Vanishing and exploding gradients

Deep Learning intermedio con PyTorch

Michal Oleszak

Machine Learning Engineer

Vanishing gradients

Gradients get smaller and smaller during backward pass
Earlier layers get small parameter updates
Model doesn't learn

Graphs showing gradient size versus layer index: for earlier layers, the gradients are smaller

Exploding gradients

Gradients get bigger and bigger
Parameter updates are too large
Training diverges

Graphs showing gradient size versus layer index: for earlier layers, the gradients are larger

Solution to unstable gradients

Proper weights initialization
Good activations
Batch normalization

Three steps

Weights initialization

layer = nn.Linear(8, 1)
print(layer.weight)

Parameter containing:
tensor([[-0.0195,  0.0992,  0.0391,  0.0212,
         -0.3386, -0.1892, -0.3170,  0.2148]])

Weights initialization

Good initialization ensures:

Variance of layer inputs = variance of layer outputs
Variance of gradients the same before and after a layer

How to achieve this depends on the activation:

For ReLU and similar, we can use He/Kaiming initialization

Weights initialization

import torch.nn.init as init

init.kaiming_uniform_(layer.weight)
print(layer.weight)

Parameter containing:
tensor([[-0.3063, -0.2410,  0.0588,  0.2664,
          0.0502, -0.0136,  0.2274,  0.0901]])

He / Kaiming initialization

init.kaiming_uniform_(self.fc1.weight)
init.kaiming_uniform_(self.fc2.weight)
init.kaiming_uniform_(
  self.fc3.weight,
  nonlinearity="sigmoid",
)

He / Kaiming initialization

import torch.nn as nn
import torch.nn.init as init

class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(9, 16)
        self.fc2 = nn.Linear(16, 8)
        self.fc3 = nn.Linear(8, 1)


        init.kaiming_uniform_(self.fc1.weight)
        init.kaiming_uniform_(self.fc2.weight)
        init.kaiming_uniform_(
          self.fc3.weight,
          nonlinearity="sigmoid",
        )





    def forward(self, x):
        x = nn.functional.relu(self.fc1(x))
        x = nn.functional.relu(self.fc2(x))
        x = nn.functional.sigmoid(self.fc3(x))
        return x

Activation functions

A plot showing the ReLU function. For values below zero, the line is horizontal at zero; for values above zero, the curve the line takes a positive slope.

Often used as the default activation
nn.functional.relu()
Zero for negative inputs - dying neurons

A plot showing the ELU function. It is similar to the ReLU function but with a curved and smooth transition from negative values to the positive region.

nn.functional.elu()
Non-zero gradients for negative values - helps against dying neurons
Average output around zero - helps against vanishing gradients

Batch normalization

After a layer:

Normalize the layer's outputs by:
- Subtracting the mean
- Dividing by the standard deviation
Scale and shift normalized outputs using learned parameters

Model learns optimal inputs distribution for each layer:

Faster loss decrease
Helps against unstable gradients

Batch normalization

class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(9, 16)
        self.bn1 = nn.BatchNorm1d(16)

        ...


    def forward(self, x):
        x = self.fc1(x)
        x = self.bn1(x)
        x = nn.functional.elu(x)

        ...

Let's practice!

Deep Learning intermedio con PyTorch