Discover the different flavours of quantum machine learning in this curated guide.
Quantum neural networks
It turns out that parametrized or variational quantum circuits can be trained similar to neural networks, which earned them the name 'quantum neural networks' (QNN) and has sparked extensive research into their behaviour as machine learning models—however the early optimism around usage of these models on near-term hardware has not stood up to scrutiny. Here you can explore the basic ideas of these models.
Integrating QNNs with machine learning software
Since quantum neural networks typically depend on classical software to perform an optimization task (or other analyses), it becomes important to integrate quantum neural networks with classical machine learning software. This can include general machine learning toolkits such as Scikit-learn, but also frameworks designed around gradient descent and backpropagation such as JAX and PyTorch.
Scaling and benchmarking challenges with QNNs
After an initial period of optimism, it became increasingly clear that naive QNN model designs face issues of scalability and do not stand out in performance compared to classical models and on near-term quantum hardware. It is a critically unresolved question of where variational circuits will fit in the fault-tolerant regime, and how to integrate them into more sophisticated quantum model designs.
Traditional quantum algorithms for machine learning
Since the beginning of QML, researchers have tried to use famous quantum subroutines—like the HHL algorithm, amplitude amplification, and more recent additions such as quantum singular value decomposition or QROM—to speed up machine learning. Early speedup claims advertised sub-linear runtime in the size of the data by making strong (and sometimes dubious) assumptions on how the data is stored and loaded. Since most classical machine learning algorithms already have linear runtimes in the size of the data, it is becoming clear that quantum computers cannot just speed up existing models, but have to solve issues that current machine learning struggles with—possibly in the small-data regime.
Quantum learning advantages
Some known quantum advantages, like Shor's algorithm or reconstructing states from quantum measurements, can be used to construct formal learning problems for which quantum speedups can be proven. This has become a vibrant field of academic research, but it is still unclear whether the highly abstract results can help guide the search for real-world QML algorithms.
Quantum machine learning and symmetries
The secret sauce of many quantum algorithms, like Shor's, is to process information in 'Fourier space' using the Quantum Fourier Transform. Fourier transforms are deeply linked to groups and symmetries—think of the symmetry of a periodic signal—and are increasingly important concepts in deep learning, for example to build symmetry-aware models or to understand the dynamics of learning. An exciting new area in QML asks if methods like the QFT can unlock fundamentally different approaches to machine learning.
Quantum neural networks
It turns out that parametrized or variational quantum circuits can be trained similar to neural networks, which earned them the name 'quantum neural networks' (QNN) and has sparked extensive research into their behaviour as machine learning models—however the early optimism around usage of these models on near-term hardware has not stood up to scrutiny. Here you can explore the basic ideas of these models.
Integrating QNNs with machine learning software
Since quantum neural networks typically depend on classical software to perform an optimization task (or other analyses), it becomes important to integrate quantum neural networks with classical machine learning software. This can include general machine learning toolkits such as Scikit-learn, but also frameworks designed around gradient descent and backpropagation such as JAX and PyTorch.
Scaling and benchmarking challenges with QNNs
After an initial period of optimism, it became increasingly clear that naive QNN model designs face issues of scalability and do not stand out in performance compared to classical models and on near-term quantum hardware. It is a critically unresolved question of where variational circuits will fit in the fault-tolerant regime, and how to integrate them into more sophisticated quantum model designs.
Traditional quantum algorithms for machine learning
Since the beginning of QML, researchers have tried to use famous quantum subroutines—like the HHL algorithm, amplitude amplification, and more recent additions such as quantum singular value decomposition or QROM—to speed up machine learning. Early speedup claims advertised sub-linear runtime in the size of the data by making strong (and sometimes dubious) assumptions on how the data is stored and loaded. Since most classical machine learning algorithms already have linear runtimes in the size of the data, it is becoming clear that quantum computers cannot just speed up existing models, but have to solve issues that current machine learning struggles with—possibly in the small-data regime.
Quantum learning advantages
Some known quantum advantages, like Shor's algorithm or reconstructing states from quantum measurements, can be used to construct formal learning problems for which quantum speedups can be proven. This has become a vibrant field of academic research, but it is still unclear whether the highly abstract results can help guide the search for real-world QML algorithms.
Quantum machine learning and symmetries
The secret sauce of many quantum algorithms, like Shor's, is to process information in 'Fourier space' using the Quantum Fourier Transform. Fourier transforms are deeply linked to groups and symmetries—think of the symmetry of a periodic signal—and are increasingly important concepts in deep learning, for example to build symmetry-aware models or to understand the dynamics of learning. An exciting new area in QML asks if methods like the QFT can unlock fundamentally different approaches to machine learning.
Quantum neural networks
It turns out that parametrized or variational quantum circuits can be trained similar to neural networks, which earned them the name 'quantum neural networks' (QNN) and has sparked extensive research into their behaviour as machine learning models—however the early optimism around usage of these models on near-term hardware has not stood up to scrutiny. Here you can explore the basic ideas of these models.
Integrating QNNs with machine learning software
Since quantum neural networks typically depend on classical software to perform an optimization task (or other analyses), it becomes important to integrate quantum neural networks with classical machine learning software. This can include general machine learning toolkits such as Scikit-learn, but also frameworks designed around gradient descent and backpropagation such as JAX and PyTorch.
Scaling and benchmarking challenges with QNNs
After an initial period of optimism, it became increasingly clear that naive QNN model designs face issues of scalability and do not stand out in performance compared to classical models and on near-term quantum hardware. It is a critically unresolved question of where variational circuits will fit in the fault-tolerant regime, and how to integrate them into more sophisticated quantum model designs.
Traditional quantum algorithms for machine learning
Since the beginning of QML, researchers have tried to use famous quantum subroutines—like the HHL algorithm, amplitude amplification, and more recent additions such as quantum singular value decomposition or QROM—to speed up machine learning. Early speedup claims advertised sub-linear runtime in the size of the data by making strong (and sometimes dubious) assumptions on how the data is stored and loaded. Since most classical machine learning algorithms already have linear runtimes in the size of the data, it is becoming clear that quantum computers cannot just speed up existing models, but have to solve issues that current machine learning struggles with—possibly in the small-data regime.
Quantum learning advantages
Some known quantum advantages, like Shor's algorithm or reconstructing states from quantum measurements, can be used to construct formal learning problems for which quantum speedups can be proven. This has become a vibrant field of academic research, but it is still unclear whether the highly abstract results can help guide the search for real-world QML algorithms.
Quantum machine learning and symmetries
The secret sauce of many quantum algorithms, like Shor's, is to process information in 'Fourier space' using the Quantum Fourier Transform. Fourier transforms are deeply linked to groups and symmetries—think of the symmetry of a periodic signal—and are increasingly important concepts in deep learning, for example to build symmetry-aware models or to understand the dynamics of learning. An exciting new area in QML asks if methods like the QFT can unlock fundamentally different approaches to machine learning.