Yann LeCun

Great couple of weeks for the lab, and for Andrew Bender, PhD in particular! Two of Andrew's papers diving deep into the meaning of neural alpha oscillations are now published. The first, out in PNAS, is "Differential representations of spatial location by aperiodic and alpha oscillatory activity in working memory" [1]. There's been a ton of super cool work in the last decade or so showing that the amplitude and distribution of visual cortical alpha rhythms encode the contents and spatial location of information held in working memory. Our lab has shown that our ability measure neural oscillations is confounded by non-oscillatory aperiodic activity, so we were curious how much of this encoding is *really* oscillatory. Analyzing seven different EEG datasets from Ed Awh and Ed Vogel's labs, we found that, after correcting for aperiodic activity, alpha encoding actually *improved*! And that time-resolved aperiodic activity seemed more to encode the stimulus while it was on the screen. We're now looking much more deeply into this idea that sensory information that often comes into our brains is aperiodic, and encoded that way, but then somehow transforms into an oscillatory neural code for maintenance. The second paper, in the Journal of Cognitive Neuroscience, is "Resting-state Alpha and Mu Rhythms Change Shape across Development But Lack Diagnostic Sensitivity for Attention-Deficit/Hyperactivity Disorder and Autism" [2]. LinkedIn's word limits keep me from being able to dive too deeply into this one but WOW did Andrew and Natalie Schaworonkow do an incredible amount of beautiful, technically sophisticated work. They analyzed EEG data from thousands of children, classified alpha band activity either as visual cortical alpha or sensorimotor mu based on their spatial topography, and then quantified the fine-scale waveform features of both. (By the way, the pre-prints of both these papers are on freely available on bioRxiv if you can't access the published versions.) [1] https://lnkd.in/eN9rxdZZ [2] https://lnkd.in/eNCjyQYh

97

2 Comments

Where does complexity really reside — in the *dynamics* or in the interaction *structure*? I’ve put together a short note reflecting on the recent critique about the graph vs hypergraph representation of higher-order interactions from Tiago Peixoto, Leto Peel, Thilo Gross and Manlio De Domenico (https://lnkd.in/emrN2r7h). This note examines structural–dynamical equivalence and asks under what conditions a generic, nonlinear graph dynamics admit a simpler, linear representation running on a more complex, higher-order structure. The approach involves simple Carleman linearization arguments and aims to provides a nuanced perspective on this debate, where the choice of representation is perhaps a matter of modelling convenience rather than principle. https://lnkd.in/ejSnkgFC

64

12 Comments

This was the winner of the NeurIPS 2025 best paper award... H/T Raymond de Lacaze Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free (Alibaba Grfoup et al, May 2025) Paper: https://lnkd.in/gT9fHPu6 Abstract: "Gating mechanisms have been widely utilized, from early models like LSTMs and Highway Networks to recent state space models, linear attention, and also softmax attention. Yet, existing literature rarely examines the specific effects of gating. In this work, we conduct comprehensive experiments to systematically investigate gating-augmented softmax attention variants. Specifically, we perform a comprehensive comparison over 30 variants of 15B Mixture-of-Experts (MoE) models and 1.7B dense models trained on a 3.5 trillion token dataset. Our central finding is that a simple modification-applying a head-specific sigmoid gate after the Scaled Dot-Product Attention (SDPA)-consistently improves performance. This modification also enhances training stability, tolerates larger learning rates, and improves scaling properties. By comparing various gating positions and computational variants, we attribute this effectiveness to two key factors: (1) introducing non-linearity upon the low-rank mapping in the softmax attention, and (2) applying query-dependent sparse gating scores to modulate the SDPA output. Notably, we find this sparse gating mechanism mitigates 'attention sink' and enhances long-context extrapolation performance." GitHub: https://lnkd.in/g4AqKqTV Hugging Face: https://lnkd.in/gqFfk2ae

7

Pleased to share this new publication in the Journal of Open Research Software, part of an ongoing collaboration between the CulturePlex Lab and the Donostia International Physics Center (DIPC). Titled WikiTextGraph, we present a Python tool where it extracts text and graph of Wikipedia across multiple languages (English, Spanish, Greek, Dutch, German, Hindi, Vietnamese, Basque, Polish, and Italian). #Python #OpenSource #DataScience #KnowledgeGraphs #WikiTextGraph

22

In his latest VSM Data Science Lab article, Dr. James McCaffrey presents a complete end-to-end demo of Tsetlin Machine binary classification implemented in C#. Tsetlin Machine models combine elements of propositional logic, rule-based systems, and finite state automata. The article walks through data encoding, model architecture, training, evaluation, and prediction using the Iris Dataset -- a detailed guide for developers exploring alternatives to neural network–based systems.

2

1 Comment

📢 New arXiv preprint available, where we propose evTransFER, a transfer learning framework for event-based cameras, taking advantage of a feature extractor designed to encode the spatio-temporal dynamics, and then transferring the trained encoder weights. 💡 We apply this framework to address the problem of Facial Expression Recognition, where we learn first to reconstruct faces from events (thus learning the spatio-temporal dynamics of faces), and then the learnt encoder is transferred to the Face Expression Recognition problem. A key aspect is that this does not require additional annotations, just "pairs" of event sequences and frames during training. We achieve an accuracy of 93.5%, a 25.6% increase over the state of the art. 🌐 https://lnkd.in/ezyyttZv Joint work with Ignacio Bugueño Córdova 📜 evTransFER: A Transfer Learning Framework for Event-based Facial Expression Recognition Abstract: Event-based cameras are bio-inspired vision sensors that asynchronously capture per-pixel intensity changes with microsecond latency, high temporal resolution, and high dynamic range, providing valuable information about the spatio-temporal dynamics of the scene. In the present work, we propose evTransFER, a transfer learning-based framework and architecture for face expression recognition using event-based cameras. The main contribution is a feature extractor designed to encode the spatio-temporal dynamics of faces, built by training an adversarial generative method on a different problem (facial reconstruction) and then transferring the trained encoder weights to the face expression recognition system. We show that this proposed transfer learning method greatly improves the ability to recognize facial expressions compared to training a network from scratch. In addition, we propose an architecture that incorporates an LSTM to capture longer-term facial expression dynamics, and we introduce a new event-based representation, referred to as TIE, both of which further improve the results. We evaluate the proposed framework on the event-based facial expression database e-CK+ and compare it to state-of-the-art methods. The results show that the proposed framework evTransFER achieves a 93.6\% recognition rate on the e-CK+ database, significantly improving the accuracy (25.9\% points or more) when compared to state-of-the-art performance for similar problems. RISLab #eventbased #computervision #deeplearning INSTITUTO DE CIENCIAS DE LA INGENIERÍA ICI-UOH Universidad de O'Higgins

29

When Cognitive Vision made National News: The VAMPIRE Project In 2002 the Pattern Recognition Lab was peacefully preparing to start the work on a new EU Project: Visual Active Memory Processes and Interactive REtrieval - VAMPIRE for short - when the dpa ticker suddenly flashed the sensational line "Universität Erlangen erforscht Vampire." Within minutes, the institute's venerable phone system began to glow like a garlic-free Christmas tree. Journalists from boulevard tabloids to public-service broadcasters demanded to know whether the undead were now an EU priority and - if so - how many would be available for on-camera interviews. https://lnkd.in/dPXg9HkJ

29

🚗✨ Paper Accepted at CVPR 2026! ✨🚗 Our paper on 📏 uncertainty-aware perception evaluation for autonomous driving systems (ADS) has been accepted to the CVPR 2026 main conference. Why this matters 🔍 ADS safety depends on perception that stays reliable across distance and tough weather / lighting 🌧️🌙. But perception outputs are often stochastic, and standard metrics don’t capture uncertainty well—especially when it affects downstream safety decisions. What we propose 🧠 ✅ Perception Characteristics Distance (PCD): an uncertainty-aware metric measuring the farthest distance where an object can be reliably detected. ✅ average PCD (aPCD): summarizes overall perception capability by averaging PCD across multiple thresholds. Dataset: SensorRainFall 📌 Collected on the Virginia Smart Road using camera + radar + LiDAR 📷📡🌐 under: Weather: clear / rainy ☀️🌧️ Illumination: daylight / streetlight / nighttime 🌤️💡🌙 Includes ground-truth distance, 2D boxes, and segmentation masks 🧾 Key takeaway 📈 Across state-of-the-art models, aPCD reveals robustness differences across weather and illumination that conventional metrics often miss. Resources 🔗 📂 Dataset: https://lnkd.in/e7g7q6dd 💻 Code: https://lnkd.in/etuhZF9J 📄 Preprint (arXiv): https://lnkd.in/eTcG_NTm Collaborators 🤝 Boyu Jiang, Zhengzhi Lin, Lanxin Xiang, Loren Stowe, Feng Guo — thank you for all your work on this project 🙌 Special thanks to Dr. Feng Guo for his support and guidance 💡 📍 See you at CVPR 2026 in Denver! 🏔️ #CVPR2026 #Denver #AutonomousDriving #Perception #Uncertainty #ComputerVision #ADAS #ADS #Dataset #Safety

20

1 Comment