Topic: Generator spam thread (Read 12692 times)

JoshuaFH · « **Reply #30 on:** May 13, 2019, 08:36:21 am »

When will T2T stop being funny? All it does is spit out comedy gold:

Quote

I'm a weirdo with a beard! Oh no, I'm a BEARDO D: Yes!

Abe, I had no idea there were still animals roaming the city!

A BEARD O: My beard looks like your hair, don't you think?

A BEARDO O: I bet it does! A BEARD O: What makes you think, though?

Abe, we're in love. Abe, your hair sounds like mine.

A BEARDO D: Oh, please, I want to go to America with a big beard! I mean, I want to wear a baseball cap to the World's Fair!

A BEARD O: The World's Fair is not in America. Don't forget about that!

A BEARD O: Who are you?

Abe O: I'm a BEARDO D: Yes! Abe O: No, really! A BEARDO D: What do I look like?

Abe O: Why did I ask?

A BEARDO D: I'm a BEARDO

TamerVirus · « **Reply #31 on:** May 13, 2019, 09:25:24 am »

Selections from the Dark Knight Rises

Spoiler: Bane of The Universe (click to show/hide)

Spoiler: Bane's mask (click to show/hide)

Spoiler: Feeling in charge (click to show/hide)

methylatedspirit · « **Reply #32 on:** May 13, 2019, 10:51:59 am »

I'm feeling particularly immature tonight, so enjoy some dick jokes, I guess.

Spoiler: A confuses computers by telling them to suck 'thy' dick. (click to show/hide)

Spoiler: A's conversation about their computer-related dicksuckery devolves into a series of increasingly incomprehensible questions. (click to show/hide)

ChairmanPoo · « **Reply #33 on:** May 13, 2019, 02:22:19 pm »

Spoiler (click to show/hide)

I have this idea. I'm going to give a text to the neural network and then feed the output back to it several times. That way we'll get a feedback loop in which it generates it's own content.

I used a neural network from DeepMind as the training example but it should easily work well.

The final piece of the puzzle is to feed the output of the final output feed back into the Neural Network so that it creates its own content again. How do we achieve this? Well, there are a few ways to do and as I've shown here, most people haven't thought about how to do that yet.

The first thing we need to do is to use deep learning techniques like Convolutional Neural Networks. You can check out the Wikipedia article for Convolution Neural Networks if you're unfamiliar with these super useful and promising techniques. They can be used in a wide variety of ways and with a variety of different problem domains and there are a lot of examples available for you to choose from.

In this article we're going to use a simple example from the video game series Far Cry and combine it with machine learning to explore how to model games through Deep Learning. We'll see how to combine different models to create more complex models, how to compare results, and finally, and most importantly, how to apply them to improve our models.

Convolutional Neural Networks (CNNs) as we know them

What makes CNNs different from regular neural networks (CNNs) are the way in which they are formed. This has implications for our decision making. Imagine that, we're trying to find the most accurate prediction for which the player needs to change positions in a game; a good idea can take hours to achieve but will be correct 50% of the time with the best of intentions! For this reason CNNs should be built with high resolution (i.e. 2N data points) to allow for extremely accurate predictions.

When building CNNs, consider that convolutional networks are useful for generating large quantities of "residuals" and for storing that residual data and for manipulating it later. For example, we could generate a new batch of 1000 neurons each and train them in an existing batch. This training process uses an algorithm called CNN-A, which is able to output a residual in the form of a "column" of 1000 neurons per column. These residual neurons must be combined with the first batch to create a larger batch for further validation.

If our network model has residual neurons that are not the neurons we were going to train directly from the original batch we use to train and then use these neurons later (again, in a separate batch), each neuron can be individually combined with it's predecessor to generate a new batch with only 2 inputs. This combined batch can then be further processed to create the final "batch" in the same manner we would train it individually.

However if we are only trained to learn and do a large number of steps, this may only produce an output that is a few dozen neurons. The best way to ensure we minimize this number is to train only a larger number of neurons in parallel from a previous batch (at a much higher gradient). So we could do this by starting with a single batch (we would train these neurons individually from scratch). To improve the model performance we would repeat these steps.