>>1013188
>there's more to ML than NN
agreed, that is why I havent tried making one of those yet, and am starting with simpler models like custom estimators and generators. I saw something about tensorflow in Vault7 i think, and it was on my radar since then, but just started actually getting into it despite not having great hardware yet. you're working on models that write code? very cool. that is on my to do list as well, and part of why I want a model that is good at reading PDFs, so it can digest programming books and understand the contexts involved. genetic algotithms are one I planned to learn about in more depth when working on game playing and other problem solving models. but also that's part of why I wonder about combining algorithms for custom systems. writing code sounds like a problem solving task. what kind of training data does it have, just it's own assessment of it's own generated code? how do you define it's goal? you mentioned it working for simple problems, I'm wondering how you would tell it to write a video game in the SNES processor's native ASM ;) if it could even make a bootable ROM, that would be impressive. that is on my to do list, once simpler python/C programming is accomplished. documentation for game console SDKs
I have a vague plan to combine an adversarial network of researchers/spiders with estimators that decide if the data is useful or not, skim the data to determine which specialist models to train on the new data (is it language related, programming, psychology, nature, defense, etc), evolve those estimators and spiders using genetic optimization; feed the network with new data, and build in requests for human interaction to reinforce or correct it's training (like the captcha tests where it asks you which squares have traffic signs) when it isn't sure about something. it would have a drive for the reward of correct assessments about new data, and for collecting more quantity & quality of data. that's kinda my goal, a botnet that attempts an approach at generalized intelligence by having different cores that cooperate.
I know some people who like pokemon, and I saw some people online making scripts to generate new pokemon, and to recognize pokemon then speak data like the pokedex in the cartoon.
I tried a few pokemon generators and they both looked like garbage, but I admittedly only ran for ~100 epochs, not 5000, still I decided to train an estimator and then maybe use it as part of the model to be used by a generator, instead of using the GAN model and both discriminator & generator both training from scratch
That is one project on my to do list,a pokemon master AI, starting with a custom pokedex trained on game graphics (2d & 3d), the cartoon series, cards, fan art, plushies, etc - every format possible. writing an estimator that can take large resolution and ~64x64 sprites will be part of that. if it starts with a deeper understanding of what they all look like, the next step of generation may be more successful? I DL'd about 5 different pokedex sourcecode masters from github, and am still looking over the code before running any of them.
NVidia recently announced an epic graphic generator. they obviously have better hardware and access to data, but it's still a good indication of where technology is at. anybody had luck yet with generation of graphics via ML?
besides pokemon, deep exploit / deep-pwning looks very interesting. I've been setting up a VM network to let it loose on, isolated from any interaction with the host, should be fun to see what happens. A machine learning automated system that uses metasploit on Kali. anybody tried it yet?
also, if anybody knows a good way to train a system on PDF and other documents, I haven't yet found a simple solution for that. one goal is to train a chatbot using a free ebook library, then spiders on sites like gutenberg, so it can be a study partner. it needs to be able to use Natural Language Processing to gain comprehension of the information. I think an NN may work well for this, because I hear they work well for unsupervised learnig on unlabeled data, but am not sure. any ideas?