Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

ECE586BH Lecture1

Download as pdf or txt
Download as pdf or txt
You are on page 1of 36

ECE586BH: Interplay between Control and

Machine Learning

Bin Hu
ECE , University of Illinois Urbana-Champaign

Lecture 1, Fall 2023


Feedback Control Machine learning

• dynamical systems • statistics/optimization


• robustness • large-scale (big data)
• safety-critical • performance-driven
• model-based design • train using data
• CDC/ACC/ECC • NeurIPS/ICML/ICLR

2
Artificial Intelligence Revolution

Safety-critical applications!
3
Flight Control Certification

Ref: J. Renfrow, S. Liebler, and J. Denham. “F-14 Flight Control Law Design,
Verification, and Validation Using Computer Aided Engineering Tools,” 1996.
4
Feedback Control Machine learning

Unified and automated tools for a repeatable and trustable


design process of next generation intelligent systems

5
Example: Robustness is crucial!
• Deep learning: Small adversarial perturbations can fool the classifier!

• Optimization: The oracle can be inexact! xk+1 = xk − α(∇f (xk )+ek )


• Decision and control: Model uncertainty and sim-to-real gap matter!

6
Control for Learning
Control theory addresses unified analysis and design of dynamical systems.

LTI systems Markov jump systems Lur’e systems

ξk+1 = Aξk + Buk ξk+1 = Aik ξk + Bik uk ξk+1 = Aξk + Bφ(Cξk )

 T
AT P B

Pn A PA − P
AT P A − P ≺ 0 T
i=1 pij Ai Pi Ai ≺ Pj ≺M
BTP A BTP B

Pros: Unified testing conditions when problem parameters are changed.


Cons: For control, we only need to solve the conditions numerically.
Control for learning: Algorithms and networks treated as control systems
• Neural networks as generalized Lur’e systems
• Stochastic learning algorithms as generalized Lur’e systems

Key message: Robustness can be addressed in a unified manner!


7
Learning for Control
Control theory addresses unified analysis and design of dynamical systems.

LTI systems MJLS Lur’e systems

ξk+1 = Aξk + Buk ξk+1 = Aik ξk + Bik uk ξk+1 = Aξk + Bφ(Cξk )

 T
AT P B

T
Pn T A PA − P
A PA − P ≺ 0 i=1 pij Ai Pi Ai ≺ Pj ≺M
BTP A BTP B

Many control design methods rely on convex conditions (BGFB1994).


What about problems that cannot be formulated as convex optimization?
• Direct policy search (e.g. min J(K)) is nonconvex!

Learning for control: Tailoring nonconvex learning theory to push robust control
theory beyond the convex regime
8
Outline

• Control for Learning


• Control methods for certifiably robust neural networks
• A control perspective on stochastic learning algorithms

• Learning for Control


• Global convergence of direct policy search on robust control

9
Robust Control Theory
- ∆

v w

P 

1. Approximate the true system as “a linear system + a perturbation”


2. ∆ can be a troublesome element: nonlinearity, uncertainty, or delays
3. Rich control literature including standard textbooks
• Zhou, Doyle, Glover, “ Robust and optimal control,” 1996
4. Many tools: small gain, passivity, dissipativity, Zames-Falb multipliers, etc
5. The integral quadratic constraint (IQC) framework [Megretski, Rantzer
(TAC1997)] provides a unified analysis for “ LTI P + troublesome ∆”
6. Recently, IQC analysis has been extended for more general P
7. Typically, the stability is tested by a SDP condition
10
Quadratic Constraints from Robust Control
• Lur’e system: ξk+1 = Aξk + B∆(Cξk ).
• EX: Gradient method (A = I, B = −αI, C = I, and ∆ = ∇f )
• Question: How to prove that the above Lur’e system converges? We are
looking at the following set of coupled sequences {ξk , wk , vk }

{(ξ, w, v) : ξk+1 = Aξk + Bwk , vk = Cξk } ∩ {(ξ, w, v) : wk = ∆(vk )}


• Key idea: Quadratic constraints! Replace the troublesome nonlinear
element ∆ with the following quadratic constraint:
(  T   )
vk vk
{(v, w) : wk = ∆(vk )} ⊂ (v, w) : M ≤0 ,
wk wk
where M is constructed from the property of ∆.
• If we can show that any sequence from the set below converges,
(  T   )
vk vk
(ξ, w, v) : ξk+1 = Aξk + Bwk , vk = Cξk , M ≤0 ,
wk wk
then we are done.
11
Quadratic Constraints from Robust Control
Now we are analyzing the sequence from the following set:
(  T   )
vk vk
(ξ, w, v) : ξk+1 = Aξk + Bwk , vk = Cξk , M ≤0
wk wk

Theorem
If there exists a positive definite matrix P and 0 < ρ < 1 s.t.
 T T
A P A − ρ2 P AT P B
   
C 0 C 0
 M
BTP A BTP B 0 I 0 I
T
then ξk+1 P ξk+1 ≤ ρ2 ξkT P ξk and limk→∞ ξk = 0.
T  T   T  T
A P A − ρ2 P AT P B
    
ξk ξk ξ C 0 C 0 ξk
≤ k M
wk BTP A BTP B wk wk 0 I 0 I wk
| {z } | {z }
T TP ξ T 
P ξk+1 −ρ2 ξk
 
ξk+1 k
v v
 k  M  k ≤0
wk wk
This condition is a semidefinite program (SDP) problem!
12
Illustrative Example: Gradient Descent Method
• Rewrite the gradient method xk+1 = xk − α∇f (xk ) as:
(xk+1 − x? ) = (xk − x? ) − α ∇f (xk )
| {z } | {z } | {z }
ξk+1 ξk wk

• If f is L-smooth and m-strongly convex, then by co-coercivity:


T 
x − x? −(m + L)I x − x?
  
2mLI
≤0
∇f (x) −(m + L)I 2I ∇f (x)
| {z }
M

• We have A = I, B = −αI, C = I, and the following SDP


 T T 
A P A − ρ2 P A T P B
  
C 0 C 0
 M
BTP A BTP B 0 I 0 I

(1 − ρ2 )p −αp
   
• This leads to −2mL m + L
+ ⊗I 0
−αp α2 p m+L −2
• Choose (α, ρ, p) to be ( L1 , 1 − m 2 2 L−m 1 2
L , L ) or ( L+m , L+m , 2 (L + m) ) to
? ?
recover standard rates, i.e. kxk+1 − x k ≤ (1 − m/L)kxk − x k
• For this proof, is strong convexity really needed? No! Regularity condition!
13
Illustrative Example: Gradient Descent Method
• We have shown kxk+1 − x? k ≤ (1 − m/L)kxk − x? k
• Is it a contraction, i.e. kxk+1 − x0k+1 k ≤ (1 − m/L)kxk − x0k k?
• (xk+1 − x0k+1 ) = (xk − x0k ) − α (∇f (xk ) − ∇f (x0k ))
| {z } | {z } | {z }
ξk+1 ξk wk

• If f is L-smooth and m-strongly convex, then by co-coercivity:


T 
x − x0 x − x0
  
2mLI −(m + L)I
≤0
∇f (x) − ∇f (x0 ) −(m + L)I 2I ∇f (x) − ∇f (x0 )
| {z }
M

• We have A = I, B = −αI, C = I, and the same SDP


(1 − ρ2 )p −αp
   
−2mL m + L
+ ⊗I 0
−αp α2 p m+L −2

• Choose (α, ρ, p) to be ( L1 , 1 − m 2
L,L )
2
or ( L+m , L−m 1 2
L+m , 2 (L + m) ) to give
the contraction result!
• For this proof, is strong convexity really needed? Yes!
14
Outline

• Control for Learning


• Control methods for certifiably robust neural networks
• A control perspective on stochastic learning algorithms

• Learning for Control


• Global convergence of direct policy search on robust control

15
Deep Learning for Classification
Deep learning has revolutionized the fields of AI and computer vision!
• Input space X ⊂ Rd to a label space Y := {1, . . . , H}.
• Predict labels from image pixels
• Neural network classifier function f := (f1 , . . . , fH ) : X → RH such that
the predicted label for an input x is arg maxj fj (x).
• Input-label (x, y) is correctly classified if arg maxj fj (x) = y.

16
Deep Learning Models

Deep learning models: f (x) = xD+1 and x0 = x


• Feedforward: xk+1 = σ(Wk xk + bk ) for k = 0, 1, · · · , D
• Residual network: xk+1 = xk − σ(Wk xk + bk ) for k = 0, 1, · · · , D
• Many other structures: transformers, etc
Deep learning models are expressive and generalize well, achieving state-of-the-art
results in computer vision and natural language processing. However, ...
17
Adversarial Attacks and Robustness

• For correct labels (i.e. arg maxj fj (x) = y), one may find kτ k ≤ ε s.t.
arg maxj fj (x + τ ) 6= y (small perturbation lead to wrong prediction)
• Small perturbation can fool modern deep learning models!
• How to deploy deep learning models into safety-critical applications?
• Certified robustness: A classifier f is certifiably robust at radius ε ≥ 0 at
point x with label y if for all τ such that kτ k ≤ ε : arg maxj fj (x + τ ) = y

18
1-Lipschitz Networks for Certified Robustness
• Tsuzuku, Sato, Sugiyama (NeurIPS2018): Let f be L-Lipschitz. If we have

Mf (x) := max(0, fy (x) − max
0
f y 0 (x)) > 2Lε
y 6=y

then we have for every τ such that kτ k2 ≤ ε: arg maxj fj (x + τ ) = y



• Perturbation smaller than Mf (x)/ 2L cannot deceive f for datapoint x!
• If each layer of a network is 1-Lipchitz, the entire network is 1-Lipschitz.

• For each data point, we test whether Mf (x) > 2ε, and then count the
percentage of data points that is guaranteed to be guarded for perturbation
smaller than ε (which is the certified accuracy for that ε).
• We need to train a Lipschitz neural network with good prediction margins!
Previous approaches:
 
• Spectral normalization (MKKY2018): xk+1 = σ 1
kWk k2 Wk xk + bk
• Orthogonality (TK2021, SF2021): xk+1 = σ(Wk xk + bk ) with WkT Wk = I
• Convex potential layer (MDAA2022): xk+1 = xk − kW2 k2 Wk σ(WkT x + bk )
k 2
1

• AOL (PL2022): xk+1 = σ(Wk diag( j |Wk Wk |ij ) 2 xk + bk )
T
P
19
My Focus: Principles for 1-Lipschitz Networks
Theorem (AHDAH2023)
If there exists nonsingular diagonal Tk s.t. WkT Wk  Tk , then we have
−1
1. The layer xk+1 = σ(Wk Tk 2 xk + bk ) is 1-Lipschitz for any 1-Lipschitz σ.
2. The layer xk+1 = xk − 2Wk Tk−1 σ(WkT x + bk ) is 1-Lipschitz if σ is ReLU.
−1 −1 −1
kxk+1 − x0k+1 k2 ≤ kWk Tk 2 (xk − x0k )k2 = (xk − x0k )T Tk 2 WkT Wk Tk 2 (xk − x0k )
| {z }
≤kxk −x0k k2

The second statement can be proved using the quadratic constraint argument.

A Unification of Existing 1-Lipschitz Neural Networks


• Spectral normalization: Statement 1 with Tk = kWk k22 I
• Orthogonal weights: Statement 1 with Tk = I and WkT Wk = I
• CPL: Statement 2 with Tk = kWk k22 I
• AOL: Statement 1 with Tk = diag( nj=1 |WkT Wk |ij )
P

• Control Theory (SLL): Tk = diag( nj=1 |WkT Wk |ij qj /qi ).


P
20
Experimental Results
4 versions of SDP-based Lipchitz Network (SLL) (S, M, L, XL)

Natural Provable Accuracy (ε)


Datasets Models
Accuracy 36 72 108
1
255 255 255

Cayley Large 43.3 29.2 18.8 11.0 -


SOC 20 48.3 34.4 22.7 14.2 -
SOC+ 20 47.8 34.8 23.7 15.8 -
CPL XL 47.8 33.4 20.9 12.6 -
AOL Large 43.7 33.7 26.3 20.7 7.8
CIFAR100
SLL Small 45.8 34.7 26.5 20.4 7.2
SLL Medium 46.5 35.6 27.3 21.1 7.7
SLL Large 46.9 36.2 27.9 21.6 7.9
SLL X-Large 47.6 36.5 28.2 21.8 8.2

• Competitive results over CIFAR100 and TinyImageNet


• Many extensions: Lipschitz deep equilibrium models, neural ODEs, etc
21
Quadratic Constraints for Lipschitz Networks
• Residual network: xk+1 = xk − Gk σ(WkT xk + bk ) for k = 0, 1, · · · , D.
• 1-Lipschitz layer: How to enforce kxk+1 − x0k+1 k ≤ kxk − x0k k?
• (xk+1 − x0k+1 ) = (xk − x0k ) −Gk σ(WkT xk + bk ) − σ(WkT x0k + bk )

| {z } | {z } | {z }
ξk+1 ξk wk

• We will use the property of σ to construct Mk such that we only need to


look at the following set with Ak = I and Bk = −Gk :
(  T   )
ξk ξk
(ξ, w) : ξk+1 = Ak ξk + Bk wk , Mk ≤0 ,
wk wk

• Then we can ensure kξk+1 k ≤ kξk k via enforcing a SDP for the set:

 T  T  T
AT AT
  
Ak P Ak − P k P Bk  M =⇒ ξk Ak Ak − I k Bk ξk
k |{z} ≤0
BkT P Ak BkT P Bk wk BkT Ak BkT Bk wk
P =I | {z }
kξk+1 k2 −kξk k2 =kxk+1 −x0k+1 k2 −kxk −x0k k2

22
Quadratic Constraints for Lipschitz Networks
• Since σ is slope-restricted on [0, 1], the following scalar-version incremental
quadratic constraint holds with m = 0 and L = 1:
T 
a − a0 a − a0
  
2mL −(m + L)
≤0
σ(a) − σ(a0 ) −(m + L) 2 σ(a) − σ(a0 )
| {z }
 

0 −1 
−1 2

• The vector-version quadratic constraint: For diagonal Γk  0, we have


T 
vk − vk0 vk − vk0
  
0 −Γk
≤0
σ(vk ) − σ(vk0 ) −Γk 2Γk σ(vk ) − σ(vk0 )
| {z }
Xk

• Choosing vk = WkT xk + bk and vk0 = WkT x0k + bk , we have


T
WkT (xk − x0k ) WkT (xk − x0k )
  
Xk ≤0
σ(WkT xk + bk ) − σ(WkT x0k + bk ) σ(WkT xk + bk ) − σ(WkT x0k + bk )
23
Quadratic Constraints for Lipschitz Networks
T
−Γk WkT
      
ξ ξk Wk 0 0 0
• We get k Mk ≤ 0 with Mk =
wk wk 0 I −Γk 2Γk 0 I
| {z }
 
0 −Wk Γk 
−Γk WkT

2Γk

Theorem
If there exists diagonal Γk  0 such that
   
0 −Gk 0 −Wk Γk

−GT k GTk Gk −Γk WkT 2Γk

then the residual network xk+1 = xk − Gk σ(WkT xk + bk ) is 1-Lipschitz.

• Analytical solution: Gk = Wk Γk and Γk WkT Wk Γk  2Γk .


• Suppose Γk is nonsingular, and Tk = 2Γ−1k . Then the residual network
xk+1 = xk − 2Wk Tk−1 σ(WkT xk + bk ) is 1-Lipschitz as long as Tk  WkT Wk
• Ref: Araujo, Havens, Delattre, Allauzen, H.. A unifying algebraic
perspective on Lipschitz neural networks, ICLR, 2023. (Spotlight) 24
Outline

• Control for Learning


• Control Methods on Certifiably Robust Neural Networks
• A Control Perspective on Stochastic Learning Algorithms

• Learning for Control


• Global convergence of direct policy search on robust control

25
History: Computer-Assisted Proofs in Optimization
In the past ten years, much progress has been made in leveraging SDPs to assist
the convergence rate analysis of optimization methods.
• Drori and Teboulle (MP2014): numerical worst-case bounds via the
performance estimation problem (PEP) formulation
• Lessard, Recht, Packard (SIOPT2016): numerical linear rate bounds using
integral quadratic constraints (IQCs) from robust control theory
• Taylor, Hendrickx, Glineur (MP2017): interpolation conditions for PEPs
• H., Lessard (ICML2017): first SDP-based analytical proof for Nesterov’s
accelerated rate
• H., Seiler, Ranzter (COLT2017): first paper on SDP-based convergence
proofs for stochastic optimization using jump system theory and IQCs
• Van Scoy, Freeman, and Lynch (LCSS2017): first paper on control-oriented
design of accelerated methods: triple momentum
Taken further by different groups
• inexact gradient methods, proximal gradient methods, conditional gradient
methods, operator splitting methods, mirror descent methods, distributed
gradient methods, monotone inclusion problems 26
Stochastic Methods for Machine Learning
• Many learning tasks (regression/classification) lead to finite-sum ERM
n
1X
minp fi (x)
x∈R n i=1
where fi (x) = li (x) + λR(x) (li is the loss, and R avoids over-fitting).
• Stochastic gradient descent (SGD): xk+1 = xk − α∇fik (xk )
• Inexact oracle: xk+1 = xk − α(∇fik (xk ) + ek ) where kek k ≤ δk∇fik (xk )k
(the angle θ between (ek + ∇fik (xk )) and ∇fik (xk ) satisfies | sin(θ)| ≤ δ)
• Algorithm change: SAG (SRF2017) vs. SAGA (DBL2014)
n
!
k+1 k ∇fik (xk ) − yikk 1X k
SAG: x =x −α + y
n n i=1 i
n
!
k+1 k k k 1X k
SAGA: x = x − α ∇fik (x ) − yik + y
n i=1 i

∇fi (xk ) if i = ik

k+1
where yi :=
yik otherwise
• Markov assumption: In reinforcement learning, {ik } can be Markovian 27
My Focus: Unified Analysis of Stochastic Methods
Assumption




• fi smooth, f RSI 





• ik is IID or Markovian 




 Bound

• Oracle is exact or inexact 



• many other possibilities  • Ekxk − x? k2 ≤ c2 ρk + O(α)



 • Ekxk − x? k2 ≤ c2 ρk

Method

• Other forms





• SGD 





• SAGA-like methods 





• Temporal difference learning
How to automate rate analysis of stochastic learning algorithms? Use
numerical semidefinite programs to support search for analytical proofs?

assumption + method =⇒ bound


28
My Focus: Stochastic Methods for Learning
In the deterministic setting, we just need to show that the trajectories generated
by optimization methods belong to the following set:
(  T   )
vk vk
(ξ, w, v) : ξk+1 = Aξk + Bwk , vk = Cξk , Mj ≤ Λj , j ∈ Π
wk wk

What to do for stochastic optimization (e.g. xk+1 = xk − α∇fik (xk ) where


ik ∈ {1, · · · , n} is sampled)?
• Stochastic quadratic constraints: Show that the trajectories generated by
stochastic optimization methods belong to the following set:
(  T   )
vk vk
(ξ, w, v) : ξk+1 = Aξk + Bwk , vk = Cξk , E Mj ≤ Λj , j ∈ Π
wk wk

• Jump system approach: Show that the trajectories generated by stochastic


optimization methods belong to the following set:
(  T   )
vk vk
(ξ, w, v) : ξk+1 = Aik ξk + Bik wk , vk = Cik ξk , Mj ≤ Λj , j ∈ Π
wk wk

where Aik ∈ {A1 , · · · , An }, Bik ∈ {B1 , · · · , Bn }, and Cik ∈ {C1 , · · · , Cn }


29
Stochastic Quadratic Constraints
Suppose we can show that the trajectories generated by stochastic optimization
methods belong to the following set:
(  T   )
vk vk
(ξ, w, v) : ξk+1 = Aξk + Bwk , vk = Cξk , E Mj ≤ Λj , j ∈ Π
wk wk

Theorem
If there exists a positive definite matrix P , non-negative λj and 0 < ρ < 1 s.t.
 T T
A P A − ρ2 P AT P B
 X   
C 0 C 0
 λj Mj
BTP A BTP B 0 I 0 I
j∈Π

T
P ξk+1 ≤ ρ2 EξkT P ξk +
P
then Eξk+1 j∈Π λj Λj .
T   X  T
AT P A − ρ 2 P AT P B
   
ξk ξk vk v
≤ λj Mj k
wk BTP A BTP B wk wk wk
| {z } j∈Π
T
ξk+1 TP ξ
P ξk+1 −ρ2 ξk k

Then take expectation and apply the expected quadratic constraints!


30
Main Result: Analysis of Biased SGD
• Consider xk+1 = xk − α(∇fik (xk ) + ek ) with kek k2 ≤ δ 2 k∇fik (xk )k2 + c2
• If c = 0, the bound means the angle θ between (ek + ∇fik (xk )) and
∇fik (xk ) satisfies | sin(θ)| ≤ δ
 
• Rewritten as (xk+1 − x? ) = (xk − x? ) + −αI −αI ∇fik (xk )
 
| {z } | {z } ek
ξk+1 ξk | {z }
wk

• Assume the restricted secant inequality ∇f (x) (x − x ) ≥ mkx − x? k2


T ?

• Assume fi is L-smooth, i.e. k∇fi (x) − ∇fi (x? )k ≤ Lkx − x? k


T 
xk − x? xk − x?
  
2mI −I 0
• 1st QC: E ∇fik (xk )  −I 0 0 ∇fik (xk ) ≤ |{z}0
ek 0 0 0 ek Λ1
| {z }
M1

?
T 
−2L2 I 0 xk − x?
  
xk − x 0 n
2X
• 2nd QC: E ∇fik (xk )  0 I 0 ∇fik (xk ) ≤ k∇fi (x? )k2
n i=1
ek 0 0 0 ek
| {z } | {z }
M2 Λ2
31
Main Result: Analysis of Biased SGD
• We can rewrite kek k2 ≤ δ 2 k∇fik (xk )k2 + c2 as
T 
xk − x? xk − x?
  
0 0 0
E ∇fik (xk ) 0 −δ 2 I c2
0 ∇fik (xk ) ≤ |{z}
ek 0 0 I ek Λ3
| {z }
M3

• We have A = I, B = −αI
 
−αI , C = I, and the following SDP
 T 3 T
A P A − ρ2 P AT P B
 X   
C 0 C 0
 λ j Mj
BTP A BTP B 0 I 0 I
j=1

• Biased SGD satisfies Ekxk+1 − x? k2 ≤ ρ2 Ekxk − x? k2 + λ2 Λ2 + λ3 c2 if


1 − ρ2
     2 
−α −α −2m 1 0 2L 0 0
 −α α 2 − δ 2 λ3 α 2  + λ1  1 0 0 + λ2  0 −1 0  0
2 2
−α α α + λ3 0 0 0 0 0 0
32
Main Result: Analysis of Biased SGD
• Given Ekx0 − x? k2 ≤ U0 , set Uk+1 = min(ρ2 Uk + λ2 Λ2 + λ3 c2 ) with
1 − ρ2
     2 
−α −α −2m 1 0 2L 0 0
 −α α 2 − δ 2 λ3 α 2  + λ1  1 0 0 + λ2  0 −1 0  0
2 2
−α α α + λ3 0 0 0 0 0 0
then we have Ekxk − x? k2 ≤ Uk . This leads to a sequential SDP problem.
• This problem has an exact solution
 p p 2
Uk+1 = α c2 + δ 2 Λ2 + 2L2 δ 2 Uk + (1 − 2mα + 2L2 α2 )Uk + Λ2 α2

c2 +δ 2 Λ2 m(c2 (2L2 −m2 )+(1−δ 2 )Λ2 m2 )


• limk→∞ Uk = m2 −2δ 2 L2 + (m2 −2δ 2 L2 )2 α + O(α2 )
2 2 2
• Rate = 1 − m −2δ m
L
α + O(α2 )
• For different assumptions, modify (Mj , Λj )!
• H., Seiler, and Lessard. Analysis of biased stochastic gradient descent using
sequential semidefinite programs. Mathematical Programming, 2021
• Syed, Dall’Anese, H.. Bounds for the tracking error and dynamic regret of
inexact online optimization methods: A unified analysis via sequential SDPs.
33
Jump System Approach
n  T
1 X AT 2
AT
 X   
i P Ai − ρ P i P Bi  λj
C 0
Mj
C 0
T
n i=1 B i P Ai BiT P Bi 0 I 0 I
j∈Π

Pros:
• General enough to handle many algorithms: H., Seiler, Rantzer (COLT2017)
Method Ãik B̃ik C̃
" #
eik eTik
 
In − eik eTik 0̃  T 
SAGA 0̃ 1
−α
n (e − neik )
T
1 −αeTik
" #
eik eTik
 
In − eik eTik 0̃  T 
SAG 0̃ 1
−αn (e − eik )
T
1 −αn eik
T

• General enough to handle Markov {ik }: Syed and H. (NeurIPS2019), Guo


and H. (ACC2022a,2022b)
Cons:
• SDPs are much bigger than the ones obtained from stochastic quadratic
constraints, and we have to exploit SDP structures for simplifications
34
Control for Learning: Summary

• Iterative learning algorithms and neural network layers can be thought as


feedback control systems.

• The quadratic constraint approach from control theory can be leveraged to


formulate SDP conditions for machine learning research.

• Different from the study in control, now we want to obtain analytical


solutions of the SDPs!

35
Outline

• Control for Learning


• Control methods on certifiably robust neural networks
• A control perspective on stochastic learning algorithms

• Learning for Control


• Global convergence of direct policy search on robust control

36

You might also like