MDZip: Neural Compression of Molecular Dynamics Trajectories for Scalable Storage and Ensemble Reconstruction 1. MDZip is a novel neural compression framework that achieves over 95% reduction in storage size for molecular dynamics (MD) trajectories while preserving essential dynamical information. This makes it possible to share and store large-scale MD data without significant loss of detail. 2. The framework uses convolutional autoencoders trained on individual systems to reconstruct atomic trajectories with high geometric fidelity from compact latent representations. It is physics-agnostic but still accurately preserves ensemble-level features such as RMSD fluctuations, pairwise distance distributions, and principal components. 3. A key innovation is the use of residual (skip-connected) autoencoders, which consistently improve reconstruction accuracy and reduce outliers compared to traditional autoencoders. This results in fewer structural deviations and better preservation of local and global properties. 4. MDZip supports customizable compression–accuracy trade-offs, allowing users to balance storage efficiency and reconstruction quality based on their specific needs. It also enables secure and FAIR-compliant data sharing by requiring only the trained model parameters, scalers, topology, and compressed trajectory for decompression. 5. While local structural deviations can impair energetic fidelity, the study shows that short energy minimization can partially recover physically reasonable conformations. This suggests practical strategies for enhancing the physical realism of reconstructed trajectories. 6. The framework is evaluated on a diverse benchmark of proteins, protein–peptide complexes, and nucleic acids, demonstrating its broad applicability and robustness across different biomolecular systems. It offers a scalable solution to current storage limitations in MD data management. 📜Paper: biorxiv.org/content/10.1101/… #MolecularDynamics #NeuralCompression #DataStorage #Bioinformatics #ComputationalBiology

📘 New Theoretical Framework Drop: Stable & Convexified Information Bottleneck → informationbottleneck.com This work addresses critical instabilities in the classical IB formulation — particularly the non-smooth behavior of p(z|x) under β-phase transitions. ✅ Convexified objective with entropy regularization ✅ Smooth representation flow with no sudden collapse ✅ Symbolic reformulation aligned with mutual information geometry ✅ Compatible with deterministic & variational approximations ✅ Supports continuous β-path tracking (∂L/∂β ∈ C⁰) ✅ Ready for integration in high-dimensional encoders (e.g., ViT, ResNets) Now being applied to symbolic bottlenecks, interpretability flows, and compressed multi-modal fusion. 📎 Read the full theoretical breakdown and join the post-IB evolution. #InformationBottleneck #ConvexOptimization #MutualInformation #DeepLearningTheory #VariationalInference #SymbolicML #RateDistortion #InformationTheory #NeuralCompression #BetaPhaseTransition #RepresentationStability #FarukAlpay #IBFramework #ThermodynamicLearning #DataEncoding #SignalDeformation #LayerwiseCompressibility #IBObjective #AdaptiveRepresentation #PhaseSmoothness #DeterministicBottleneck #XAI #MLFormalism #VariationalBound #JensenGap #NeuralInformationFlow #DifferentiableInformation #InformationTopology #MLTheory #StatisticalLearning #AIResearch

Call for papers #ICML2025 Workshop on ML4Wireless!🛜 Let's bridge the gap between machine learning and communication! 👉 More info at: sites.google.com/uniroma1.it… @icmlconf #icml2025 #icml #icmlworkshop #neuralcompression #GenAI

Introducing the first #NeuralCompression benchmark for astrophysics data! #ICLR2025 🚀🌌

We're thrilled to announce our #NeuralCompression workshop at @NeurIPSConf! Join us for our third event, following ICML2023 and ICLR2021. Send us your papers—from theory to SOTA algorithms to preliminary but promising ideas! Deadline: TBD; check the website.

AMD reveals a new neural texture compression system aimed at reducing storage and VRAM requirements, similar to Nvidia's method from last year. #NeuralCompression #Vram #Gaming haywaa.com/article/amds-neur…

We just pushed the NeuralCompression 0.3.0 release. The primary update for this one is utilities for training generative compression models in PyTorch: - Perceptual loss metrics (LPIPS, DISTS, FID/256) - HiFiC autoencoder architecture - Discriminator superclass (including ILLM)

#NeuralCompression enthusiasts: our #ICML workshop's call for paper is out! Visit our website for all details! 🤖🗜️👇

My talk on the challenges of practical #NeuralCompression is now on Youtube: youtube.com/watch?v=v6DWk7ch…. I still can't believe it's not better compressed 😉 Thanks, @ICBINBWorkshop, for hosting me!

Happy to announce @julberner has introduced a new feature into NeuralCompression on bits-back with diffusion models. It achieves a rate close to the (negative) ELBO of the paper “Improved Denoising Diffusion Probabilistic Models”. Description here! ⬇️.

📢📢New feature in #NeuralCompression repo: Bits-Back compression for diffusion models! Compress image data 🖼️ using diffusion models at an effective rate close to the (negative) ELBO. See: github.com/facebookresearch/… Some context ⏩[1/4]

Very happy to share our repository on NeuralCompression! So far we have implementations the scale hyperprior image compression model, the DVC video compression model, and arithmetic/rANS entropy coders in JAX that can run on the CPU or GPU. github.com/facebookresearch/…

Definately worth a listen 👏👏👌@TPMPodcast @Chews_Health @AnninaBSchmid #neuralcompression #nerveregeneration #neuropathy