Deep Learning Toolbox

Design, train, analyze, and simulate deep learning networks

Deep Learning Toolbox provides functions, apps, and Simulink blocks for designing, training, and simulating deep neural networks. You can visualize and interpret predictions, verify network properties, and compress networks with pruning, projection, or quantization. You can also generate C/C++, CUDA^®, and HDL code for trained networks (with MATLAB Coder, GPU Coder, or Deep Learning HDL Toolbox).

The toolbox provides interfaces to other AI frameworks, enabling inference in MATLAB and Simulink and allowing the import of models from PyTorch^®, TensorFlow^®, Keras™, and ONNX^® into MATLAB.

The Deep Network Designer app lets you design, import, edit, and analyze networks. The Time Series Modeler app lets you train and compare models for time series prediction without writing code.

Open and Explore an Interactive Example in MATLAB Online

Photo of a PCB titled “Predicted Defects” with three annotations labeled “missing_hole.”

Design Engineered Systems

Create and use explainable, robust, and scalable deep learning models for virtual sensors, automated visual inspection, reduced order modeling, wireless communications, computer vision, and other applications.

AI for Engineers: Building an AI System (3:40)

Documentation | Examples

Screenshot of the Time Series Modeler app comparing trained deep learning models. The LSTM prediction closely matches the ground truth.

Build with Low-Code Apps

Visualize, build, train, and compare AI models for time-series data using the Time Series Modeler app. Accelerate network design and analysis for built-in and Python-based models using the Deep Network Designer app.

Interactively Design Networks with a Low-Code App (3:30)

Documentation | Examples

Virtual sensor modeling screenshots and a line graph comparing truth, EKF, deep learning FNN, and deep learning LSTM.

Integrate Models in Simulink

Use deep learning with Simulink to test the integration of deep learning models into larger systems. Simulate models based on MATLAB or Python to assess model behavior and system performance.

Developing and Embedding AI-Based SOC Estimation for BMS Using MATLAB (7:52)

Documentation | Examples

Flowchart showing the import of models from TensorFlow, ONNX, and PyTorch, and the export of models to TensorFlow and ONNX.

Interoperate with PyTorch and ONNX

Exchange deep learning models with Python-based deep learning frameworks. Import PyTorch, TensorFlow, and ONNX models, and export networks to TensorFlow and ONNX. Co-execute Python-based models in MATLAB and Simulink.

Interoperability Between Deep Learning Toolbox, TensorFlow, PyTorch, and ONNX

Documentation | Examples

Verification and validation process for AI-enabled systems, based on and adapted from an EASA diagram.

Verify and Explain Model Behavior

Verify robustness and reliability of MATLAB, PyTorch, and ONNX deep learning models using formal methods and out-of-distribution detection with the AI Verification Library. Visualize training progress and activations and explain network predictions.

What Is Explainable AI? (3:45)

Documentation | Examples

Screenshot of the Deep Network Quantizer showing a net layer graph, calibration statistics, and a validation summary.

Compress Neural Networks

Compress deep learning networks with pruning, projection, or quantization to reduce their memory footprint and increase inference performance. Assess inference performance and accuracy using the Deep Network Quantizer app.

Quantizing a Deep Learning Network in MATLAB (2:50)

Documentation | Examples

Diagram showing MATLAB and Simulink code generation for deploying deep learning models to CPUs, GPUs, MCUs, and FPGAs.

Generate Code for Deployment

Automatically generate optimized C/C++ code (with MATLAB Coder) and CUDA code (with GPU Coder) for deployment to CPUs and GPUs. Generate synthesizable Verilog^® and VHDL^® code (with Deep Learning HDL Toolbox) for FPGAs and SoCs.

Generate Generic C/C++ Code for Deep Learning Networks in Simulink (5:10)

Documentation | Examples

Product Resources:

Documentation Examples Videos Technical articles Functions Models Tutorials Requirements Release notes

Mercedes-Benz Simulates Hardware Sensors with Deep Neural Networks

“This was the first time we were simulating sensors with neural networks on one of our powertrain ECUs. Without MATLAB and Simulink, we would have to use a tedious manual coding process that was very slow and error-prone.”

View more customer stories

Deep Learning Toolbox FAQs

Deep Learning Toolbox provides functions, apps, and Simulink blocks for designing, implementing, and simulating deep neural networks such as CNNs, LSTMs, and transformers.

You can create convolutional neural networks (CNNs), LSTMs, GANs, and transformers, or perform transfer learning with pretrained models.

Yes, you can import PyTorch, TensorFlow, and ONNX models for inference, transfer learning, simulation, and deployment, and export networks to TensorFlow and ONNX with a single line of code.

The Deep Network Designer app lets you design, edit, and analyze networks interactively, import pretrained models, and export networks to Simulink.

The Time Series Modeler app lets you visualize, preprocess, and load time series data; design, edit, and train neural networks interactively; and export networks to Simulink for time series regression and forecast problems such as virtual sensor modeling.

Yes, the toolbox supports GPU acceleration to speed up deep learning training and computations.

Yes, you can automatically generate optimized C/C++, CUDA, and HDL code for trained networks for deployment to CPUs, GPUs, MCUs, and FPGAs.

You can visualize training progress and activations; use Grad-CAM, D-RISE, and LIME to explain network results; and verify the robustness and reliability of deep neural networks with the AI Verification Library for Deep Learning Toolbox.

Yes, you can analyze networks for compression using the Deep Network Designer app, and then compress networks with quantization, projection, or pruning to reduce memory footprint and increase inference performance using the Deep Network Quantizer app.

Yes, you can integrate deep learning models into Simulink to test integration with larger systems and simulate models based on MATLAB or Python.