Pinned
Pasquale Minervini
11.6K posts
Pasquale Minervini
@PMinervini
Research in ML/NLP at @EdinburghNLP (tenured faculty at @EdinburghUni), Co-Founder @Miniml_AI, @ELLISforEurope Scholar, neuralnoise.com
- Based on our NeurIPS 2021 paper (arxiv.org/abs/2106.01798), we developed a PyTorch library that easily allows you to include discrete combinatorial solvers into larger neural architectures, and back-prop through them, in one line of code! 🚀 Check it out! github.com/uclnlp/torch-i…
GIF - If you're a 1st/2nd year PhD student, please make a website about yourself and your work! No need to go into debt; just use GitHub Pages and JekyllIf you’re a guy in your early 20s, buy at least 40 springer textbooks. Go into debt if you have to
- Replying to @PMinervinito clarify -- I didn't mean to shame the authors of these papers; the real issue is AI reviewers, what we see here is just the authors trying to defend against that in some way (the proper way would be identifying poor reviews and asking the AC or meta-reviewer to discard them)
- Hey everyone! I'm joining the School of Informatics at the University of Edinburgh (@EdinburghNLP, @infated, @EdinburghUni) as a faculty member in September 2022, and have funding for PhD students in NLP and ML! 🧑🎓🚀 neuralnoise.com/2021/call-for-… (🧵1/5)
- For the folks working on reasoning with neural models -- just a reminder that if you train a ResNet to produce the shortest path in a map, accuracy will be ~0%, while if you incorporate a shortest path algorithm into the model, you quickly go above 90%! (arxiv.org/abs/2106.01798)
- Conditional Theorem Provers are scalable neuro-symbolic reasoning models that learn to recursively select and generate rules on-the-fly conditioned on the goal via gradient-based optimisation! To appear at #ICML2020, Arxiv arxiv.org/abs/2007.06477 Slide neuralnoise.com/icml20_talk.pdf 1/N
- Replying to @ilyasutthe fact that GPUs are much better than all other NN methods suggests that the brain might be using GPUs too
- Am I doing this right? Or should I use torch.einsum to achieve consciousness instead?
it may be that today's large neural networks are slightly conscious - Our paper “Differentiable Reasoning on Large Knowledge Bases and Natural Language” will appear at #AAAI2020 as an oral! We scale neural theorem provers to massively large KBs and corpora. Paper and TF/Py🔥 code will follow :) W/ @backprop2seed, @_rockt, @riedelcastro, @egrefen
- OMG our @NeurIPSConf paper (arxiv.org/abs/2106.01798) got featured on @ykilcher's YouTube channel!! 🚀🚀🚀 This is quite a life goal for me -- Yannic's paper explanations are extremely clear (and fun!), and I always listen to them when I'm digging into a new topic 🤗🔥New Video🔥How to backpropagate through an algorithm? Seems crazy, but this paper shows it's actually possible for a large class of algorithms, such as k-subset, ILP, and many graph algorithms. Watch my (amateur 🙃) attempt at an explanation here: youtu.be/W2UT8NjUqrk
arxiv.orgImplicit MLE: Backpropagating Through Discrete Exponential Family...Combining discrete probability distributions and combinatorial optimization problems with neural network components has numerous applications but poses several challenges. We propose Implicit...
