Data Re-uploading Techniques for Expressivity

In the previous section, we explored the mathematical foundations of quantum feature maps, where classical data $x$ is mapped to a quantum state $|\phi(x)\rangle = U(x)|\psi_0\rangle$ using an encoding circuit $U(x)$. Typically, this encoding happens once, often at the beginning of the quantum circuit, followed by a parameterized processing unit $W(\theta)$. While straightforward, this "single upload" strategy can sometimes limit the complexity of the functions your quantum model can learn. The expressivity of the model, meaning the range and richness of functions it can represent, is constrained by this initial, fixed mapping combined with the variational circuit.

To create potentially more powerful quantum machine learning models capable of capturing intricate patterns in data, we can employ Data Re-uploading. Instead of encoding the data just once, we interleave data encoding layers with layers of parameterized quantum gates.

The Structure of Data Re-uploading Circuits

Imagine breaking down the variational part of your quantum circuit $W(\theta)$ into smaller blocks or layers, say $W_1(\theta_1), W_2(\theta_2), \dots, W_L(\theta_L)$. In a data re-uploading scheme, the classical input data $x$ is encoded repeatedly before each (or some) of these variational blocks.

The general structure looks like this:

$$|\psi(x, \theta)\rangle = W_L(\theta_L) U(x) W_{L-1}(\theta_{L-1}) U(x) \dots W_1(\theta_1) U(x) |\psi_0\rangle$$

Here:

• $|\psi_0\rangle$ is the initial state (e.g., $|0\dots 0\rangle$).
• $U(x)$ is the data encoding circuit, mapping the classical data $x$ into operations on the qubits. This can be any standard encoding like angle encoding, but crucially, it's applied multiple times.
• $W_i(\theta_i)$ are layers of parameterized quantum gates (e.g., rotations, CNOTs) with trainable parameters $\theta_i$.
• $L$ is the number of layers or re-uploading steps.

The main idea is that the parameters $\theta_i$ in later layers $W_i$ now act on a quantum state that has been influenced by both previous parameters $\theta_{j<i}$ and multiple applications of the data encoding $U(x)$.

A schematic representation of a data re-uploading circuit. Classical data $x$ is encoded via $U(x)$ multiple times, interspersed with parameterized processing layers $W(\theta_i)$.

Enhancing Expressivity

Why does this help? Think about how the final quantum state is constructed. Each application of $U(x)$ potentially rotates or modifies the state vector in Hilbert space in a way that depends on the input $x$. The parameterized layers $W(\theta_i)$ then process this state. When $U(x)$ is applied again, it acts on this already processed state.

This interleaving allows for the creation of much more complex functions of the input data $x$ within the quantum state amplitudes. If $U(x)$ and $W(\theta)$ involve non-commuting gates (which they usually do), the order matters, and the repeated interaction generates intricate dependencies between $\theta$ and $x$.

Consider a simple analogy: in classical neural networks, applying non-linear activation functions repeatedly allows the network to learn complex, non-linear decision boundaries. Data re-uploading can be seen as providing a somewhat analogous mechanism in the quantum domain. Each $U(x)$ layer re-introduces the input data's influence, allowing the $W(\theta_i)$ layers to progressively build up more complex correlations and features related to $x$. Research suggests that such classifiers can, in principle, approximate any continuous function of the input data, much like universal approximation theorems for classical networks.

Benefits

The primary benefit of data re-uploading is the potential for increased expressivity. This means the model might be able to:

• Represent more complex functions or decision boundaries.
• Achieve good performance with potentially fewer qubits compared to single-upload models trying to learn the same function.
• Create effective feature maps that might be hard to replicate classically.

However, there are important design choices and potential challenges:

• Circuit Design: How many re-uploading layers $L$ are optimal? What specific gates should $U(x)$ and $W(\theta_i)$ contain? These choices depend heavily on the specific problem and available quantum hardware. There's no one-size-fits-all answer.
• Trainability: While expressivity increases, so can the complexity of training. Deep re-uploading circuits might be more susceptible to optimization difficulties like barren plateaus (which we will discuss in detail in Chapter 4). The gradients might vanish, making training very slow or impossible.
• Overfitting: Highly expressive models can sometimes overfit the training data, failing to generalize to new, unseen data. Regularization techniques might become necessary.
• Hardware Resources: Each re-uploading step adds depth to the circuit, increasing the computation time and susceptibility to noise on current quantum devices. Designing hardware-efficient $U(x)$ and $W(\theta_i)$ layers is significant (covered further in Chapter 7).

Role in Quantum Models

Data re-uploading is not just an encoding trick; it's a fundamental component in constructing certain types of QML models, particularly Quantum Neural Networks (QNNs) and Variational Quantum Classifiers (VQCs). The structure naturally lends itself to a layer-wise interpretation, analogous to classical deep learning architectures. We will see these structures reappear when we discuss VQAs (Chapter 4) and QNNs (Chapter 5).

In summary, data re-uploading provides a powerful technique to enhance the expressivity of quantum machine learning models by repeatedly encoding classical data within a parameterized quantum circuit. This allows the model to learn more complex relationships between the input data and the desired output, potentially leading to better performance on challenging tasks, provided that considerations regarding circuit design, trainability, and hardware limitations are carefully managed. The next sections will explore in more detail the properties of these richer feature maps created through techniques like data re-uploading.