Mediaspace scheduled maintenance: Aug 25, 2026 07:00 - 12:00 AM. During this time, videos will be temporarily unavailable. Check status updates.

Lecture

Neural Tangent Kernel: Generalization in Deep Learning

In course

PHYS-754: Lecture series on scientific machine learning

This lecture presents ongoing work on how scientific questions can be tackled using machine learning. Machine learning enables extracting knowledge from data computationally and in an automatized way. We will learn on examples how this is influencing the v ...

Description

This lecture explores the concept of the neural tangent kernel (NTK) in deep learning, showing how it converges to a constant value as the neural network width approaches infinity. It discusses the impact of network size on generalization, highlighting the role of random noise and fluctuations during training and initialization. The lecture also delves into the asymptotic behavior of generalization as the number of parameters increases, demonstrating the variance of the output and its relationship with network size. Various measures of generalization are analyzed, revealing patterns such as peaks and decays. The presentation concludes with insights on the behavior of wide neural networks under gradient descent and their evolution into linear models.

This video is available exclusively on Mediaspace for a restricted audience. Please log in to MediaSpace to access it if you have the necessary permissions.

Watch on Mediaspace

Instructors (10)

Anne-Clémence Corminboeuf

David Richard Harvey

Alexander Mathis

Alexander studied pure mathematics with a minor in logic and theory of science at the Ludwig Maximilians University in Munich. For his PhD also at LMU, he worked on optimal coding approaches to elucidate the properties of grid cells. As a postdoctoral fellow with Prof. Venkatesh N. Murthy at Harvard University and Prof. Matthias Bethge at Tuebingen AI, he decided to study olfactory behaviors such as odor-guided navigation, social behaviors and the cocktail party problem in mice. During this time, he increasingly got interested sensorimotor behaviors beyond olfaction and started working on proprioception, motor adaption, as well as computer vision tools for measuring animal behavior. In his group, he is interested in elucidating how the brain gives rise to adaptive behavior. One of the major goals is to synthesize large datasets into computationally useful information. For those purposes, he develops algorithms and systems to analyze animal behavior (e.g. DeepLabCut), neural data, as well as creates experimentally testable computational models.

Philippe Schwaller

Matthieu Wyart

Giuseppe Carleo

Giuseppe Carleo is a computational quantum physicist, whose main focus is the development of advanced numerical algorithms tostudy challenging problems involving strongly interacting quantum systems.He is best known for the introduction of machine learning techniques to study both equilibrium and dynamical properties,based on a neural-network representations of quantum states, as well for the time-dependent variational Monte Carlo method.He earned a Ph.D. in Condensed Matter Theory from the International School for Advanced Studies (SISSA) in Italy in 2011.He held postdoctoral positions at the Institut d’Optique in France and ETH Zurich in Switzerland, where he alsoserved as a lecturer in computational quantum physics.In 2018, he joined the Flatiron Institute in New York City in 2018 at the Center for Computational Quantum Physics (CCQ), working as a Research Scientist and project leader, and also leading the development of the open-source project NetKet.Since September 2020 he is an assistant professor at EPFL, in Switzerland, leading the Computational Quantum Science Laboratory (CQSL).

Official source

Related concepts (32)

Linear model

In statistics, the term linear model is used in different ways according to the context. The most common occurrence is in connection with regression models and the term is often taken as synonymous with linear regression model. However, the term is also used in time series analysis with a different meaning. In each case, the designation "linear" is used to identify a subclass of models for which substantial reduction in the complexity of the related statistical theory is possible.

Convolutional neural network

Convolutional neural network (CNN) is a regularized type of feed-forward neural network that learns feature engineering by itself via filters (or kernel) optimization. Vanishing gradients and exploding gradients, seen during backpropagation in earlier neural networks, are prevented by using regularized weights over fewer connections. For example, for each neuron in the fully-connected layer 10,000 weights would be required for processing an image sized 100 × 100 pixels.

Generalized linear model

In statistics, a generalized linear model (GLM) is a flexible generalization of ordinary linear regression. The GLM generalizes linear regression by allowing the linear model to be related to the response variable via a link function and by allowing the magnitude of the variance of each measurement to be a function of its predicted value. Generalized linear models were formulated by John Nelder and Robert Wedderburn as a way of unifying various other statistical models, including linear regression, logistic regression and Poisson regression.

Deep learning

Deep learning is part of a broader family of machine learning methods, which is based on artificial neural networks with representation learning. The adjective "deep" in deep learning refers to the use of multiple layers in the network. Methods used can be either supervised, semi-supervised or unsupervised.

Linear regression

In statistics, linear regression is a linear approach for modelling the relationship between a scalar response and one or more explanatory variables (also known as dependent and independent variables). The case of one explanatory variable is called simple linear regression; for more than one, the process is called multiple linear regression. This term is distinct from multivariate linear regression, where multiple correlated dependent variables are predicted, rather than a single scalar variable.

Related lectures (35)

Provable Benefits of Overparameterization in Model Compression

Explores the provable benefits of overparameterization in model compression, emphasizing the efficiency of deep neural networks and the importance of retraining for improved performance.

Introduction to Machine Learning: Supervised Learning

Introduces supervised learning, covering classification, regression, model optimization, overfitting, and kernel methods.

Convolutional Neural Networks

Covers Convolutional Neural Networks, including layers, training strategies, standard architectures, tasks like semantic segmentation, and deep learning tricks.

Introduction to Machine Learning: Linear Models

Introduces linear models for supervised learning, covering overfitting, regularization, and kernels, with applications in machine learning tasks.

Neural Networks: Multilayer Learning