Antonio Montano 🪄’s Post

An overview of vector fields and flow matching on Euclidean space and Riemannian manifolds, concepts familiar to those with basic knowledge of differential geometry.

"Flow Matching Guide and Code" by Yaron Lipman, Marton Havasi, Peter Holderrieth, Neta Shaul,  Matthew Le, Brian Karrer, Ricky Chen, David Lopez-Paz, Heli Ben Hamu, Itai Gat "Flow Matching (FM) is a recent framework for generative modeling that has achieved state-of-the-art performance across various domains, including image, video, audio, speech, and biological structures. This guide offers a comprehensive and self-contained review of FM, covering its mathematical foundations, design choices, and extensions. By also providing a PyTorch package featuring relevant examples (e.g., image and text generation), this work aims to serve as a resource for both novice and experienced researchers interested in understanding, applying and further developing FM." Paper: https://lnkd.in/dhBQepeh Code: https://lnkd.in/dakgF6gd #machinelearning

If you work with Generative AI, Flow Matching is an essential method that has proven efficacious. Check out this excellent guide and accompanying code to familiarize yourself if you haven’t already.

I am going to present our new #ICML2024 paper “LayerMerge: Neural Network Depth Compression through Layer Pruning and Merging” in Vienna on July 23, 2024. Feel free to visit our poster session! I would love to talk with you and discuss our work. 🖼️ Poster location: Hall C 4-9, Number 305 🕜 Time: Tue 23 Jul 1:30 p.m. CEST — 3 p.m. CEST ❗️TL;DR: LayerMerge reduces depth of the CNN and diffusion models by jointly pruning convolution and activation layers and merging resulting consecutive convolution layers. 💡 Overview: Pruning the depth of the large models while maintaining performance can drastically save inference time and memory consumptions. Traditional methods for reducing network depth usually follow one of two approaches: 1️⃣ Pruning Convolution Layers: Aggressively removes parameters, risking loss of important information. 2️⃣ Pruning Activation Layers and Merging Layers: Eliminates redundant activation layers and merges resulting consecutive convolution layers, potentially increasing kernel sizes and negating speed gains. ⚡️ LayerMerge addresses these issues by jointly pruning convolution layers and activation functions. It optimizes which layers to remove, speeding up inference while minimizing performance loss. Since this selection process involves an exponential search space, we formulate a novel surrogate optimization problem and efficiently solve it via dynamic programming. 🔥 Our results show that LayerMerge outperforms current methods for reducing network depth in tasks including image classification and generation. We attach the qualitative demo that shows the efficacy of LayerMerge. I express my deep gratitude to my co-authors, Marwa Halabi, Mingi Ji, Hyun Oh Song. 👉 Paper: https://lnkd.in/gz92Dq9F 👉 Code: https://lnkd.in/gtJ5Appg 👉 Project page: https://lnkd.in/g8DBDJ82  👉 Poster: https://lnkd.in/ghDuiPfJ

🔍 Exploring the Regularity of Characteristic Functions in Besov-Type and Triebel-Lizorkin-Type Spaces 📈 A recent study by Wen Yuan, Winfried Sickel, and Dachun Yang delves into the optimal regularity of characteristic functions within Besov-type and Triebel-Lizorkin-type spaces. This research is pivotal for understanding the membership conditions for characteristic functions associated with complex geometric domains, such as snowflake domains and spiral-type domains. ### Key Highlights: - Optimal Regularity Determination: The authors establish necessary and sufficient conditions for characteristic functions to belong to these spaces under specific geometric constraints. - Applications in PDEs: The findings have implications in partial differential equations, particularly concerning pointwise multiplier problems. - Characterization of Domains: The study provides insights into how different types of domains (thick vs. weakly thick) affect the regularity properties of their associated characteristic functions. ### Implications: This research not only enhances our theoretical understanding but also opens avenues for practical applications in mathematical analysis, particularly within function spaces. 🔗 Dive deeper into this fascinating topic by exploring their findings on how these mathematical structures interact with real-world applications! #AI #Algorithms #ArtificialIntelligence #BesovSpaces #DL #DS #DataScience #DeepLearning #FunctionalAnalysis #ML #MachineLearning #Mathematics #PDEs #Research #Tech #Technology #TriebelLizorkinSpaces Source: https://lnkd.in/efcUXuig

💥💥💥 Can LLMs do research-level mathematics? There’s a moving frontier between what can and cannot be done with LLMs. 👉 Lattice-Valued Bottleneck Duality Robert Ghrist, Julian Gould, Miguel Lopez Abstract This note reformulates certain classical combinatorial duality theorems in the context of order lattices. For source-target networks, we generalize bottleneck path-cut and flow-cut duality results to edges with capacities in a distributive lattice. For posets, we generalize a bottleneck version of Dilworth's theorem, again weighted in a distributive lattice. These results are applicable to a wide array of non-numerical network flow problems, as shown. All results, proofs, and applications were created in collaboration with AI language models. An appendix documents their role and impact. 👉 https://lnkd.in/dYbZPddc #machinelearning

Ariadne’s Graph (AriGraph) integrates semantic and episodic memories into a memory graph, enabling an LLM-based agent, Ariadne, to learn and utilize a world model from environmental interactions, enhancing spatial orientation and exploration. Evaluated using TextWorld tasks (navigation puzzles, cleaning, and cooking), Ariadne outperformed other memory approaches and reinforcement learning baselines by extracting semantic triplets from observations and updating the graph. The agent’s planning and decision-making benefited from a memory retrieval system, allowing efficient navigation and task execution. Comparisons with human players and reinforcement learning models highlighted Ariadne's superior performance, particularly in complex tasks. https://lnkd.in/gDqX4RRe

Positional embeddings are a fundamental component in transformer models, providing critical positional information to the model. This blog will explore the mathematical concepts behind positional embeddings, particularly focusing on sinusoidal positional embeddings, and will work through a detailed example to illustrate these concepts. #ArtificialIntelligence #Math https://lnkd.in/gyFfdgWW

#QTNano: Machine learning as a tool for chemical space exploration broadens horizons to work with known and unknown molecules. At its core lies molecular representation, an essential key to improve learning about structure-property relationships. Recently, contrastive frameworks have been showing impressive results for representation learning in diverse domains. Therefore, this paper proposes a contrastive framework that embraces multimodal molecular data. Specifically, our approach jointly trains a graph encoder and an encoder for the simplified molecular-input line-entry system (SMILES) string to perform the contrastive learning objective. Since SMILES is the basis of our method, i.e., we built the molecular graph from the SMILES, we call our framework as SMILES Contrastive Learning (SMICLR). https://lnkd.in/dqXpfbyQ

AI can now create a replica of your personality "Led by Joon Sung Park, a Stanford PhD student in computer science, the team recruited 1,000 people who varied by age, gender, race, region, education, and political ideology. They were paid up to $100 for their participation. From interviews with them, the team created agent replicas of those individuals. As a test of how well the agents mimicked their human counterparts, participants did a series of personality tests, social surveys, and logic games, twice each, two weeks apart; then the agents completed the same exercises. The results were 85% similar." MIT Tech. Review: https://lnkd.in/d_hmBfsn paper: Generative Agent Simulations of 1,000 People https://lnkd.in/dsZYNaDQ