Machine Learning Research
Researching Machine Learning
How do
Basically, all of the test cases need to be the same length (number of pixels, characters, etc) and have a normalized output (again same length, and stable index across possible inputs)
Example ways of normalizing test cases
- Random crop of a fixed size out of images
- Images placed randomly on a random noise buffer of the same size
- A byte array of ~1000 random characters, and the log line placed at a random offset
- Array of random assembly opcodes, with a disassembly of the function placed at a random offset
Example ways to remove bias
- Shift the hue of the image
- Rotate the image
- Random placement on field of noise
Tags?
To classify tags, just have each tag as a bit in the output array:
{pixel 1x1, pixel 1x2, ...} { tag_has_house?, tag_has_dog?, tag_has_cat?}
That way, the output from a test will be the likely hood of each tag.
Upsides:
- One test run for any number of tags
- Only one model for any number of tags
Downsize:
- Model needs to be retrained when a new tag is applied to the training set
Need to research
- GAN: https://en.wikipedia.org/wiki/Generative_adversarial_network
- RNN: https://en.wikipedia.org/wiki/Recurrent_neural_network
- Reversing the network
- https://github.com/galeone/tfgo
- others in https://github.com/avelino/awesome-go#machine-learning
- Object identification
- https://github.com/plutov/packagemain/tree/master/04-tensorflow-image-recognition
Share this post
Twitter
Facebook
Reddit
LinkedIn
StumbleUpon
Pinterest
Email