lambeq use cases
lambeq
covers a wide range of experiment use cases (Fig. 4) in three broad categories:
quantum simulations on classical hardware;
actual runs on quantum hardware;
evaluation of tensor networks on classical hardware.
The above figure introduces a couple of concepts that might need further explanation for users new to quantum computing:
shotbased run/simulation: Unlike classical computers, quantum computers are inherently nondeterministic. This means that running a quantum circuit only once and using the output for some task would produce unreliable results. The solution is to run the same circuit many times (or shots), exploiting statistical aggregation. The inherent uncertainty of quantum computers is greatly increased by the limitations of current NISQ devices, which are prone to noise, errors, and environmental interference.
noisy simulation: A noisy simulation uses a noise model that tries to approximate the negative effect of noise, errors, and environmental interference that are inherent in current NISQ devices. It is the closest you can get to an actual quantum run from a simulation running on classical hardware.
Table 1 provides a concise reference for the most common scenarios, together with the recommended lambeq
models and trainers to use for each of them, while the following subsections present each case in more detail.
Use case 
Configurations 


Exact nonshot based simulation of quantum circuits on classical hardware 
NumpyModel with QuantumTrainer 

Noiseless shotbased simulation of quantum circuits on classical hardware 
TketModel with QuantumTrainer , 

Noisy shotbased simulation of quantum circuits on classical hardware 
TketModel with QuantumTrainer 

Evaluation of quantum circuits on a quantum computer 
TketModel with QuantumTrainer 

Evaluation of classical, tensorbased models 

Hybrid classical/quantum simulation of quantum circuits on classical hardware 
Exact (non shotbased) simulation of quantum circuits on classical hardware
 Description:
Perform a simple, noiseless, nonshotbased simulation of a quantum run on classical hardware.
 Configuration:
NumpyModel
withQuantumTrainer
.PennyLaneModel
withPytorchTrainer
.
 When to use:
As a first proofofconcept for a quantum model configuration
As a simple baseline for comparing with quantum runs
When fast training speeds are required
Computation with NISQ devices is slow, noisy and limited, so it is still not practical to do extensive training and comparative analyses on them. For this reason, and especially at the early stages of modelling, proofsofconcept are usually obtained by running simulations on classical hardware. The simplest possible way to simulate a quantum computation on a classical computer is by using linear algebra; since quantum gates correspond to complexvalued tensors, each circuit can be represented as a tensor network where computation takes the form of tensor contraction. The output of the tensor network gives the ideal probability distribution of the measurement outcomes on a noisefree quantum computer and is only a rough approximation of the sampled probability distribution obtained from a NISQ device. An “exact simulation” of this form usually serves as a simple baseline or the first proof of concept for testing a quantum configuration, and in lambeq
is implemented by the NumpyModel
class, and by the PennyLaneModel
with the attribute backend_config={'backend'='default.qubit', 'shots'=None}
.
See also:
Shotbased simulation of quantum circuits on classical hardware
 Description:
Noisy or noiseless shotbased simulations on classical hardware using tket or PennyLane backends.
 Configuration:
TketModel
withQuantumTrainer
.PennyLaneModel
withPytorchTrainer
.
 When to use:
As a faithful approximation of an actual quantum run
When the available actual quantum machines are still small for the kind of experiment you have in mind
When a faithful approximation of a quantum run is needed, one should use a proper shotbased simulation, optionally including a noise model that is appropriate for the specific kind of quantum hardware. In fact, a noisy shotbased simulation is as close as we could get to an actual quantum run. For example, in order to run an architectureaware simulation on an IBM machine, we could use a TketModel
initialised with a Qiskit noise model:
from pytket.extensions.qiskit import IBMQEmulatorBackend
from lambeq import TketModel
all_circuits = train_circuits + dev_circuits + test_circuits
device_name = 'ibmq_washington' # need credentials to access this device
backend = IBMQEmulatorBackend(device_name)
backend_config = {
'backend': backend,
'compilation': backend.default_compilation_pass(2),
'shots': 8192
}
model = TketModel.from_diagrams(all_circuits, backend_config=backend_config)
As another example, simulating a noisy run on a Honeywell machine with a PennyLaneModel
would require the following initialisation:
from lambeq import PennyLaneModel
all_circuits = train_circuits + dev_circuits + test_circuits
backend_config = {'backend': 'honeywell.hqs',
'device': 'H1',
'shots': 1000,
'probabilities': True,
'normalize': True}
model = PennyLaneModel.from_diagrams(all_circuits,
backend_config=backend_config)
If you have not previously done so, it will be necessary to save your Honeywell account email address to the PennyLane configuration file in order to use the ‘honeywell.hqs’ backend:
import pennylane as qml
qml.default_config["honeywell.global.user_email"] = "my_Honeywell/Quantinuum_account_email"
qml.default_config.save(qml.default_config.path)
Using a noise model in our simulations is not always necessary, especially in the early stages of modelling when it is often useful to assess the expected performance of the model in ideal conditions, ignoring the effects of noise and environmental interference. By default PennyLaneModel
uses a noiseless simulation, and a shotbased simulation can be initialised as below:
from lambeq import PennyLaneModel
backend_config = {'shots': 1000}
model = PennyLaneModel.from_diagrams(all_circuits,
backend_config=backend_config)
See also:
Evaluation of quantum circuits on a quantum computer
 Description:
Perform actual quantum runs using tket or PennyLane backends.
 Configuration:
TketModel
withQuantumTrainer
.PennyLaneModel
withPytorchTrainer
.
 When to use:
The real thing, use it whenever possible!
As soon as you are satisfied with the results of the simulations, it’s time for the ultimate test of your model on a real quantum machine. For this, you will need an account on a platform that provides quantum services, such as IBM Quantum.
Note
While providers usually offer free plans which allow some limited access to their resources, depending on your experimental needs a paid subscription might be required. Table 2 summarises some popular quantum platforms that are currently available to the public.
Platform 
Technology 

Annealing, trapped ions, superconducting qubits, photonics 

Superconducting qubits 

Superconducting qubits 

Trapped ions 

Superconducting qubits 

Trapped ions, superconducting qubits, neutral atoms 

Trapped ions 

Superconducting qubits 
See also:
Evaluation of classical tensorbased models
 Description:
Perform tensorbased experiments on classical hardware using PyTorch.
 Configuration:
PytorchModel
withPytorchTrainer
. When to use:
As a proofofconcept for validating sentence modelling at a high level
As a classical baseline to compare with similarly structured quantum models
For enhancing models with neural parts and other ML features
While lambeq
is primarily aimed at the design and execution of NLP models on quantum hardware, in practice it is more than a QNLP toolkit: it is a modelling tool capable of representing language at many different levels of abstraction, including syntax trees, string/monoidal diagrams, strict pregroup diagrams, and quantum circuits. For example, the abstract representation given by a string diagram can be directly translated into a tensor network and executed on classical hardware. This can be useful for providing comparison and benchmarking between quantum models and similar classical implementations.
Furthermore, using the PyTorch backend via PytorchModel
provides access to a wide range of robust deep learning features, allowing you to combine your tensorbased models with neural parts (e.g. embeddings or classifiers) in an effortless way.
See also:
Hybrid classical/quantum simulations on classical hardware
 Description:
Hybrid neural/classical/quantum configurations based on PennyLane and PyTorch.
 Configuration:
PennyLaneModel
withPytorchTrainer
. When to use:
To mix neural nets (or other classical models) and quantum circuits into hybrid models
To exploit the rich functionality and options provided by the PennyLane toolkit
PennyLane is currently one of the most complete quantum ML toolkits available, covering almost every possible training use case. One of its big strengths is allowing the combination of quantum and classical parts in models, in what is usually referred to as hybrid QML. PennyLane integrates smoothly with PyTorch; for example in lambeq
it is possible to use a PennyLaneModel
in conjunction with a PytorchTrainer
to perform a wide range of experiments.
See also: