Viktor Petukhov, PhD
My scientific interestsfor collaborations
  • Enhancing research by NLP tools
    My current research interests are about enhancing human research capabilities with AI, data analysis and tech in general. At the moment I’m looking into various ways to use NLP to simplify working with existing literature, such as search, summarization and automated reasoning (broadly, the agenda is inspired by AllenAI and Ought)
  • AI Safety
    I am also quite interested in AI Safety, explainability of neural networks and formal guarantees on machine learning algorithms. I havea vast experience of working with graphs, including Graph Signal Processing, which I’d be interested to apply to studying neural networks, particularly using the framework of Complexity Theory.
  • AI + politics
    Expanding on the idea of enhancing the human capabilities, one especially interesting application is policy-making and political research. Data science is uniquely suited for contributing towards evidence based policy-making, as well as providing trustworthy information for general audience. Here I’m particularly excited about work of Our World in Data and bellingcat. So, I’m looking for ways to apply data scrapping and NLP to find insights on pressing political and social issues.
My relevant projects
  • Modularity of neural networks (code)
    With a team from AI Safety Camp we worked on enforcing modular structure for neural networks. We developed severaly ways to adjust a loss function to achieve that and demonstrated it on a simple MLP, as well as on a word2vec model.
    Show more ↓
    In my part, I extended and applied the Graph Spectral Regularization method created by Tong, et. al for enforcing arbitrary graph structures. First, it allowed to convert word2vec embedding into a visual field, which could be easily inspected by eyes (similar to how CNNs are analysed). Second, we could train word2vec so the neurons were groupped into five modules, which, supposedly, correspond to high-level features of the texts.
  • Visualization of collections of documents
    During my PhD I was looking for information on areas I don’t understand. There, conventional search engines give a very small subset of the information on the topic. It leaves you very little idea of what you miss and even how much missing information is there in total. So, we created a tool for building an overview of large collections of documents.
Share your contacts to discuss your ideas and needs
If there is no specific request, but you are interested in how machine learning can be useful to you — feel free to write too
If you need to attach any materials — send your request to e-mail
viktor.petukhov@pm.me