Tensor Mathematics & Backpropagation

calendarDec 12, 2024 · 6 min read · ai-knowhow tensors backpropagation mathematics deep-learning  ·
Share on: linkedincopy

Tensor mathematics fundamentals and backpropagation theory with detailed mathematical derivations.

Tensor Basics

What is a Tensor?

A tensor is a generalization of scalars, vectors, and matrices to higher dimensions:

  • Scalar (0D tensor): $x \in \mathbb{R}$
  • Vector (1D tensor): $\mathbf{x} \in \mathbb{R}^n$
  • Matrix (2D tensor): $\mathbf{X} \in \mathbb{R}^{m \times n}$
  • 3D Tensor: $\mathcal{X} \in \mathbb{R}^{m \times n \times p}$
  • nD Tensor: $\mathcal{X} \in \mathbb{R}^{d_1 \times d_2 \times \cdots \times d_n}$

Tensor Operations

Element-wise Operations

Addition: $$ (\mathbf{A} + \mathbf{B}){ij} = A{ij} + B_{ij} $$

Multiplication (Hadamard product): $$ (\mathbf{A} \odot \mathbf{B}){ij} = A{ij} \cdot B_{ij} $$

Matrix Multiplication

$$ (\mathbf{AB}){ij} = \sum{k=1}^{n} A_{ik}B_{kj} $$

Tensor Contraction

Generalization of matrix multiplication to higher dimensions:

$$ \mathcal{C}{i_1...i_m,k_1...k_p} = \sum{j_1,...,j_n} \mathcal{A}{i_1...i_m,j_1...j_n} \mathcal{B}{j_1...j_n,k_1...k_p} $$

Gradients and Derivatives

Scalar Derivative

$$ \frac{\partial f}{\partial x} = \lim_{h \to 0} \frac{f(x+h) - f(x)}{h} $$

Vector Gradient

For $f: \mathbb{R}^n \to \mathbb{R}$:

$$ \nabla f(\mathbf{x}) = \begin{bmatrix} \frac{\partial f}{\partial x_1} \ \frac{\partial f}{\partial x_2} \ \vdots \ \frac{\partial f}{\partial x_n} \end{bmatrix} $$

Jacobian Matrix

For $\mathbf{f}: \mathbb{R}^n \to \mathbb{R}^m$:

$$ \mathbf{J} = \begin{bmatrix} \frac{\partial f_1}{\partial x_1} & \cdots & \frac{\partial f_1}{\partial x_n} \ \vdots & \ddots & \vdots \ \frac{\partial f_m}{\partial x_1} & \cdots & \frac{\partial f_m}{\partial x_n} \end{bmatrix} $$

Hessian Matrix

Second-order partial derivatives:

$$ \mathbf{H}_{ij} = \frac{\partial^2 f}{\partial x_i \partial x_j} $$

Chain Rule for Backpropagation

Univariate Chain Rule

$$ \frac{dy}{dx} = \frac{dy}{du} \cdot \frac{du}{dx} $$

Multivariate Chain Rule

For $z = f(y_1, ..., y_m)$ where $y_i = g_i(x_1, ..., x_n)$:

$$ \frac{\partial z}{\partial x_j} = \sum_{i=1}^{m} \frac{\partial z}{\partial y_i} \frac{\partial y_i}{\partial x_j} $$

Vector Chain Rule

$$ \frac{\partial \mathbf{z}}{\partial \mathbf{x}} = \frac{\partial \mathbf{z}}{\partial \mathbf{y}} \frac{\partial \mathbf{y}}{\partial \mathbf{x}} $$

Backpropagation Algorithm

Forward Pass

Given input $\mathbf{x}$, compute activations layer by layer:

Layer $l$: $$ \begin{aligned} \mathbf{z}^{[l]} &= \mathbf{W}^{[l]} \mathbf{a}^{[l-1]} + \mathbf{b}^{[l]} \ \mathbf{a}^{[l]} &= \sigma(\mathbf{z}^{[l]}) \end{aligned} $$

Where:

  • $\mathbf{W}^{[l]}$: Weight matrix for layer $l$
  • $\mathbf{b}^{[l]}$: Bias vector for layer $l$
  • $\sigma$: Activation function
  • $\mathbf{a}^{[0]} = \mathbf{x}$: Input

Backward Pass

Compute gradients layer by layer (from output to input):

Output Layer $L$: $$ \delta^{[L]} = \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[L]}} = \frac{\partial \mathcal{L}}{\partial \mathbf{a}^{[L]}} \odot \sigma'(\mathbf{z}^{[L]}) $$

Hidden Layer $l$: $$ \delta^{[l]} = (\mathbf{W}^{[l+1]})^T \delta^{[l+1]} \odot \sigma'(\mathbf{z}^{[l]}) $$

Gradients: $$ \begin{aligned} \frac{\partial \mathcal{L}}{\partial \mathbf{W}^{[l]}} &= \delta^{[l]} (\mathbf{a}^{[l-1]})^T \ \frac{\partial \mathcal{L}}{\partial \mathbf{b}^{[l]}} &= \delta^{[l]} \end{aligned} $$

Activation Functions and Derivatives

Sigmoid

$$ \sigma(x) = \frac{1}{1 + e^{-x}} $$

Derivative: $$ \sigma'(x) = \sigma(x)(1 - \sigma(x)) $$

Tanh

$$ \tanh(x) = \frac{e^x - e^{-x}}{e^x + e^{-x}} $$

Derivative: $$ \tanh'(x) = 1 - \tanh^2(x) $$

ReLU

$$ \text{ReLU}(x) = \max(0, x) $$

Derivative: $$ \text{ReLU}'(x) = \begin{cases} 1 & \text{if } x > 0 \ 0 & \text{if } x \leq 0 \end{cases} $$

Leaky ReLU

$$ \text{LeakyReLU}(x) = \max(\alpha x, x), \quad \alpha \in (0, 1) $$

Derivative: $$ \text{LeakyReLU}'(x) = \begin{cases} 1 & \text{if } x > 0 \ \alpha & \text{if } x \leq 0 \end{cases} $$

Softmax

$$ \text{softmax}(\mathbf{z})i = \frac{e^{z_i}}{\sum{j=1}^{K} e^{z_j}} $$

Jacobian: $$ \frac{\partial \text{softmax}(\mathbf{z})_i}{\partial z_j} = \text{softmax}(\mathbf{z})i (\delta{ij} - \text{softmax}(\mathbf{z})_j) $$

Cost Functions

Mean Squared Error (MSE)

$$ \mathcal{L}(\mathbf{y}, \hat{\mathbf{y}}) = \frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2 $$

Gradient: $$ \frac{\partial \mathcal{L}}{\partial \hat{y}_i} = -\frac{2}{n}(y_i - \hat{y}_i) $$

Binary Cross-Entropy

$$ \mathcal{L}(y, \hat{y}) = -[y \log(\hat{y}) + (1-y) \log(1-\hat{y})] $$

Gradient: $$ \frac{\partial \mathcal{L}}{\partial \hat{y}} = -\frac{y}{\hat{y}} + \frac{1-y}{1-\hat{y}} $$

Categorical Cross-Entropy

$$ \mathcal{L}(\mathbf{y}, \hat{\mathbf{y}}) = -\sum_{i=1}^{K} y_i \log(\hat{y}_i) $$

Gradient: $$ \frac{\partial \mathcal{L}}{\partial \hat{y}_i} = -\frac{y_i}{\hat{y}_i} $$

Softmax + Cross-Entropy (Combined)

$$ \frac{\partial \mathcal{L}}{\partial z_i} = \hat{y}_i - y_i $$

This simplification is why softmax and cross-entropy are often combined!

Optimization Algorithms

Gradient Descent

$$ \theta_{t+1} = \theta_t - \eta \nabla_\theta \mathcal{L}(\theta_t) $$

Where $\eta$ is the learning rate.

Stochastic Gradient Descent (SGD)

Update using single sample or mini-batch:

$$ \theta_{t+1} = \theta_t - \eta \nabla_\theta \mathcal{L}_i(\theta_t) $$

Momentum

$$ \begin{aligned} \mathbf{v}t &= \beta \mathbf{v}{t-1} + (1-\beta) \nabla_\theta \mathcal{L}(\theta_t) \ \theta_{t+1} &= \theta_t - \eta \mathbf{v}_t \end{aligned} $$

Typical: $\beta = 0.9$

RMSprop

$$ \begin{aligned} \mathbf{s}t &= \beta \mathbf{s}{t-1} + (1-\beta) (\nabla_\theta \mathcal{L}(\theta_t))^2 \ \theta_{t+1} &= \theta_t - \frac{\eta}{\sqrt{\mathbf{s}t + \epsilon}} \nabla\theta \mathcal{L}(\theta_t) \end{aligned} $$

Adam (Adaptive Moment Estimation)

$$ \begin{aligned} \mathbf{m}t &= \beta_1 \mathbf{m}{t-1} + (1-\beta_1) \nabla_\theta \mathcal{L}(\theta_t) \ \mathbf{v}t &= \beta_2 \mathbf{v}{t-1} + (1-\beta_2) (\nabla_\theta \mathcal{L}(\theta_t))^2 \ \hat{\mathbf{m}}_t &= \frac{\mathbf{m}_t}{1-\beta_1^t} \ \hat{\mathbf{v}}_t &= \frac{\mathbf{v}t}{1-\beta_2^t} \ \theta{t+1} &= \theta_t - \frac{\eta}{\sqrt{\hat{\mathbf{v}}_t} + \epsilon} \hat{\mathbf{m}}_t \end{aligned} $$

Typical: $\beta_1 = 0.9$, $\beta_2 = 0.999$, $\epsilon = 10^{-8}$

Batch Normalization

Forward Pass

$$ \begin{aligned} \mu_B &= \frac{1}{m} \sum_{i=1}^{m} x_i \ \sigma_B^2 &= \frac{1}{m} \sum_{i=1}^{m} (x_i - \mu_B)^2 \ \hat{x}_i &= \frac{x_i - \mu_B}{\sqrt{\sigma_B^2 + \epsilon}} \ y_i &= \gamma \hat{x}_i + \beta \end{aligned} $$

Backward Pass

$$ \begin{aligned} \frac{\partial \mathcal{L}}{\partial \gamma} &= \sum_{i=1}^{m} \frac{\partial \mathcal{L}}{\partial y_i} \hat{x}i \ \frac{\partial \mathcal{L}}{\partial \beta} &= \sum{i=1}^{m} \frac{\partial \mathcal{L}}{\partial y_i} \ \frac{\partial \mathcal{L}}{\partial \hat{x}i} &= \frac{\partial \mathcal{L}}{\partial y_i} \gamma \ \frac{\partial \mathcal{L}}{\partial \sigma_B^2} &= \sum{i=1}^{m} \frac{\partial \mathcal{L}}{\partial \hat{x}i} (x_i - \mu_B) \cdot \frac{-1}{2}(\sigma_B^2 + \epsilon)^{-3/2} \ \frac{\partial \mathcal{L}}{\partial \mu_B} &= \sum{i=1}^{m} \frac{\partial \mathcal{L}}{\partial \hat{x}i} \frac{-1}{\sqrt{\sigma_B^2 + \epsilon}} + \frac{\partial \mathcal{L}}{\partial \sigma_B^2} \frac{-2}{m} \sum{i=1}^{m} (x_i - \mu_B) \ \frac{\partial \mathcal{L}}{\partial x_i} &= \frac{\partial \mathcal{L}}{\partial \hat{x}_i} \frac{1}{\sqrt{\sigma_B^2 + \epsilon}} + \frac{\partial \mathcal{L}}{\partial \sigma_B^2} \frac{2(x_i - \mu_B)}{m} + \frac{\partial \mathcal{L}}{\partial \mu_B} \frac{1}{m} \end{aligned} $$

Dropout

Forward Pass (Training)

$$ \mathbf{r} \sim \text{Bernoulli}(p), \quad \tilde{\mathbf{h}} = \mathbf{r} \odot \mathbf{h} $$

Forward Pass (Inference)

$$ \tilde{\mathbf{h}} = p \mathbf{h} $$

Backward Pass

$$ \frac{\partial \mathcal{L}}{\partial \mathbf{h}} = \mathbf{r} \odot \frac{\partial \mathcal{L}}{\partial \tilde{\mathbf{h}}} $$

Convolutional Layer Backpropagation

Forward Pass

$$ y_{i,j} = \sum_{m=0}^{k-1} \sum_{n=0}^{k-1} w_{m,n} \cdot x_{i+m,j+n} + b $$

Backward Pass

Gradient w.r.t. weights: $$ \frac{\partial \mathcal{L}}{\partial w_{m,n}} = \sum_{i,j} \frac{\partial \mathcal{L}}{\partial y_{i,j}} \cdot x_{i+m,j+n} $$

Gradient w.r.t. input: $$ \frac{\partial \mathcal{L}}{\partial x_{i,j}} = \sum_{m,n} \frac{\partial \mathcal{L}}{\partial y_{i-m,j-n}} \cdot w_{m,n} $$

Example: Full Backprop Through 2-Layer Network

Network Architecture

$$ \begin{aligned} \mathbf{z}^{[1]} &= \mathbf{W}^{[1]} \mathbf{x} + \mathbf{b}^{[1]} \ \mathbf{a}^{[1]} &= \text{ReLU}(\mathbf{z}^{[1]}) \ \mathbf{z}^{[2]} &= \mathbf{W}^{[2]} \mathbf{a}^{[1]} + \mathbf{b}^{[2]} \ \hat{\mathbf{y}} &= \text{softmax}(\mathbf{z}^{[2]}) \ \mathcal{L} &= -\sum_i y_i \log(\hat{y}_i) \end{aligned} $$

Backpropagation Steps

Step 1: Output gradient $$ \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[2]}} = \hat{\mathbf{y}} - \mathbf{y} $$

Step 2: Layer 2 gradients $$ \begin{aligned} \frac{\partial \mathcal{L}}{\partial \mathbf{W}^{[2]}} &= \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[2]}} (\mathbf{a}^{[1]})^T \ \frac{\partial \mathcal{L}}{\partial \mathbf{b}^{[2]}} &= \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[2]}} \end{aligned} $$

Step 3: Backprop to layer 1 $$ \frac{\partial \mathcal{L}}{\partial \mathbf{a}^{[1]}} = (\mathbf{W}^{[2]})^T \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[2]}} $$

Step 4: Through ReLU $$ \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[1]}} = \frac{\partial \mathcal{L}}{\partial \mathbf{a}^{[1]}} \odot \mathbb{1}[\mathbf{z}^{[1]} > 0] $$

Step 5: Layer 1 gradients $$ \begin{aligned} \frac{\partial \mathcal{L}}{\partial \mathbf{W}^{[1]}} &= \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[1]}} \mathbf{x}^T \ \frac{\partial \mathcal{L}}{\partial \mathbf{b}^{[1]}} &= \frac{\partial \mathcal{L}}{\partial \mathbf{z}^{[1]}} \end{aligned} $$

Related Snippets