Embedded deep learning for
everybody, anywhere, anytime

DeLTA-Lite makes it possible to build and deploy embedded deep learning models for palm-sized devices within one day and without any programming knowledge.

Advantages of DeLTA-Lite

Using DeLTA-Lite you can build an embedded deep learning model * by preparing just 2 things.

*The prediction system.

Annotated Data

Annotated Data
Labelled training data. (we can support you in creating the data with our data preparation service)

Target Hardware

Annotated Data
Devices that deep learning models run on

DeLTA-Lite includes cutting-edge technologies, methods, and tool-chains developed by LeapMind. You can use them without programming. All you need is your ‘training data’ and ‘hardware’ to build embedded deep learning models.

The DeLTA-Lite Process

Dataset Dataset
Design of the deep learning model
Training the model
Compressing the model
Converting to C-code
Compiling for hardware
reduces the 3-month process to a 1-day process!
Deploying onto hardware Deploying onto hardware

Until now it took a lot of time and money just to conduct field studies on the application of deep learning, letalone integrating it into the company’s operations. By using DeLTa-Lite, the time and cost is drastically reduced, making it easy to experiment with and integrate deep learning technology.

How to use DeLTA-Lite

You can build a model through the web browser interface in just 4 easy steps.

  • 1.Choose an application type.

    1 Choose an application type.

  • 2.Upload your training data.

    2 Upload your training data.

  • 3.Start the training.

    3 Start the training.

  • 4.Download your trained model.

    4 Download your trained model.

After selecting the tasks and uploading your training data it takes about 1 day * until your model is ready to be downloaded - just by following those simple steps.

*The exact time depends on the task and the amount of training data

Use cases

1. Food industry

1. Food industry

Deep Learning models can be trained to identify contaminations and anomalies based on images as part of the food production line. This can significantly reduce workload and maintenance cost imposed by manual inspection and operation of traditional sensors.

2. Automotive industry

2. Automotive industry

Based on cameras installed in a car, deep learning models can be trained to identify situations both inside and outside the vehicle, providing the driver with accessibility, functionality, and security features.

3. Construction industry

3. Construction industry

The task of identifying anomalies in construction or factories had previously been done manually by workers. By using deep learning it becomes possible to automate those processes based on images taken by drones or fix-point cameras.

4. Advertising

4. Advertising

Using deep learning to classify and analyze online images , new market trends can be identified. This can become an incentive for new projects and products, as well as marketing strategies.

Obstacles for building embedded deep learning models

In most cases, creating embedded deep learning models takes a long time.


  • Project planning

Picking a deep learning model

  • Investigating available approaches

Obtaining training data

  • Collecting data
  • Data annotation / labelling

Design, implementation,
training and fine tuning

  • Design of deep learning models
  • Model implementation
  • Training the model
  • Fine tuning hyperparameters
  • Model compression
  • Compilation

Deploying on hardware

  • Deploying
  • Verification and testing

Those are the basic steps for building an embedded deep learning model which can easily take more than 3 month.

To build an embedded deep learning model the following expert knowledge is essential

Model design

Model design

Expert knowledge on deep learning is necessary for good model design.

Model implementation, training and tuning

Model implementation, training and tuning

High level engineering is required for the implementation and training of high quality models.

Model compression

Model compression

A high level of research and development is essential to compressing a model without decreasing its performance.

Convert and compile

Convert and compile

Expert level knowledge of hardware design is necessary for embedding a model into devices like FPGAs.

Those expertise hardware and software techniques/skills/knowledges are required to make/build embedded deep learning models and deploying it/those onto a hardware.