@azettaai just did a cool thing

have you ever wondered what your classifier is doing? well it is just a vortex system link: tahabouhsine.com/vortex/ repo: github.com/mlnomadpy/vortex/

have you ever wondered: - wtf is the classifier doing? - wtf is backpropagation doing? - wtf is the meaning of life? well here is a beautiful animation to explain the first two questions :D this part of the explainability toolkit we are building @mlnomads

we are in @wandb don't mind the 2 hours sleep schedule face find me if you wanna chat about white box nn

wanna have fun this summer? try to train the smallest best model you can on small scale datasets in the donwscaling laws challenge from @mlnomads the prize? friends you will make along the way links bellow

what if your second brain was...agentic???? cool right, well that's what i wasted today implementing in @bashnota now you can turn any llm out there into your own agent just ask your question, and the agent will generate and execute code inside the nota analyse the multimodal results and generate a conclusion here is an unedited 2x speedup agent in work using gemini model

just finished the integration of @kaggle jupyter servers with @bashnota now you can create agi in style amma share the demo tomorrow am tired af but you can try it, already pushed it to the demo offline[.]bashnota[.]com enjoy

Hello everyone, This tweet is about a side research project I’ve been busy with for the past two months @mlnomads Through this work, I aim to create a white-box deep learning, where you can trace the geometry/topology of your input from the start to the logits. Before diving into details, please note: this work is still in progress. It needs large-scale validation, so don’t take anything as ultimate truth. Unfortunately, due to unforeseen circumstances, I can’t finish it alone, so I’m looking for your feedback. As part of exploring contrastive learning (preprint: Simo Loss AFCL), I introduced a new operation: the ⵟ-product (yat), which combines notions of orthogonality/parallelism and distance. Unlike the dot product, it works in a pseudo-metric space, capturing direction and distance while naturally inducing non-linearity. Using this, I developed a phyisics inspired "neuron", called the ⵟ-neuron, which is the backbone of the first version of Neural-Matter Networks (NMNs) In NMNs, we no longer need activation functions. Non-linear behavior is achieved intrinsically through the ⵟ-product. The network operates with only softmax for class probability distribution for the output layer and attention (I also proposed softermax, which remove the need for exp operation). The layer code in flax linen and pytorch are available in the repo, i am gonna share the next part of the blog soon, but before that, I would love to get your feedback on my thoughts and the current state of the work. Also, during this work I used resources from both the Google Developer Expert program and Google AI/ML Developer Programs. Thank you for all the support everyone.

We're building a crowdsourced Moroccan Darija-English translation dataset to advance language models for Moroccan Darija! 🎉 🔗 Project: darijaai.mlnomads.com 🔗 GitHub: github.com/ElhoubeBrahim/col… 🔗 Blog post: blog.mlnomads.com/darijaai-p… #AISprint

wa7ed chher hadi bdina kankhedmou fl community 3la publishing blog posts 3la l ML b darija. link: blog.mlnomads.com/ml-enginee… author: @elhou_io #ml #darija

Happening tomorrow #mlnomads

We are back! Today 8PM GMT 1 Topics lii ghadii nkhedrou 3lihoum ou ghadi nkhedmou bihoum: - Self-Supervised Learniing - Contrastive Learning - LoRA - Sehd li kayn Link: mlnomads.com/discord See you there!

Scientists, designers, storytellers, makers, technologists, and innovators: we need you! Join the @SpaceApps Challenge, an #OpenScience hackathon, Oct. 7-8, 2023, to solve challenges we face on Earth and in space. bit.ly/44JtwPj