What is this set all about?
About AI creating creativity from text generation, picture generation to movie script creation. The purpose is to help
... [Show More] augment our creativity
How does DL excel at creativity?
Our perceptual modalities, our language and artwork all have a statistical structure and learning it is what the technology excels at. They can learn the statistical latent space of images, music, and stories and sample from this space
Brainpower
Read More
What does text generation with LSTM do?
It grabs an RNN and create new text. This spans from automatic smart replies from chatbots, speech synthesis to creating new music
Language Model
Any network that can model the probability of the next token given the previous ones 'a cat on the ma'. It means to capture the latent space of the language, its statistical structure
Token
a word or character
Sample from it in a language model
Generating new sequences. Feed the model initial strings of text called conditioning data and ask it to generate the next character or the next word. You can even add the generated output back int the input data
Model 1: Character-level neural language model
We will start with the initial text 'The cat sat on the m'. It will go thru a language model that uses a probability distribution of the next character. It then undergoes a sampling strategy process where a sampled character is generated and put back into the initial text. Again, it goes thru the language model, a probability distribution and a sampling strategy where another character is generated. See page 294
Importance of sampling strategy
The way you choose next character is important. Greedy sampling - always choosing the most likely next character. Doing this, however, results in repetitive strings that dont look like a coherent language.
Stochastic Sampling: Introducing randomness in the sampling process, by sampling from the probability distribution for the next character. In this setup if e has a probability of 0.2, it will be picked 20% of the time. One problem with this strategy, you cannot control the randomness.
Low vs High Entropy
When we draw the next character, that sample will depend on the amount of randomness we elect to use or entropy. Low entropy will give us more predictable human like sequences whereas high entropy will give us either super creative or non-sense text. Its up to us to judge the optimal entropy for our end goals
Heres how to generate new text from our model built given a trained model and a seed text snippet
1. Draw from the model a probability distirbution for the next character, given the generated text available so far
2. Reweight the distribution to a certain temperature
3. Sample the next character at random according to the reweighted distribution
4. Add the new character at the end of the avaialable text
How to make a better character-level LSTM
As you add more data, more layers in the model, and for more epochs, you will achieve a more coherent and realistic model. Based on how you judge randomness will dictate how the next character is picked.
Inner-workings of Deep Dream
Runs a convnet in reverse, performing gradient ascent on the input to the convnet in order to maximize the activation of a specific filter in an upper layer
- Try to maximize activation of entire layers
- You start on a preexisting visual pattern, distorting elements of the image
- The input images are processed at different scales called octaves, which improves the quality of the visualizations
Do the different types of convnets, VGG16, VGG19, Xception, ResNet50 influence how the DeepDream looks?
Yes because different convnet architectures have built different learned features. Inception is known to be the best for this.
Deep dreams process can be also applied to:
Speech, music and more
Neural Style Transfer
Applying style of reference image to a target image while conserving the conetnt of the target image
Gram Matrix
The inner product of the feature maps of a given layer.
The inner product
Representing a map of the correlations between the layer features. These features capture the statistics of the patterns of a particular spatial scale, which empirically correspond to the appearance of the textures found at this scale
Difference w/ VGG19 over VGG16
It has three more convolutional layers. [Show Less]