology, ive just put it on hold for a year while I build some skills in ML and also do a project in it.
How can any person rationally think this is a good Idea?
Why aren't you focusing on actually getting one thing finished and while doing that trying to join some research group to do research?

That's a stupid way to look at it. ML lets you do things that statistics can't. The knowledge that makes you good at designing neural networks is not useful for collecting and interpreting statistics, and vice versa.

You don't need it. If you want into field, you can simply apply as domain expert.

>That's not how it works. You are not likely to find a job after a year of doing nothing related to the job unless there are special circumstances, and "i studied machine learning" is not one of them.

Medical training in the UK is very different to what you think it is. There will be no bias against me for turning down a job, it's a really standardized process and I nailed the interview 6 months ago, i got like 96/100 and since then i've added more to my CV. It's very common in the UK for doctors to take a so called "gap year" many do it for work related/recreational travel by working in australian emergency rooms for a year, others do a teaching degree for a year, some just take a year off doing ad-hoc work because they haven't decided on a career.

>Don't bother, you are clearly retarded. Why would you even think about doing this? Jumping on some random technology and praying that, even if it somehow becomes extremely relevant, that you are the one to survive the market? Why not join a research group if you are genuinely interested? Or go to an already established company and make your luck there? Its such an enormously retarded decision to take a risk on a thing you literally know nothing about...

well I actually will be doing some ML research during my year out. This time next year I should have finished training some medical image model (whatever that means, hopefully in the next two months i'll understand it all).

i'm quite surprised by the shock people have that I turned down a radiology job. Its really quite low risk, I've already got my medical degree (or MD as you guys call it) and i've acquired a bunch of core skills which will allow me to work in any emergency room or medical ward in my country. that's not going away anytime soon

>radiology
HAHAHA, enjoy literally having cancer in 30 years. Good on you to turn down that radiology residency, now go do a specialty that won't kill you before you can upload your mind to the cloud.

Also, interventional cardiologists and neurosurgeons are stealing your jobs away anyway.
>t. cardiology: thanks amerifats for eating all that shit last Thanksgiving, now I can dish out more Lipitor while Stacey from Pfizer gives me my third blowjob for the day after paying for lunch.

Not OP but how do I graduate from toy problems in tensorflow and go to the next step.
Right now I am learning a bunch of basic knowledge I suppose they teach in more formal trainings (stuff like LSTM, dropout, autoencoders...) but whenever I try to do something non trivial with those I either fail because I lack data or because my poorfag tier machine takes too long to optimize with reasonable batch sizes.

I have a severe case of sour grapes and I consider focusing instead on "quick" reinforcement learning but I'm overwhelmed by the number of research papers I can't understand.

It's stupid to talk about Machine Learning as it were distinct from statistics in any way. They're doing absolutely the same thing, reasoning about information in data. It just happens to be that statistics is traditionally not the sexiest discipline of mathematics, so some wise guys decided to call themselves data scientists

>well I actually will be doing some ML research during my year out. This time next year I should have finished training some medical image model (whatever that means, hopefully in the next two months i'll understand it all).
>i'm quite surprised by the shock people have that I turned down a radiology job. Its really quite low risk, I've already got my medical degree (or MD as you guys call it) and i've acquired a bunch of core skills which will allow me to work in any emergency room or medical ward in my country. that's not going away anytime soon
The point is that what you are doing is enormously counterproductive, if you want to do research, then go and do research.
But there is literally no point in stopping being a doctor to do machine learning, you are doing nothing but shooting yourself in the foot.

Trying to transition from being a doctor, to doing machine learning is just enormously retarded, it is much more valuable to go and do the research you want to do and then learn from the people you are doing research with, while researching. If you do that your question doesn't even arise, any company doing medical imaging and machine learning will want to hire you and if you have some great Idea with a large potential market you will have enough connections to have a competent team.

>i'm quite surprised by the shock people have that I turned down a radiology job.
No its people telling you that it is a bad idea to go from being a doctor to doing machine learning.

All right man, my suggestion for you is please follow UC Berkeley's CS 188.
It's an excellent course for understanding the general principles in AI.
It'll be very useful for you to learn and understand the concepts on an intuitive level.

>t. used it to refresh my memory of AI as a PhD student in my 1st year

>Jumping into a field whose ultimate logical endgame is machines learning how to do your job better.

Bullshit. The main tool of neural networks is optimization, via gradient descent. Statistics does not ever use that. The #1 issue with neural networks is vanishing gradient when you train the network. Statistics has nothing of this sort because in rthe first place it does not do anything that neural networks do. The process of making a neural network is first designing a huge difficult mathematical formula and then altering its numeric parameters iteratively, by plugging inputs into it and comparing outputs to expected outputs. This is the core, the most basic part about neural networks, and there is no analog to it in statistics. Either you are uninformed, or you are blatantly lying.

am I tripping or what? literally met an MD with pretty much exactly the same story. D'you live in Edinburgh rn OP?

Neural networks are just function approximators, which for one reason or another work well for high dimensional data. Classical Statistics also rely on optimization techniques. Again Stochastic gradient descent is commonly used, because it works in high dimensional spaces (not very well indeed, it is occasionally outperformed by Random search aka Genetic Algorithms). So if I fit a linear function via Least Squares to my data, am I doing Machine Learning? What when I use Stochastic Gradient descent? I also could design a Neural net, that does just the same linear regression (which is by the way a frequently used introductory example in neural network lectures) In any case one does the same thing. You search for the regression (or classification) function that approximates your data the best.

> Machine Learning is just a fancy new word for statistics
Please never say this to anyone else
It just shows how ignorant you are

as for the OP
You are looking for deep learning specifically.
Better skip stupid machine learning (in case you're studying years old kernel machines you are on a very wrong track)
Learn CNNs, resnet architectures and similar stuff
Remember, for radiology you need to identify images and that involves Deep learning, not just Machine learning

>will the market already be dominated by deepmind and whoever else is doing it
You're already very late user, there are many such startups already.

Hello sir
ever heard of deep learning? It's a subset of Machine learning which is nowhere near statistics you speak of.
CNNs, MLP nets, Spiking nets
Classifying dogs from cats, it's actual learning and nothing like statistics

Poor guy, machine learning things are good but today Deep Learning methods are cruching everything

youtube.com/watch?v=CVYCmB1e8J8

youtube.com/watch?v=kw2Ire8ecMU

Pytourch is more friendly for newcomers and research, facebook sponsored course because pytourch 1.0 announce.

udacity.com/course/deep-learning-pytorch--ud188

Neural Networks and Deep learning Charu C aggarwal is very good book for newcomers.

More traditional book is
deeplearningbook.org

publications.lib.chalmers.se/records/fulltext/248445/248445.pdf

Other medical deep learning examples book is

Deep Learning and Convolutional Neural Networks For Medical Image Computing Le Lu Yefeng Zheng Gustavo Carneiro.

Machine learning is a dead end AI attempt just like expert systems, plugging more GPUs will not bring us truly intelligent systems. If anything ML is just a more computationally expensive statistics.

Have you read the post? I provided an example, where I showed, that Machine Learning contains the classical problems of Statistics. I also explained where neural nets fit in this picture, they're function approximators in high dimensional spaces (for example approximation of a function, that separates cats and dogs in a n*m dimensional space of pixel data). So either your definition of Machine Learning is so broad, that it contains Statistics ( which is stupid) or they're just the same thing

>Another UK user with a MD
Thought I was the only one on Jow Forums. 3 years working as a Obstetrician & Gyanaechologist now (with some side emergency room work). If your asking if ML will be over saturated in the future then the answer is that it already is. I don't feel that it is yesterday's news yet but you'd have to do something to really stand out then just learn the basics & already known techniques. As companies will just be looking for cheap labor which you probably don't want to be.

Also:
>There will be no bias against me for turning down a job
You are correct but you do realise that a company won't hold that position for a year for you to go off & learn about something that will have very little impact to the job in the short term.
>It's very common in the UK for doctors to take a so called "gap year" many do it for work related/recreational travel.
Your half right here. YES it is very common to take a gap year BUT not fresh out of school. They usually spend 2 or more years actually working before they take a gap year. Even the very few who do it right after school usually do charity work or work in A&E in other countries, which still has some prevalence to there medical training.
>i've acquired a bunch of core skills which will allow me to work in any emergency room or medical ward in my country.
Yes, everyone who does a MD has to learn those core skills don't act like your the only one in the UK who had to do it. But have you ever worked in a emergency room its very long hours for very little pay off which would probably impact your ML research more than radiology, which means that you've shot yourself in the foot for just falling for a meme subject.

Long story short I say you should have taken the job & done it gradually on the side. It may take longer to learn that way but you would have a stable income for the foreseeable future AND then maybe think about taking a gap year to learn more.

Similar position. So ok, learn Python so you can use frameworks like Tensorflow straight away.
When you are comfortable with that, I advise you to switch to Julia.
To keep things simple, syntax is pretty much the same and you can call python and c++ libraries without boilerplate.

>So if I fit a linear function via Least Squares to my data, am I doing Machine Learning? What when I use Stochastic Gradient descent?
Yes, you are doing machine learning. A tiny part of it.

You are also using a shallow one layer network for this which means none of knowledge obtained over last decade on how to train complex networks is useful.

Machine learning is way, way greater than that, and what you call statistics intersects only a little bit with neural networks.

>approximation of a function, that separates cats and dogs in a n*m dimensional space of pixel data
Convolution neural nets dont work this way
CNNs legit learn to make outlines, identify gradients, correlate various features and cluster pixels. It is learning by itself! This isn't statistics where i have explicitly write equations to differentiate red to black gradient from green to red. This is the whole point of ML, rather than writing countless if-elses, let the machine learn on its own.
If this is statistics, then your definition of statistics is too broad (which is stupid)

You may argue Reinforcement learning is not part of ML but RL is nowhere, i repeat, nowhere near any of the statistics you know of.

I hope i can make you think otherwise.

There is still space for OS initiatives. There is a lot of space for social media.

I think you are in a good path with machine learning and radiology, there will still be space in 5 years.
>Module 1: Programming with C++ and MATLAB (I am aware these arent really useful for ML, but they are useful for learning the foundations of programming like FOR loops etc and the foundations of image processing such as image convolution)
Not really useful, skip and start programming with Python and R
>Module 2: Software engineering with python (Stuff like unit testing, version control, creating ?modules? in python)
This is really straight forward and many of the books you would use to lean python would already include it.

Add Calculus and linear algebra in the first semester, you will really need optimization and differential equations.

The second semester is good but the drop the word "Basics".

>It is learning by itself!
If you can put the cool-aid down for a moment, you're still doing fitting of the function that your neural network defines, and the fact that you an find cool things in kernels is just a side effect of that. You are still fitting a function - the difference is that your function is a lot more complex.
Not the guy you're arguing with. I don't consider ML to be a branch of statistics, but saying things like learns by itself is extremely misleading.

Agreed, I went a little overboard with that

I don't know. But I fully support what you do and focus on ML. even though ML is hype right now. However you can create you own company (startup) that ML focus on medical such radiology (comparing image) using image processing. When I was in college (electrical engineering) one of my junior created a bachelor thesis about image processing on patient to determine a disease. But it would be perfect if you're a doctor and learning ML. good luck.

some an addition, user the mentioned on the first comment about matlab. it's worth to learn. because I one of my junior create that, he's using matlab.

matlab is not at all worth learning IMO
python/R ftw

Diz you learn just how to use the framework? Or did you study why it works like calculating the derivative function, doing everything by hand?

If you just learned how to use the framework blindly, you are no different than a pajeet janitor.

You can learn 'machine learning' from CMU for free, all the courses are open functionalcs.github.io/curriculum/#sec-17

It's entirely doing statistics. So basically you're going to get a masters degree in stats which may not be so bad as a radiologist. In 5 years, there could be insane developments. For example, Deep Learning is basically at a stalemate. We have all the tools and methods, but not the hardware or network ability to actually do it as conceived. So think of all the guys in the 1990s who invented things like in-memory dbms and other stuff we use today on paper, that's exactly where DL is.

However when we do have this hardware/ability, we can then do insane shit like have a phone that will tell you everything going on in a room as you enter such as what subject everybody is talking about, how long the line is to order something, what tables are open for you to sit at and talk to people and which are private, ect ect. All instantly as you walk through the door.

At that time translation will also have a breakthrough where you can just put a phone down and have a convo with somebody with a completely different language on auto translate.

So yes, there is definitely a future for this stuff but likely it will be in the hands of a very few gigantic corporations like Google/MIcrosoft

Machine learn is all about learning patterns and generalizing those patterns.

It is not as easy as you think. If it were easy, we would have already a JAV uncensoring nn. It wouldn't really need to undo the censorship with fidelity, just creating a convincing image would do the job, and it is 100% possible to achieve it with NN.

>Machine learn

Yes

Dude fuck you, you dont even know math to do ML

machine learning and specifically AI at this time are just buzz words. In current year AI DOES NOT EXIST therefore I dont really understand what could you possibly study for the whole year on the subject of AI. Seems to me like its just another university money grabbing scheme.
Machine learning is also similar situation, though its debatable what really is machine learning, a fancy word for statistics?

neural networks is absolute meme, if anyone is offering courses for this then its 200% scam

I feel bad now
I had some doubts but now I feel kind of silly for enrolling in my uni's data science masters program

Don't take Jow Forums so seriously.

You fell face down into the meme

>I'm a doctor and recently I turned down a very good job offer to become a radiologist.
>the reason why I turned the offer down was to study artificial intelligence and machine learning
>Ive learned the fundamentals of programming and soon I'll be learning the ML fundamentals.
>The question is where do I go from here?
you took a complete 180 and didn't do your research?

Attached: 1516752075494.png (733x1200, 565K)

I'm kind of with statistics user here. I've been coming into the field of ML for quite some time now on my own and already knowing classical statistics well, and I also see ML as higher-end statistics.

Both classical statisticians and machine learning claim Lasso/Ridge regression as their own, which makes it seem like a pissing contest at this point.

Who cares if CNN's learn on their own? That's simply a statistical function that systematically writes its own if-else's to manage the overwhelmingly complex routine aspect of learning to approximate some data, which is great, but only deserves its own elevated sub-field.

Assuming you are OP, I can assure you that, if you are comfortable with decent-level College math (Multi-variate Calc./diff-eq, and enough familiarity to understand the mathematical machinery of basic statistics) and know some programming, you can 100% study ML on the side if you're smart enough to have been offered a Radiology job already.

Knowing 1) only that level of math, and 2) knowing programming are enough. If you don't know either of these well, add another half-year of preparation.

Keep at it, user. It may have been bad to not study on the side, but your intuition is right that ML is going to become incredibly important in medicine.

Attached: 10499758-3x2-940x627.jpg (940x627, 63K)

You're confused.
Machine Learning is just "Turing Complete Statistics".
I would say that it is nothing like the traditional pure mathematical functions you are used to, but that is technically wrong due to category theory.

So in fact, machine learning is performed using pure functions, which are merely a much more highly dimensional version of y=mx+b.
It doesn't change that all you're doing is trying to approximate a statistical model by fitting the equation to the dataset.
There is a reason that ML uses a lot of statistics terminology, and that's because they are the same field.

en.wikipedia.org/wiki/Statistical_learning_theory