Training hopfield nets
Geometric approach
-
W = Y Y T − N p I \mathbf{W}=\mathbf{Y} \mathbf{Y}^{T}-N_{p} \mathbf{I} W=YYT−NpI
-
E ( y ) = y T W y \mathbf{E}(\mathbf{y})=\mathbf{y}^{T} \mathbf{W y} E(y)=yTWy
-
Sine : y T ( Y Y T − N p I ) y = y T Y Y T y − N N p \mathbf{y}^{T}\left(\mathbf{Y} \mathbf{Y}^{T}-N_{p} \mathbf{I}\right) \mathbf{y}=\mathbf{y}^{T} \mathbf{Y} \mathbf{Y}^{T} \mathbf{y}-N N_{p} yT(YYT−NpI)y=yTYYTy−NNp
-
So W is identical to behavior with W = Y Y T \mathbf{W}=\mathbf{Y} \mathbf{Y}^{T} W=YYT
- Energy landscape only differs by an additive constant
- Have the same eigen vectors
-
A pattern y p y_p yp is stored if:
- sign ( W y p ) = y _ p \operatorname{sign}\left(\mathbf{W} \mathbf{y}_{p}\right)=\mathbf{y}\_{p} sign(Wyp)=y_p for all target patterns
-
Training: Design W W W such that this holds
-
Simple solution: y p y_p yp is an Eigenvector of W W W
Storing k orthogonal patterns
- Let
Y
=
[
y
_
1
y
_
2
…
y
_
K
]
\mathbf{Y}=\left[\mathbf{y}\_{1} \mathbf{y}\_{2} \ldots \mathbf{y}\_{K}\right]
Y=[y_1y_2…y_K]
- W = Y Λ Y T \mathbf{W}=\mathbf{Y} \Lambda \mathbf{Y}^{T} W=YΛYT
- λ 1 , . . . , λ k \lambda_1,...,\lambda_k λ1,...,λk are positive
- for λ 1 = λ 2 = λ k = 1 \lambda_1= \lambda_2=\lambda_k= 1 λ1=λ2=λk=1 this is exactly the Hebbian rule
- Any pattern
y
y
y can be written as
- y = a 1 y 1 + a 2 y 2 + ⋯ + a N y N \mathbf{y}=a_{1} \mathbf{y}_{1}+a_{2} \mathbf{y}_{2}+\cdots+a_{N} \mathbf{y}_{N} y=a1y1+a2y2+⋯+aNyN
- W y = a 1 W y 1 + a 2 W y 2 + ⋯ + a N W y N = y \mathbf{W y}=a_{1} \mathbf{W y}_{1}+a_{2} \mathbf{W y}_{2}+\cdots+a_{N} \mathbf{W y}_{N} = y Wy=a1Wy1+a2Wy2+⋯+aNWyN=y
- All patterns are stable
- Remembers everything
- Completely useless network
- Even if we store fewer than
N
N
N patterns
- Let Y = [ y _ 1 y _ 2 … y _ K r _ K + 1 r _ K + 2 … r _ N ] Y=\left[\mathbf{y}\_{1} \mathbf{y}\_{2} \ldots \mathbf{y}\_{K} \mathbf{r}\_{K+1} \mathbf{r}\_{K+2} \ldots \mathbf{r}\_{N}\right] Y=[y_1y_2…y_Kr_K+1r_K+2…r_N]
- W = Y Λ Y T W=Y \Lambda Y^{T} W=YΛYT
- r _ K + 1 r _ K + 2 … r _ N \mathbf{r}\_{K+1} \mathbf{r}\_{K+2} \ldots \mathbf{r}\_{N} r_K+1r_K+2…r_N are orthogonal to y 1 y 2 … y K \mathbf{y}_1 \mathbf{y}_2 \ldots \mathbf{y}_K y1y2…yK
- λ 1 = λ 2 = λ k = 1 \lambda_1= \lambda_2=\lambda_k= 1 λ1=λ2=λk=1
- Problem arise because eigen values are all 1.0
- Ensures stationarity of vectors in the subspace
- All stored patterns are equally important
General (nonorthogonal) vectors
- w j i = ∑ p ∈ { p } y i p y j p w_{j i}=\sum_{p \in\{p\}} y_{i}^{p} y_{j}^{p} wji=∑p∈{p}yipyjp
- The maximum number of stationary patterns is actually exponential in N N N (McElice and Posner, 84’)
- For a specific set of
K
K
K patterns, we can always build a network for which all
K
K
K patterns are stable provided
k
≤
N
k \le N
k≤N
- But this may come with many “parasitic” memories
Optimization
- Energy function
- E = − 1 2 y T W y − b T y E=-\frac{1}{2} \mathbf{y}^{T} \mathbf{W} \mathbf{y}-\mathbf{b}^{T} \mathbf{y} E=−21yTWy−bTy
- This must be maximally low for target patterns
- Must be maximally high for all other patterns
- So that they are unstable and evolve into one of the target patterns
- Estimate
W
W
W such that
- E E E is minimized for y 1 , . . . , y P y_1,...,y_P y1,...,yP
- E E E is maximized for all other y y y
- Minimize total energy of target patterns
- E ( y ) = − 1 2 y T W y W ^ = argmin W ∑ y ∈ Y P E ( y ) E(\mathbf{y})=-\frac{1}{2} \mathbf{y}^{T} \mathbf{W y} \quad \widehat{\mathbf{W}}=\underset{\mathbf{W}}{\operatorname{argmin}} \sum_{\mathbf{y} \in \mathbf{Y}_{P}} E(\mathbf{y}) E(y)=−21yTWyW =Wargmin∑y∈YPE(y)
- However, might also pull all the neighborhood states down
- Maximize the total energy of all non-target patterns
- E ( y ) = − 1 2 y T W y E(\mathbf{y})=-\frac{1}{2} \mathbf{y}^{T} \mathbf{W} \mathbf{y} E(y)=−21yTWy
- W ^ = argmin W ∑ y ∈ Y P E ( y ) − ∑ y ∉ Y P E ( y ) \widehat{\mathbf{W}}=\underset{\mathbf{W}}{\operatorname{argmin}} \sum_{\mathbf{y} \in \mathbf{Y}_{P}} E(\mathbf{y})-\sum_{\mathbf{y} \notin \mathbf{Y}_{P}} E(\mathbf{y}) W =Wargmin∑y∈YPE(y)−∑y∈/YPE(y)
- Simple gradient descent
-
W = w + η ( ∑ y ∈ Y P y y T − ∑ y ∉ Y P y y T ) \mathbf{W}=\mathbf{w}+\eta\left(\sum_{\mathbf{y} \in \mathbf{Y}_{P}} \mathbf{y} \mathbf{y}^{T}-\sum_{\mathbf{y} \notin \mathbf{Y}_{P}} \mathbf{y} \mathbf{y}^{T}\right) W=w+η(∑y∈YPyyT−∑y∈/YPyyT)
-
minimize the energy at target patterns
-
raise all non-target patterns
- Do we need to raise everything?
-
Raise negative class
-
Focus on raising the valleys
- If you raise every valley, eventually they’ll all move up above the target patterns, and many will even vanish
-
How do you identify the valleys for the current W W W?
-
Initialize the network randomly and let it evolve
-
It will settle in a valley
-
Should we randomly sample valleys?
- Are all valleys equally important?
- Major requirement: memories must be stable
- They must be broad valleys
-
Solution: initialize the network at valid memories and let it evolve
- It will settle in a valley
- If this is not the target pattern, raise it
-
What if there’s another target pattern downvalley
-
no need to raise the entire surface, or even every valley
- Raise the neighborhood of each target memory
Storing more than N patterns
- Visible neurons
- The neurons that store the actual patterns of interest
- Hidden neurons
- The neurons that only serve to increase the capacity but whose actual values are not important
- The maximum number of patterns the net can store is bounded by the width N N N of the patterns…
- So lets pad the patterns with
K
K
K “don’t care” bits
- The new width of the patterns is N + K N+K N+K
- Now we can store N + K N+K N+K patterns!
- Taking advantage of don’t care bits
- Simple random setting of don’t care bits, and using the usual training and recall strategies for Hopfield nets should work
- However, to exploit it properly, it helps to view the Hopfield net differently: as a probabilistic machine
A probabilistic interpretation
- For binary y the energy of a pattern is the analog of the negative log likelihood of a Boltzmann distribution
- Minimizing energy maximizes log likelihood
- E ( y ) = − 1 2 y T W y P ( y ) = Cexp ( − E ( y ) ) E(\mathbf{y})=-\frac{1}{2} \mathbf{y}^{T} \mathbf{W y} \quad P(\mathbf{y})=\operatorname{Cexp}(-E(\mathbf{y})) E(y)=−21yTWyP(y)=Cexp(−E(y))
Boltzmann Distribution
- E ( y ) = − 1 2 y T W y − b T y E(\mathbf{y})=-\frac{1}{2} \mathbf{y}^{T} \mathbf{W} \mathbf{y}-\mathbf{b}^{T} \mathbf{y} E(y)=−21yTWy−bTy
- P ( y ) = Cexp ( − E ( y ) k T ) P(\mathbf{y})=\operatorname{Cexp}\left(\frac{-E(\mathbf{y})}{k T}\right) P(y)=Cexp(kT−E(y))
- C = 1 ∑ y exp ( − E ( y ) k T ) C=\frac{1}{\sum_{\mathrm{y}} \exp \left(\frac{-E(\mathbf{y})}{k T}\right)} C=∑yexp(kT−E(y))1
- k k k is the Boltzmann constant, T T T is the temperature of the system
- Optimizing
W
W
W
- E ( y ) = − 1 2 y T W y W ^ = argmin W ∑ y ∈ Y P E ( y ) − ∑ y ∉ Y P E ( y ) E(\mathbf{y})=-\frac{1}{2} \mathbf{y}^{T} \mathbf{W} \mathbf{y} \quad \widehat{\mathbf{W}}=\underset{\mathbf{W}}{\operatorname{argmin}} \sum_{\mathbf{y} \in \mathbf{Y}_{P}} E(\mathbf{y})-\sum_{\mathbf{y} \notin \mathbf{Y}_{P}} E(\mathbf{y}) E(y)=−21yTWyW =Wargmin∑y∈YPE(y)−∑y∈/YPE(y)
- Simple gradient descent
- W = W + η ( ∑ y ∈ Y P α y y y T − ∑ y ∉ Y P β ( E ( y ) ) y y T ) \mathbf{W}=\mathbf{W}+\eta\left(\sum_{\mathbf{y} \in \mathbf{Y}_{P}} \alpha_{\mathbf{y}} \mathbf{y} \mathbf{y}^{T}-\sum_{\mathbf{y} \notin \mathbf{Y}_{P}} \beta(E(\mathbf{y})) \mathbf{y} \mathbf{y}^{T}\right) W=W+η(∑y∈YPαyyyT−∑y∈/YPβ(E(y))yyT)
- α y \alpha_y αy more importance to more frequently presented memories
- β ( E ( y ) ) \beta (E(y)) β(E(y)) more importance to more attractive spurious memories
- Looks like an expectation
- W = W + η ( E y ∼ Y P y y T − E y ∼ Y y y T ) \mathbf{W}=\mathbf{W}+\eta\left(E_{\mathbf{y} \sim \mathbf{Y}_{P}} \mathbf{y} \mathbf{y}^{T}-E_{\mathbf{y} \sim Y} \mathbf{y} \mathbf{y}^{T}\right) W=W+η(Ey∼YPyyT−Ey∼YyyT)
- The behavior of the Hopfield net is analogous to annealed dynamics of a spin glass characterized by a Boltzmann distribution