Adj_grad hebbian model

# class MLP_adjacent(nn.Module):

#     def __init__(self):

#         super().__init__()

#         self.W_1 = nn.Parameter(torch.zeros(10, 3), requires_grad=False).to(device)
#         self.f_1 = universal_approximator(10, requires_grad=True).to(device)

#         self.W_2 = nn.Parameter(torch.zeros(10, 10), requires_grad=False).to(device)
#         self.f_2 = universal_approximator(10, requires_grad=True).to(device)

#         self.W_3 = nn.Parameter(torch.zeros(2, 10), requires_grad=False).to(device)
#         self.f_3 = universal_approximator(2, requires_grad=True).to(device)

#         self.λ = 1e-3

#         init_weights(self)

#     def update(self):

#         self.W_1 = self.W_1 + self.λ*dW1
#         self.W_1 = self.W_1 + self.λ*dW1
#         self.W_1 = self.W_1 + self.λ*dW1


#     def forward(self, X):

#         Y1 = self.f_1(self.W_1@X)
#         dW1 = criterion_rule(X, Y1, self.W_1)
#         self.W_1 = self.W_1 + self.λ*dW1
#         # updated pass
#         Y1 = self.f_1(self.W_1@X)

#         Y2 = self.f_2(self.W_2@Y1)
#         dW2 = criterion_rule(Y1, Y2, self.W_2)
#         self.W_2 = self.W_2 + self.λ*dW2
#         # updated pass
#         Y2 = self.f_2(self.W_2@Y1)

#         Y3 = self.f_3(self.W_3@Y2)
#         dW3 = criterion_rule(Y2, Y3, self.W_3)
#         self.W_3 = self.W_3 + self.λ*dW3
#         # updated pass
#         Y3 = self.f_3(self.W_3@Y2)

#         return Y3