Deep Learning No Coding

Our product empowers people to learn about and work with deep learning and neural networks without any coding.

In Brief

Have you ever wished you had deep learning technology available at your fingertips? Possibilities are endless, here are a few examples for business and personal use:

  • pattern recognition and forecasting based on your data set
  • data classification
  • predicting finance markets and sports results
  • creating your own music based on the music you like
  • applying artistic styles to your photos
  • teaching computer how to play certain games.

To do these things, you need to know the math behind them and have programming skills to write code. Our product (DLNC) allows you to apply the magic of deep learning without the need to write any code. Those new to deep learning can learn hands-on amazing abilities of this technology. For those already familiar with deep learning, DLNC will save you tons of time from coding, data processing, debugging neural networks, keeping track of training runs, along with many other features conducive to the professional application of deep learning. The three-minute video below shows how simple and easy it is to build from scratch, train, and use your own neural network with DLNC.

Neural style transfer converts a photo into the style of a painting

Neural style transfer converts a photo into the style of a painting

Ideas for Inspiration

Let’s imagine you want to come up with an idea for a story or a blockbuster movie. Wouldn’t it be cool if a computer can give you a hint based on a collection of top-rated movies? Great! Let’s build our own neural network that would do just that. So, we loaded a bunch of plot taglines from The Movie Database and will show you what the computer had come up with on its own. For best result, please watch the video below in full screen mode.

Please note that the current user interface is a functional prototype. We plan to substantially improve the look, feel, and usability of our user interface for the final release.

The entire process took less than 10 minutes with DLNC and is simpler than working with most video editing software.  Yes, users will need to understand the various deep learning concepts and components in the software, but that’s the fun and educational part of working with deep learning, and our hands-on training in DLNC will help with that.

To achieve a similar result without DLNC, someone who is not a software developer would have to learn to code in a programming language that supports deep learning, then learn a deep learning library like Torch or TensorFlow, a process that can take months to years of training and thousands to tens of thousands of dollars in training cost to master.

The Big Picture

Applications with intelligence are already all around us, and software coding could soon take a back seat to machine learning. The economic and technological divides in our society are widening. The high concentration of deep learning and artificial intelligence expertise at a few large tech companies could expand these divides. We want to live in a future where vital technologies, like deep learning, are not in the hands of a few. We think that anyone sufficiently motivated and empowered can apply deep learning to become a teacher of machines. With that in mind, we are working to provide a software platform that allows users to build and train neural networks with no coding, along with a marketplace where users can share useful data or neural networks for free or pay; so that powerful deep learning technologies are readily available to all.

What is DLNC?

DLNC (Deep Learning No Coding) is a software platform that simplifies the process of creating and applying deep learning solutions to a range of tasks and problems.  DLNC will help users to create AI applications faster and easier by allowing users to work with all aspects of deep learning through a visual user interface.  To acquaint new users with deep learning, our application will be accompanied by a series of fun hands-on exercises and labs designed to build intuition and understanding of all aspects of deep learning.  We will also provide users with a number of pre-built and/or pre-trained neural networks so that users can get productive in working with deep learning right out of the gate.

For advanced users already familiar with deep learning, our application provides an accelerated workflow with many useful features, such as automatic hyperparameter tuning, useful charts and visualizations, record keeping of all training runs, and more.

Our approach is to create an easy to use deep neural network builder and trainer application. Users will build and train deep neural network via point-and-click and drag-and-drop instead of writing code. We will provide support for all types of neural works currently in use, such as convolutional neural networks for image processing, recurrent neural network for text and sequenced data processing, and generative adversarial networks for content generation, and reinforcement learning, or even a combination of different network types. Our neural network builder application will include great visualizations to help build our users’ intuition on how well their neural networks work. Thus, our software will not only make working with deep learning easier, it will also train our users to work better with this technology over time and assist with improving the tuning and selection of deep learning approaches best suited for their data and application.

The Benefits of DLNC

We believe people with various backgrounds can work with deep learning having the right tools and guidance. That is what we are working to make possible. While there are a lot of high level math involved in deep learning and neural networks, at a practical level there are many simple principles and ideas that anyone can grasp, if they have the right software tool to help, like DLNC. For a comparison, witness the massive amount of 3D visual content in the world right now: the myriads of movies, apps, and games. These contents are created by many thousands of artists using software tools to bring worlds and characters to life in 3D without knowing much or even any of the complex math underlying 3D graphics.

We want to create the same kind of empowerment for people of all backgrounds to harness the power of deep learning. In addition to the powerful software that we’re building, we are also providing hands-on and easy to follow instructions on the key ideas and components behind the power of deep learning and neural networks with our DLNC Guide to Deep Learning (see our FAQ below for more details). And no artistic skill or experience is required to work with deep learning :).

Deep learning is about teaching machines with data. Preparing data for machine consumption is an intricate task given the enormous variety of data that could be used for machine learning. We will provide a variety of useful data processing routines for various data types, from images to text to raw numbers to videos, as well as support for composing multiple processing steps. Users will have great flexibility and power in how to process data for use in their models. For users new to deep learning, we’ll have data sets with fully configured processing to get them started in deep learning right away. We, along with our user community, will also add new data sets fully configured with all the proper processing steps to our platform and marketplace over time.

Our powerful model builder allows easy construction of neural networks via drag-and-drop. Users can build and manage very deep network with ease and can even link up multiple network types, like combining recurrent and convolution layers, to create complex, state of the art models that can be trained and used on the edge or in the cloud. All layer parameters are visible at a glance and are fully editable. Any layer parameter can be turned into a hyperparameter to allow for different values or even dynamic decay or alteration during training.

Deep learning experts can jump right in and build complex networks in minutes, while our hands-on labs and exercises will take new users step-by-step through various network and layer components to build understanding and intuition.  By working gradually from simple networks to more complex models, users will gain practical know-how for constructing custom neural networks.  A variety of pre-built / pre-tuned networks and models will be shipped with the software to get users productive with deep learning right away.

Our flexible and expandable training interface provides users with full control over the training of a wide variety of models and data. Helpful metrics and visualization of training progress is also provided, as well as automatic hyperparameter tuning to help with iterating toward great model performance without constant supervision.  We also keep meticulous records of each training run, including all associated hyperparameter values and performance metrics.  Our goal is to provide a well-organized, simple to use yet fully informative training workflow to make this time-consuming process as efficient as possible.

Our development environment runs identically on Windows and Linux right now, with macOS support on the way.  We will provide full single and multi GPU support for training or inference either on local computers or in the cloud.

For users who are also software developers, the Professional Edition of our software platform will allow for the export of deep learning functionalities for deployment to mobile devices, personal computers, and the cloud, either stand-alone or as part of another application.  The Professional Edition will also support distributed training and inference of large models across multiple computers with centralized management and reporting.

The Marketplace

Our goal is to engage many more people in working with AI technologies and letting more people to benefit from the growing use of AI. Integral to our approach is a marketplace where our user community can easily share AI components with other users for free or pay. We hope to build a large user base to help keep prices low and allow all our users instant access to many useful models and data sets. This way many users can benefit from deep learning innovations without ever needing to build their own neural networks.  Furthermore, engaging the cumulative brainpower of many more users (who currently have no access to deep learning at all) will naturally speed up innovations in deep learning and spur the cost-effective applications of this technology beyond the confines of large tech corporations and to small businesses and individuals.

The DLNC prototype training interface with hyperparameter tuning, training progress indication, chart visualization, and training history including all hyperparameter values and test results.

The DLNC prototype training interface with hyperparameter tuning, training progress indication, chart visualization, and training history including all hyperparameter values and test results.

Key Features of DLNC, All Editions

  • Easy access to a variety of public data sets for deep learning.
  • Easy intake of custom data sets.
  • Easy sequencing of processing steps and setting data processing parameters.
  • Data visualization before and after processing.
  • Powerful visual neural network builder with a toolbox of layers grouped conveniently for different tasks.
  • Building large neural networks by placing and rearranging layers with ease via drag-and-drop.
  • Customization of layer parameters made easy with point-and-click.
  • Visual model builder where previously built networks can be combined into sophisticated models that are highly difficult to manage directly by coding.
  • Flexible and expandable training interface is made simple with all the hyperparameters available at a glance for tweaking.
  • Automatic hyperparameter tuning to help with iterating toward great model performance without constant supervision.
  • Helpful metrics and visualization of training progress, charts and estimates.
  • Training time reduction by setting validation accuracy targets and abandoning under-performing training runs.
  • Support for transfer learning, including pre-loading of existing weights before training and selecting of specific layers to train.
  • Full history log of every training run so that user can see immutable packages of a model with a particular set of hyperparameters.
  • Training snapshots which users can use to resume the process.
  • Visualization of layer outputs and weights.
  • Ships with a variety of pre-configured data sets and pre-tuned models ready for immediate use.
  • Support for training and inference on single or multiple GPUs locally or in the cloud.
  • Built-in internationalization with more languages added on demand.

Features Specific to the Professional Edition

  • Export trained neural networks as applications and libraries that can be deployed to mobile devices, personal computers, and the cloud and used without DLNC.
  • Plug-in architecture for adding new capabilities via external code.
  • Source code generation for TensorFlow and potentially other deep learning platforms.
  • Support for distributed training and inference on multiple computers locally or in the cloud with a centralized management console.
Copyright © 2017