Loss landscape and symmetries in Neural Networks

Candidate: Mario Geiger

Professor: Matthieu Wyart

PhD private defense

Table of contents

Color code

Questions

Open problems

model

\(\displaystyle w\in \mathbb{R}^N\)

Classification with deep learning

Classification with deep learning

dataset

\(\displaystyle w\in \mathbb{R}^N\)

\(\displaystyle \mathcal{L} = \frac1P \sum_i \ell(f(w, x_i),y_i)\)

loss function

\(\displaystyle w\in \mathbb{R}^N\)

Classification with deep learning

\(\displaystyle \mathcal{L} = \frac1P \sum_i \ell(f(w, x_i),y_i)\)

gradient descent

\(\displaystyle w\in \mathbb{R}^N\)

Classification with deep learning

\( w^{t+1} = w^t - \nabla_w \mathcal{L}(w^t) \)

\(\displaystyle P\)

\(\displaystyle N\)

Classification with deep learning

number of parameters

number of training samples

In this talk I will use the following notation

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

• No guarantees to be convex
• High dimensional space

Why not stuck in a local minima?

• Over parameterized: \(\mathcal{L} \to 0\)
• \(\mathcal{L}\) connected at the bottom
  Freeman and Bruna (2017)
  Soudry and Carmon (2016)
  Cooper (2018)
   
• Under Parametrized: \(\mathcal{L}\) glassy
  Baity-Jesi et al. (2018)

How is the transition between under and over parameterized?

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

• State of the art neural networks \( P \ll N\)

: test/generalization error

= number of parameters

Neyshabur et al. (2017, 2018); Bansal et al. (2018); Advani et al. (2020)

Why does the test error decreases with \(N\)?

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

\(P =\) size of trainset

\(\delta =\) distance to closest neighbor

Bach (2017)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

Hestness et al. (2017)

regression + Lipschitz continuous

Luxburg and Bousquet (2004)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

• Kernel Method:
  sample complexity /= #{ invariant transformations }
  Bietti et al. (2021)
   
• Convolutional Neural Network (CNN) are invariant under translations

sample complexity = \(P\) required to get \(\epsilon\)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

Figure from (Geometric Deep Learning by Bronstein, Bruna, Cohen and Veličković)

• Translations are a small set in the space of image transformations
   
• Labels are invariant to smooth deformations or diffeomorphisms of small magnitude
  Bruna and Mallat (2013); Mallat (2016)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

Is stability to diffeomorphisms responsible for beating the curse of dimensionality?

• Recent empirical results revealed that small shifts of images can change the output a lot
  Azulay and Weiss (2018); Dieleman et al. (2016); Zhang (2019)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

Can we generalize CNN invariance to other symmetries?

• CNNs perform much better than Fully-Connected networks
   
• Many tasks involve symmetries like rotation or mirror

Paradoxes

Landscape vs Parameterization

Data Symmetries

Non-convex \(\mathcal{L}\)

Overparameterization

Curse of dimensionality

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

How is the transition between under and over parameterized?

The same question was asked in the physics of glasses

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

\(\displaystyle U = \sum_{ij} U_{ij}\)

Van der Waals

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

\(\displaystyle U = \sum_{ij} U_{ij}\)

Van der Waals

Finite range interaction

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

\(\displaystyle U = \sum_{ij} U_{ij}\)

Van der Waals

Finite range interaction

SAT/UNSAT problem

num. of constraints \(N_\Delta\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

\(\displaystyle U = \sum_{ij} U_{ij}\)

Van der Waals

Finite range interaction

SAT/UNSAT problem

num. of constraints \(N_\Delta\)

Sharp Transition:

Jamming

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Low density: flat directions

\(N_\Delta = 0\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Low density: flat directions

High density: glassy

\(N_\Delta = 0\)

\(N_\Delta \geq N\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Low density: flat directions

High density: glassy

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Low density: flat directions

High density: glassy

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Low density: flat directions

High density: glassy

Prediction for Neural Networks

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

cross entropy loss (long range)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

cross entropy loss

(long range)

Paradoxes

quadratic hinge loss

(finite range)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

margin

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Fully Connected 5 hidden layers relu, ADAM MNIST (pca d=10)

Geiger, M., Spigler, S., d’Ascoli, S., Sagun, L., Baity-Jesi, M., Biroli, G., and Wyart, M. (2019)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Fully Connected 5 hidden layers relu, ADAM MNIST (pca d=10)

Geiger, M., Spigler, S., d’Ascoli, S., Sagun, L., Baity-Jesi, M., Biroli, G., and Wyart, M. (2019)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Fully Connected 5 hidden layers relu, ADAM MNIST (pca d=10)

Geiger, M., Spigler, S., d’Ascoli, S., Sagun, L., Baity-Jesi, M., Biroli, G., and Wyart, M. (2019)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

There is sharp jamming transition, in the same universality class as ellipses

Paradoxes

How is the transition between under and over parameterized?

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

There is sharp jamming transition, in the same universality class as ellipses

Paradoxes

How is the transition between under and over parameterized?

Why not stuck in a local minima?

Crank-up the number of parameters and you are not stuck

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

There is sharp jamming transition, in the same universality class as ellipses

Paradoxes

How is the transition between under and over parameterized?

Why not stuck in a local minima?

Crank-up the number of parameters and you are not stuck

What about overfitting?

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Double Descent

Why does the test error decreases with \(N\)?

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

P=5000 Fully Connected 2 hidden layers swish
softhinge MNIST (pca d=10)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

P=5000 Fully Connected 2 hidden layers swish
softhinge MNIST (pca d=10)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

P=5000 Fully Connected 2 hidden layers swish
softhinge MNIST (pca d=10)

Paradoxes

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

We show that the predictor explode at jamming, we predict the exponent

Mario Geiger et al J. Stat. Mech. (2020) 023401

Need to understand what happens at \(N\to \infty\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Neural Tangent Kernel

Space of functions \(X\to Y\)

Paradoxes

Arthur Jacot et al. Neurips (2018)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Neural Tangent Kernel

Manifold \(f:W\to (X\to Y)\)

Paradoxes

Space of functions \(X\to Y\)

\(f(w_0)\)

Arthur Jacot et al. Neurips (2018)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Neural Tangent Kernel

Paradoxes

Space of functions \(X\to Y\)

Manifold \(f:W\to (X\to Y)\)

tangent space

In the limit \(N\to\infty\), parameters change only a little.

Arthur Jacot et al. Neurips (2018)

\(f(w_0)\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Neural Tangent Kernel

Paradoxes

Manifold \(f:W\to (X\to Y)\)

\(\displaystyle \Theta(w, x_1, x_2) = \nabla_w f(w,x_1) \cdot \nabla_w f(w, x_2)\)

tangent space

Arthur Jacot et al. Neurips (2018)

\(f(w_0)\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Neural Tangent Kernel

Paradoxes

Manifold \(f:W\to (X\to Y)\)

\(\displaystyle \frac{d}{dt}f(w, \cdot) \propto \ell_i' \Theta(w, x_i, \cdot)\)

Arthur Jacot et al. Neurips (2018)

\(\displaystyle \Theta(w, x_1, x_2) = \nabla_w f(w,x_1) \cdot \nabla_w f(w, x_2)\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Neural Tangent Kernel

Paradoxes

Manifold \(f:W\to (X\to Y)\)

\(\displaystyle \frac{d}{dt}f(w, \cdot) \propto \ell_i' \Theta(w, x_i, \cdot)\)

Arthur Jacot et al. Neurips (2018)

Well defined behavior at \(N=\infty\)

\(\displaystyle \Theta(w, x_1, x_2) = \nabla_w f(w,x_1) \cdot \nabla_w f(w, x_2)\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

• \(\| \Theta - \langle \Theta \rangle \|^2 \sim N^{-1/2}\)
• \(\Rightarrow \mathrm{Var}(f) \sim N^{-1/2}\)
• \(\Rightarrow \epsilon(N) - \epsilon_\infty \sim N^{-1/2}\)

Mario Geiger et al J. Stat. Mech. (2020) 023401

Paradoxes

Noisy convergence toward a well defined dynamics

Why does the test error decreases with \(N\)?

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Ensemble average

Generalization error of the average of 20 network's outputs

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Noisy convergence toward a well defined dynamics...
...described by the NTK in which the weights do not move!

 

no feature learning?

 

 

Neural network do learn features

Le Quoc V. IEEE (2013)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Mean-Field limit

Mei et al. (2018); Rotskoff and Vanden-Eijnden (2018); Chizat and Bach (2018); Sirignano and Spiliopoulos (2020b); Mei et al. (2019); Nguyen (2019); Sirignano and Spiliopoulos (2020a), with recent development for deeper nets, see e.g. Nguyen and Pham (2020)

Another limit, different than NTK in which the neural network can learn features

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

NTK  vs.  Mean-Field

Q1 How to quantify if the network is rather Mean-Field or NTK?

Q2 Which of the two limits perform better?

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

\(f(w, x) = \frac1{\sqrt{h}} w^{L+1} \phi(\frac1{\sqrt{h}} \dots)\)

Paradoxes

NTK           Mean-Field

\(f(w, x) = \frac1h w^{L+1} \phi(\frac1{\sqrt{h}} \dots)\)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

\(f(w, x) = \frac1{\sqrt{h}} w^{L+1} \phi(\frac1{\sqrt{h}} \dots)\)

Paradoxes

\(f(w, x) = \frac1h w^{L+1} \phi(\frac1{\sqrt{h}} \dots)\)

\(\tilde\alpha (f(w, x) - f(w_0, x))\)

(linearize to allow small weights change to have an impact)

Chizat et al. Neurips (2019)

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

Q1 How to quantify if the network is rather Mean-Field or NTK?

\(\Delta\Theta \equiv \Theta(w) - \Theta(w_0)\)

level sets delimits a smooth transition

NTK

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Paradoxes

• Double descent is also present in feature learning

Landscape vs Parameterization

Data Symmetries

Jamming

Double descent

Phase Diagram

Fully Connected  2 hidden layers

Paradoxes

= number of neurons per layer

Q2 Which of the two limits perform better?

gradient flow

FC: lazy > feature

CNN: feature > lazy

Mario Geiger et al J. Stat. Mech. (2020) 113301

Mario Geiger et al Physics Reports V924 (2021)

Open problem: Not well understood which regime performs better

Lee Jaehoon et al. (2020)

Landscape vs Parameterization

Data Symmetries

Paradoxes

• Why not stuck in a local minima?
• How is the transition between under/over parameterized?
   
• Why does the test error decreases with \(N\)?
   
• Q1 How to quantify if the network is rather Mean-Field or NTK?
• Q2 Which of the two limits perform better?

• Crank-up \(N\)
• There is sharp jamming transition, in the same universality class as ellipses
   
• Noisy convergence toward a well defined dynamics
   

• Change in NTK delimits a smooth transition controlled by \(\tilde\alpha\)

• FC: lazy > feature

  CNN: feature > lazy

  still an open problem

Landscape vs Parameterization

Data Symmetries

Data Symmetries

 

Curse of dimensionality not answered so far

 

One needs to take into account the structure of the data.

 

Data contains symmetries

Paradoxes

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

Is stability to diffeomorphisms responsible for beating the curse of dimensionality?

Aim: measure the stability toward diffeomorphisms

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Maximum entropy distribution

Euclidean Neural Network

Paradoxes

diffeomorphisms are controlled by a temperature parameter

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Maximum entropy distribution

Euclidean Neural Network

Paradoxes

diffeomorphisms are controlled by a temperature parameter

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Maximum entropy distribution

Euclidean Neural Network

Paradoxes

diffeomorphisms are controlled by a temperature parameter

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Maximum entropy distribution

Euclidean Neural Network

Paradoxes

diffeomorphisms are controlled by a temperature parameter

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

\(D_f \propto |f(\tau x) - f(x)|\)

• average over inputs
• average over diffeomorphisms of the same temperature
• normalized by the amplitude of \(f\)

Stability to diffeomorphisms

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

\(\displaystyle R_f = \frac{D_f}{G_f}\)

Euclidean Neural Network

Paradoxes

Leonardo Petrini et al arxiv 2105.02468

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

(Relative) stability to diffeomorphisms is key to beat the curse of dimensionality

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Euclidean Neural Network

Paradoxes

Can we generalize CNN invariance to other symmetries?

So far we looked at 2D images

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

Equivariance

\(f(w, {\color{blue}D(g)} x) = {\color{blue}D'(g)} f(w, x)\)

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

Daniel Worrall et al CVPR (2017)

CNN are not equivariant to rotations

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

slide taken from Tess Smidt

Paradoxes

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

CNN are easy to implement, usual machine learning libraries provide functions to do so

Equivariant Neural Network for 3D rotations are less easy to implement and requires specialized functions

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

Paradoxes

slide taken from Tess Smidt

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

e3nn

 

open source library

 

python

 

Efficient code for:

• spherical harmonics
• tensor product

Paradoxes

between any irreps of O(3)

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

e3nn

is modular and flexible

The same building blocks allow to implement

 

• message passing on graph
• voxel convolution
• SE(3)-Transformers
• SphericalCNN

Paradoxes

Cohen, Geiger, et al (1801.10130)

 Fabian Fuchs et al (2006.10503)

Weiler, Geiger et al (1807.02547)

Tess Smidt et al (1802.08219)

Landscape vs Parameterization

Data Symmetries

Diffeomorphisms

Euclidean Neural Network

One example of results obtained using e3nn

Ab initio Molecular dynamics

Paradoxes

Simon Batzner et al (2021) arxiv 2101.03164

Landscape vs Parameterization

Data Symmetries

Paradoxes

Is stability to diffeomorphisms responsible for beating the curse of dimensionality?

• We see a strong correlation between the relative stability and the test error
• It is acquired during training
   
• How is it learned?
• Is there a bound between \(\epsilon\) and \(R_f\)?

Can we generalize CNN invariance to other symmetries?

• e3nn is equivariant to translations, rotations and mirror
• Why equivariant neural network perform better than invariant ones?

Thank you for your time