MetaMerge: Model Merging with Meta Networks
Merging ViTs and LLMs using a pretrained (graph) neural net.
blank
Boris Knyazev's homepage.
Latest posts
REAM: Compressing Mixture-of-Experts LLMs
Merging experts in Mixture-of-Experts (MoE) LLMs to compress a 235B LLM.
NiNo: Learning to Accelerate Training of Neural Networks
Explaining our ICLR 2025 paper and visualizing neuron permutation symmetry.
Can we do better than Convolutional Neural Networks?
PyTorch Implementation of “Image Classification with Hierarchical Multigraph Networks” from BMVC 2019 The number of pixels in the top row is 11, 7 and 1000 times larger (from left to right) than the number of “superpixels” in the bottom row. Can we use the superpixels rather than raw pixels as input and improve on convolutional neural networks? The British Machine Vision Conference (BMVC),…
Spectral Graph Convolution Explained and Implemented Step By Step
As part of the “Tutorial on Graph Neural Networks for Computer Vision and Beyond” The Fourier basis (DFT matrix) on the left, in which each column or row is a basis vector, reshaped to 28×28 (on the right), i.e. 20 basis vectors are shown on the right. The Fourier basis is used to compute spectral convolution is signal processing. In graphs, the Laplacian basis is used described in this post.…
Anisotropic, Dynamic, Spectral and Multiscale Filters Defined on Graphs
As part of the “Tutorial on Graph Neural Networks for Computer Vision and Beyond” I’m presenting an overview of important Graph Neural Network works, by distilling key ideas and explaining simple intuition behind milestone methods using Python and PyTorch. This post continues the first part of my tutorial . Graph of Graph Neural Network (GNN) and related works. Some other important works and edges…
Tutorial on Graph Neural Networks for Computer Vision and Beyond (Part 1)
Tutorial on Graph Neural Networks for Computer Vision and Beyond I’m answering questions that AI/ML/CV people not familiar with graphs or graph neural networks typically ask. I provide PyTorch examples to clarify the idea behind this relatively new and exciting kind of model. A figure from ( Bruna et al., ICLR, 2014 ) depicting an MNIST image on the 3D sphere. While it’s hard to adapt…