Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Reelya]

Sadly, that's the stuff that you're going to end up getting if you get paid to do transcription, in the long run. The smart way to run a business like that is to feed everything through machine algorithms, have that generate metrics for how confident it is that it's transcribed it right, then anything that fails that test, you hand it off to some freelancer for a few cents a minute, get the result back, check it, then use that as further input to train the algorithm. The window of stuff you need to bring a human in is going to be ever-smaller.

Perhaps the first company just isn't as ahead on the machine-learning as the second company is, so they're passing easier stuff to humans.

[Naturegirl1999]

How would one code a machine learning neural network?

[Arx]

Your journey starts here.

At least, IMO that's one of the best available introductions to machine learning with neural nets. I'm not going to try to cram an explanation into one post.

[Naturegirl1999]

Thank you

[Magistrum]

A lot of ways. Usually you make a array, give it a the values of your input(so, color value of pixels to analize images, amplitude, frequency and time to analize sound, and so on), then you make a second array and make the values of that be a function of all the values of the input array. Back in the day people used a sigmoid function that simplifies the value to a number between 0 and  1.
Then you get a last array of possible results you want(image of kettle, image of window) each value is a sigmoid of the sum from all the values from the previous array. If the value of the position "window" is 1, then the it is saying the image is that of a window. If it is less than one you can treat that as a "confidence level" measurement. So, 0.9 would be 90% sure it is a window.
At first you fill all the arrays relations besides the input with random numbers. The network sucks, so you see what it gives you and do what is called "backpropagation", to train the machine. You take each value you want to have (say, that was a image of a kettle, so you want the value "window" to be 0 and the value "kettle" to be 1), and add little multipliers to each value of the previous array in the function to make the function in the last array reach the value you want. These multipliers are called "weights" and they are what changes in the network when it learns.
You do that a few times with different examples and eventually you weights will start settling down at some value in each position, the smaller the change with each iteration the better your network should be.

This is the basis of what people call "neural networks", since if you draw that out it looks like "neurons" triggering other neurons in layers until at they trigger a "neuron" in the last layer, your result.

Edit: Ninja'd by Arx because I did try to cram the explanation into one post

[Naturegirl1999]

A lot of ways. Usually you make a array, give it a the values of your input(so, color value of pixels to analize images, amplitude, frequency and time to analize sound, and so on), then you make a second array and make the values of that be a function of all the values of the input array. Back in the day people used a sigmoid function that simplifies the value to a number between 0 and  1.
Then you get a last array of possible results you want(image of kettle, image of window) each value is a sigmoid of the sum from all the values from the previous array. If the value of the position "window" is 1, then the it is saying the image is that of a window. If it is less than one you can treat that as a "confidence level" measurement. So, 0.9 would be 90% sure it is a window.
At first you fill all the arrays relations besides the input with random numbers. The network sucks, so you see what it gives you and do what is called "backpropagation", to train the machine. You take each value you want to have (say, that was a image of a kettle, so you want the value "window" to be 0 and the value "kettle" to be 1), and add little multipliers to each value of the previous array in the function to make the function in the last array reach the value you want. These multipliers are called "weights" and they are what changes in the network when it learns.
You do that a few times with different examples and eventually you weights will start settling down at some value in each position, the smaller the change with each iteration the better your network should be.

This is the basis of what people call "neural networks", since if you draw that out it looks like "neurons" triggering other neurons in layers until at they trigger a "neuron" in the last layer, your result.

Edit: Ninja'd by Arx because I did try to cram the explanation into one post

Thank you
I will now try to learn what and how to make an array in C++

[Il Palazzo]

I've been reading comments under a youtube video about geology that mentions the greenhouse effect.

[Reelya]

I love C++ but you should probably try starting your machine-learning thing in a language such as Python.

C++ is a compiled language, it takes quite some time to compile a program as it gets larger. Python is interpreted, so you can run it, make changes then run it again almost instantly. The trade-off is that C++ can be a lot faster when it does run, but for debugging / design work, interpreted languages are a lot faster to work with.

EDIT: I like this series of videos on machine learning (from Stanford university). What I like about this is that he doesn't just jump into explaining neurons and perceptrons and the like, he starts by teaching how to do linear regression, then logistic regression, and then he shows how you use those building blocks (which can already solve some problems) to build a neural network.

https://www.youtube.com/watch?v=PPLop4L2eGk

[dragdeler]

-

[Yoink]

WTF, I'm still sick. This dang cough keeps hanging on. Guess I'm gonna have to buy something to sort it out.
I woke up when my alarm went off then promptly went back to sleep 'cause I felt all shitty.   
Also, today is two days before the day on which I need to have clean clothes, so I guess I'd better do the damn washing.

[methylatedspirit]

I've been having suicidal thoughts lately, and today I realized that I could off myself by either consuming every Seroquel tablet I currently have (totalling to 1800 mg worth as of last count. Given that I normally take 25 mg, that's probably fatal), or jumping off a building, head-first.

Then I wondered if the fact that I absolutely can kill myself is the reason why I haven't gone ahead with this dying thing.

[Doomblade187]

I've been having suicidal thoughts lately, and today I realized that I could off myself by either consuming every Seroquel tablet I currently have (totalling to 1800 mg worth as of last count. Given that I normally take 25 mg, that's probably fatal), or jumping off a building, head-first.

Then I wondered if the fact that I absolutely can kill myself is the reason why I haven't gone ahead with this dying thing.

Glad you're here with us still.

[wierd]

I have harbored such thoughts before.

What has worked for me in the past, is reflecting on how absurd the motive for such a drastic action usually is.

For example:  A co-worker (who has been having a midlife crisis) came to work with a 'rare' flavor of energy drink. (Not regionally available, but strongly preferred)  A resident that does not respect the boundaries of other people's things took it while we were both engaged in rooms, opened it, and drank it.  This was the final straw for my co-worker, who went off the rails, and actually hurt herself in a fit of rage over the incident. (Again, this is the result of a final straw type situation, not indicative of normal behavior.)  She was discussing her desire to make this shit stop by ODing on her insulin.

Now-- Think about it rationally;  A 2$ energy drink is the final motivation for such an action?  Really?  Isn't that just a bit absurd?  When we reach our emotional coping limits, we reach the end of our coping limits, so explosive results are expected, not unreasonable. What is unreasonable is that we get pushed into such situations in the first place, and end up resorting to "final solution" thoughts or deeds over things that normally we would consider just plain stupid.

Being able to stop, reflect on that, and go "yeah, that's absurd." has been a major rallying ability for myself more than once, since it follows with "What's more absurd, is that I have gotten to this point by allowing (list of stressors) to get me here."


I know it can be hard to muster that level of rational clarity in the depths of emotional chaos, (from personal experience), but doing so is very helpful, and a source of great strength to change your position.

[Arx]

It's also contingent on the last straw actually being absurd, or the stressors being resolvable.


It's quite likely that one of my friends committed suicide last week. I can't tell for sure. It's slightly distressing how little it bothers me. I'm just so burnt out and detached from reality these days. Nothing really seems to matter, and I keep having dissociative moments.

[Doomblade187]

It's also contingent on the last straw actually being absurd, or the stressors being resolvable.


It's quite likely that one of my friends committed suicide last week. I can't tell for sure. It's slightly distressing how little it bothers me. I'm just so burnt out and detached from reality these days. Nothing really seems to matter, and I keep having dissociative moments.

if you are working, I highly recommend taking a sick day, preferably several. Take some time to recover, get closure. Also, as someone who handles grief... Detachedly, just know you're not alone in not feeling bothered as you feel you should. Grief is personal - deal with it the way you feel is right.