Quantum Neural Networks (QNN) implementation.
QNNs are parameterized quantum circuits organized in layers that can approximate arbitrary functions for both classification and regression tasks. This implementation follows Clojure principles of simplicity and data orientation.
Key Features:
Design Principles:
Architecture:
Quantum Neural Networks (QNN) implementation. QNNs are parameterized quantum circuits organized in layers that can approximate arbitrary functions for both classification and regression tasks. This implementation follows Clojure principles of simplicity and data orientation. Key Features: - Layer-based architecture using simple data maps - Explicit parameter management per layer - Sequential network composition using vectors - Integration with existing variational algorithm infrastructure - Hardware-compatible quantum operations - Support for multiple activation patterns Design Principles: - Layers as data: Each layer is a simple map describing its configuration - Explicit parameters: Clear parameter ownership and allocation per layer - Functional composition: Networks as vectors of layers processed sequentially - Data orientation: All network state represented as plain Clojure data structures Architecture: 1. Input encoding layer (feature map) 2. Hidden quantum layers (parameterized unitaries) 3. Quantum activation functions (non-linear transformations) 4. Output measurement layer (classical readout)
(allocate-parameters network)Allocate parameter indices for each layer in a QNN network.
Parameters:
Returns: Map with :network (updated with parameter indices) and :total-parameters
Allocate parameter indices for each layer in a QNN network. Parameters: - network: Vector of layer configuration maps Returns: Map with :network (updated with parameter indices) and :total-parameters
(analyze-qnn-network network)Analyze a QNN network configuration and provide insights.
Parameters:
Returns: Analysis map with network statistics and insights
Analyze a QNN network configuration and provide insights. Parameters: - network: QNN network configuration Returns: Analysis map with network statistics and insights
(apply-activation-layer circuit layer _feature-data parameters)Apply quantum activation function to circuit qubits.
Parameters:
Returns: Quantum circuit with applied activation function
Apply quantum activation function to circuit qubits. Parameters: - circuit: Input quantum circuit - layer: Activation layer configuration - feature-data: Not used for activation layers - parameters: Parameter vector for this layer Returns: Quantum circuit with applied activation function
(apply-dense-layer circuit layer _feature-data parameters)Apply dense quantum layer with parameterized rotations.
Parameters:
Returns: Quantum circuit with applied dense transformations
Apply dense quantum layer with parameterized rotations. Parameters: - circuit: Input quantum circuit - layer: Dense layer configuration - feature-data: Not used for dense layers - parameters: Parameter vector for this layer Returns: Quantum circuit with applied dense transformations
(apply-entangling-layer circuit layer _feature-data _parameters)Apply entangling layer with specified connectivity pattern.
Parameters:
Returns: Quantum circuit with applied entangling operations
Apply entangling layer with specified connectivity pattern. Parameters: - circuit: Input quantum circuit - layer: Entangling layer configuration - feature-data: Not used for entangling layers - parameters: Parameter vector for this layer (for parameterized entangling) Returns: Quantum circuit with applied entangling operations
(apply-input-layer circuit layer feature-data _parameters)Apply input layer (feature encoding) to a quantum circuit.
Parameters:
Returns: Quantum circuit with encoded features
Apply input layer (feature encoding) to a quantum circuit. Parameters: - circuit: Input quantum circuit - layer: Input layer configuration - feature-data: Classical feature vector to encode - parameters: Not used for input layers Returns: Quantum circuit with encoded features
(apply-layer circuit layer feature-data layer-parameters)Apply a single QNN layer to a quantum circuit.
Parameters:
Returns: Quantum circuit with layer operations applied
Apply a single QNN layer to a quantum circuit. Parameters: - circuit: Input quantum circuit - layer: Layer configuration map - feature-data: Classical feature data (used only by input layers) - layer-parameters: Parameters specific to this layer Returns: Quantum circuit with layer operations applied
(apply-measurement-layer circuit layer _feature-data _parameters)Apply measurement operations to circuit (for output layers).
Parameters:
Returns: Quantum circuit with measurement operations
Apply measurement operations to circuit (for output layers). Parameters: - circuit: Input quantum circuit - layer: Measurement layer configuration - feature-data: Not used for measurement layers - parameters: Not used for measurement layers Returns: Quantum circuit with measurement operations
(apply-qnn-network network feature-data parameters)Apply a complete QNN network to input data, creating a quantum circuit.
This function implements the forward pass of a QNN by sequentially applying each layer to the quantum circuit. It handles parameter extraction and layer-specific data passing.
Parameters:
Returns: Quantum circuit representing the forward pass of the QNN
Apply a complete QNN network to input data, creating a quantum circuit. This function implements the forward pass of a QNN by sequentially applying each layer to the quantum circuit. It handles parameter extraction and layer-specific data passing. Parameters: - network: Vector of layer configuration maps (with allocated parameters) - feature-data: Classical input feature vector - parameters: Complete parameter vector for the entire network Returns: Quantum circuit representing the forward pass of the QNN
(compute-loss prediction label loss-type)Compute loss between prediction and true label.
Parameters:
Returns: Loss value as a real number
Compute loss between prediction and true label. Parameters: - prediction: Model prediction (expectation value, probability vector, or bitstring) - label: True label (number for regression, integer for classification) - loss-type: Type of loss function Returns: Loss value as a real number
(count-layer-parameters layer)Count the number of parameters needed for a specific layer type.
Parameters:
Returns: Number of parameters required for this layer
Count the number of parameters needed for a specific layer type. Parameters: - layer: Layer configuration map Returns: Number of parameters required for this layer
(create-feedforward-qnn num-qubits
hidden-layers
&
{:keys [feature-map-type activation-type]
:or {feature-map-type :angle
activation-type :quantum-tanh}})Create a standard feedforward QNN configuration.
Parameters:
Returns: Complete QNN network configuration with allocated parameters
Create a standard feedforward QNN configuration. Parameters: - num-qubits: Number of qubits in the network - hidden-layers: Number of hidden layers - feature-map-type: Input encoding type (:angle, :amplitude, etc.) - activation-type: Activation function type Returns: Complete QNN network configuration with allocated parameters
(create-qnn-circuit-constructor network)Create a circuit constructor function for QNN integration with variational algorithms.
This function creates a circuit constructor compatible with the variational algorithm template. It returns a function that takes parameters and feature data, returning a complete QNN circuit.
Parameters:
Returns: Function (parameters, feature-data) -> quantum circuit
Create a circuit constructor function for QNN integration with variational algorithms. This function creates a circuit constructor compatible with the variational algorithm template. It returns a function that takes parameters and feature data, returning a complete QNN circuit. Parameters: - network: QNN network configuration (with allocated parameters) Returns: Function (parameters, feature-data) -> quantum circuit
(create-qnn-cost-function network
training-data
backend
loss-type
&
{:keys [options] :or {options {}}})Create a cost function for QNN training compatible with variational algorithms.
This function creates a cost function that takes parameters and returns a loss value. It integrates with the QClojure variational algorithm template for optimization.
Parameters:
Returns: Cost function (parameters) -> loss value
Create a cost function for QNN training compatible with variational algorithms.
This function creates a cost function that takes parameters and returns a loss value.
It integrates with the QClojure variational algorithm template for optimization.
Parameters:
- network: QNN network configuration
- training-data: Vector of {:features [...] :label ...} maps
- backend: Quantum backend for circuit execution
- loss-type: Loss function type (:mse, :cross-entropy, :hinge, etc.)
- options: Training options (shots, measurement strategy, etc.)
Returns:
Cost function (parameters) -> loss value(create-qnn-model config)Create a complete QNN model with training and inference capabilities.
This function provides a high-level interface for QNN creation, training, and inference, abstracting away the low-level details.
Parameters:
Configuration options:
Returns: QNN model map with train and predict functions
Create a complete QNN model with training and inference capabilities. This function provides a high-level interface for QNN creation, training, and inference, abstracting away the low-level details. Parameters: - config: QNN configuration map Configuration options: - :num-qubits - Number of qubits - :hidden-layers - Number of hidden layers - :feature-map-type - Input encoding type - :activation-type - Activation function type Returns: QNN model map with train and predict functions
(evaluate-qnn-accuracy network test-data parameters backend)Evaluate QNN accuracy on a dataset.
Parameters:
Returns: Accuracy as a number between 0 and 1
Evaluate QNN accuracy on a dataset. Parameters: - network: QNN network configuration - test-data: Test dataset - parameters: Optimized QNN parameters - backend: Quantum backend Returns: Accuracy as a number between 0 and 1
(extract-expectation-value result & {:keys [options] :or {options {}}})Extract expectation value from quantum measurement results.
Production-ready implementation that properly handles result-specs format from backend protocol execution. Works with both measurement-based expectations and explicit expectation value results.
For classification, computes <Z> expectation on output qubits. For regression, uses explicit expectation values from backend.
Parameters:
Returns: Expectation value as a real number
Extract expectation value from quantum measurement results. Production-ready implementation that properly handles result-specs format from backend protocol execution. Works with both measurement-based expectations and explicit expectation value results. For classification, computes <Z> expectation on output qubits. For regression, uses explicit expectation values from backend. Parameters: - result: Quantum execution result from backend with :results map - options: Optional configuration map - :observable-index - Index of observable to extract (default: 0) - :use-measurements - Compute from measurements vs use explicit expectations Returns: Expectation value as a real number
(extract-layer-parameters layer full-parameters)Extract parameters for a specific layer from the full parameter vector.
Parameters:
Returns: Parameter vector for this specific layer
Extract parameters for a specific layer from the full parameter vector. Parameters: - layer: Layer configuration with parameter indices - full-parameters: Complete parameter vector for the network Returns: Parameter vector for this specific layer
(extract-most-probable-state result & {:keys [options] :or {options {}}})Extract the most probable computational basis state.
Production-ready implementation that properly handles result-specs format from backend protocol execution. Finds the bitstring with the highest measurement count.
Parameters:
Returns: String representing the most probable bitstring, or [bitstring count] if :return-count is true
Extract the most probable computational basis state. Production-ready implementation that properly handles result-specs format from backend protocol execution. Finds the bitstring with the highest measurement count. Parameters: - result: Quantum execution result from backend with :results map - options: Optional configuration map - :return-count - Also return the count (default: false) Returns: String representing the most probable bitstring, or [bitstring count] if :return-count is true
(extract-probability-distribution result & {:keys [options] :or {options {}}})Extract probability distribution from quantum measurement results.
Production-ready implementation that properly handles result-specs format from backend protocol execution. Converts measurement counts to normalized probabilities.
Parameters:
Returns: Vector of probabilities for each computational basis state, sorted by bitstring value (e.g., [p("00"), p("01"), p("10"), p("11")])
Extract probability distribution from quantum measurement results.
Production-ready implementation that properly handles result-specs format
from backend protocol execution. Converts measurement counts to normalized
probabilities.
Parameters:
- result: Quantum execution result from backend with :results map
- options: Optional configuration map
- :num-states - Expected number of states (default: infer from outcomes)
- :normalize - Whether to normalize probabilities (default: true)
Returns:
Vector of probabilities for each computational basis state, sorted by
bitstring value (e.g., [p("00"), p("01"), p("10"), p("11")])(optimize-parameters cost-fn initial-params options)Simple parameter optimization for demonstration.
In a full implementation, this would delegate to QClojure's optimization algorithms like Nelder-Mead, COBYLA, or gradient-based methods.
Parameters:
Returns: Optimization result map
Simple parameter optimization for demonstration. In a full implementation, this would delegate to QClojure's optimization algorithms like Nelder-Mead, COBYLA, or gradient-based methods. Parameters: - cost-fn: Cost function to minimize - initial-params: Initial parameter values - options: Optimization options Returns: Optimization result map
(qnn-circuit-constructor-adapter options)Create circuit constructor function for QNN from options.
This function creates a circuit constructor that works with the enhanced variational template. The resulting function takes parameters and a feature vector, returning a complete QNN circuit.
Parameters:
Returns: Function that takes (parameters, feature-vector) and returns a quantum circuit
Create circuit constructor function for QNN from options. This function creates a circuit constructor that works with the enhanced variational template. The resulting function takes parameters and a feature vector, returning a complete QNN circuit. Parameters: - options: QNN training options containing :network Returns: Function that takes (parameters, feature-vector) and returns a quantum circuit
(qnn-dataset-constructor options)Extract training data from QNN options for the variational algorithm template.
This function adapts the QNN data format to work with the variational algorithm template, which expects a dataset source.
Parameters:
Returns: Training data vector
Extract training data from QNN options for the variational algorithm template. This function adapts the QNN data format to work with the variational algorithm template, which expects a dataset source. Parameters: - options: QNN training options containing :training-data Returns: Training data vector
(qnn-forward-pass network
feature-data
parameters
backend
&
{:keys [options] :or {options {}}})Execute the forward pass of a QNN on a quantum backend.
This function executes the complete QNN circuit on a quantum backend using the proper backend protocol with result specifications. It returns measurement results suitable for both classification and regression tasks.
Production-ready implementation:
Parameters:
Options:
Returns: Execution result map from backend containing:
Execute the forward pass of a QNN on a quantum backend. This function executes the complete QNN circuit on a quantum backend using the proper backend protocol with result specifications. It returns measurement results suitable for both classification and regression tasks. Production-ready implementation: - Uses backend protocol's execute-circuit with result-specs - Requests appropriate measurements for the task type - Compatible with real quantum hardware and cloud services Parameters: - network: QNN network configuration - feature-data: Input feature vector - parameters: Network parameter vector - backend: Quantum backend implementing QuantumBackend protocol - options: Execution options map Options: - :shots - Number of measurement shots (default: 1024) - :task-type - :classification or :regression (default: :classification) - :measurement-qubits - Qubits to measure (default: all output qubits) - :observables - Observables for expectation value measurements (default: Pauli-Z) Returns: Execution result map from backend containing: - :results - Result map with measurements, expectations, or probabilities - :final-state - Final quantum state (if state-vector available) - :execution-metadata - Backend execution metadata
(qnn-loss-function loss-type task-type)Create a loss function for QNN that works with the enhanced variational template.
This function creates loss functions suitable for both classification and regression tasks in quantum neural networks. It computes loss based on predictions and labels/targets.
Parameters:
Returns: Function that takes (predictions, labels) and returns loss value
Create a loss function for QNN that works with the enhanced variational template. This function creates loss functions suitable for both classification and regression tasks in quantum neural networks. It computes loss based on predictions and labels/targets. Parameters: - loss-type: Type of loss function (:mse, :cross-entropy, :hinge, :accuracy) - task-type: Type of ML task (:classification or :regression) Returns: Function that takes (predictions, labels) and returns loss value
(qnn-parameter-count options)Calculate required parameter count for QNN network.
This function determines how many parameters are needed for the complete QNN network by summing parameters across all layers.
Parameters:
Returns: Total number of parameters required
Calculate required parameter count for QNN network. This function determines how many parameters are needed for the complete QNN network by summing parameters across all layers. Parameters: - options: QNN training options containing :network Returns: Total number of parameters required
(qnn-prediction-extractor measurement-result
task-type
measurement-strategy
shots)Extract prediction from QNN measurement result for enhanced template.
This function is designed to work with the enhanced-variational-algorithm template's prediction extraction infrastructure. It handles both classification and regression.
Parameters:
Returns: Predicted value (integer for classification, float for regression)
Extract prediction from QNN measurement result for enhanced template. This function is designed to work with the enhanced-variational-algorithm template's prediction extraction infrastructure. It handles both classification and regression. Parameters: - measurement-result: Measurement result map - task-type: Type of ML task (:classification or :regression) - measurement-strategy: How to extract predictions (:expectation, :probability, :bitstring) - shots: Number of measurement shots Returns: Predicted value (integer for classification, float for regression)
(qnn-result-processor optimization-result options)Process QNN training results with ML-specific metrics.
This processor extracts ML performance metrics from the optimization results and adds QNN-specific analysis.
Parameters:
Returns: Enhanced result map with QNN metrics
Process QNN training results with ML-specific metrics. This processor extracts ML performance metrics from the optimization results and adds QNN-specific analysis. Parameters: - optimization-result: Result from variational optimization - options: QNN training options Returns: Enhanced result map with QNN metrics
(train-qnn backend options)Train a QNN using the enhanced variational algorithm template.
This function configures and executes quantum neural network training with support for both classification and regression tasks. The QNN is trained using the enhanced variational algorithm template with ML-specific objective handling.
The training process:
Parameters:
Required options:
Optional options (with defaults):
Returns: Map containing trained model and analysis results:
Example:
(train-qnn
backend
{:network qnn-network
:training-data [{:features [0.1 0.2] :label 0}
{:features [0.8 0.9] :label 1}]
:task-type :classification
:loss-type :accuracy
:optimization-method :cmaes
:max-iterations 100
:shots 1024})Train a QNN using the enhanced variational algorithm template.
This function configures and executes quantum neural network training with support
for both classification and regression tasks. The QNN is trained using the enhanced
variational algorithm template with ML-specific objective handling.
The training process:
1. Encodes training features into quantum states via input layers
2. Applies parameterized quantum layers (dense, entangling, activation)
3. Measures results and extracts predictions
4. Optimizes network parameters to minimize task-specific loss
Parameters:
- backend: Quantum backend for circuit execution
- options: Configuration map with required and optional keys
Required options:
- :network - QNN network configuration (vector of layer maps)
- :training-data - Vector of {:features [...] :label ...} maps
Optional options (with defaults):
- :task-type - ML task type (:classification or :regression) [default: :classification]
- :loss-type - Loss function (:accuracy, :cross-entropy, :hinge, :mse, :mae)
[default: :accuracy for classification, :mse for regression]
- :measurement-strategy - How to extract predictions (:expectation, :probability, :bitstring)
[default: :bitstring for classification, :expectation for regression]
- :optimization-method - Optimizer to use (:cmaes, :nelder-mead, :powell, :adam)
[default: :cmaes]
- :max-iterations - Maximum training iterations [default: 100]
- :tolerance - Convergence tolerance [default: 1e-6]
- :shots - Number of measurement shots [default: 1024]
- :initial-parameters - Initial parameter values (random if not provided)
Returns:
Map containing trained model and analysis results:
- :success - Training completion status
- :optimal-parameters - Trained network parameters
- :final-loss - Final loss value
- :final-accuracy - Final accuracy (for classification tasks)
- :iterations - Number of iterations performed
- :convergence-history - Loss values over iterations
- :total-runtime-ms - Total training time
- :network-info - Network architecture details
- :training-info - Training configuration summary
Example:
```clojure
(train-qnn
backend
{:network qnn-network
:training-data [{:features [0.1 0.2] :label 0}
{:features [0.8 0.9] :label 1}]
:task-type :classification
:loss-type :accuracy
:optimization-method :cmaes
:max-iterations 100
:shots 1024})(visualize-qnn-network network)Create a simple ASCII visualization of the QNN network structure.
Parameters:
Returns: String containing ASCII art representation
Create a simple ASCII visualization of the QNN network structure. Parameters: - network: QNN network configuration Returns: String containing ASCII art representation
cljdoc builds & hosts documentation for Clojure/Script libraries
| Ctrl+k | Jump to recent docs |
| ← | Move to previous article |
| → | Move to next article |
| Ctrl+/ | Jump to the search field |