The community’s demand and usage for machine learning and AI has grown rapidly. The activities of the community have led to the emergence of several new AI frameworks that significantly simplify the use of AI in applications. In this article, we will introduce some of the frameworks to help get started with AI-assisted chatbot development.
Tensor Flow
The popular open-source framework from Google supports Deep Learning and high-performance numerical computations. In TensorFlow, mathematical operations are represented as graphs. A graph is represented by the sequential flow of all operations to be performed by TensorFlow. The framework is released under the Apache 2.0 license and can therefore also be used for commercial applications.
Characteristics
- Highly scalable multi-programming interface for simple programming
- Strongly growing open source community and supported by large companies like AirBnB, eBay, Dropbox, and Coca-Cola
- Extensive documentation with examples
Advantages
- Framework is a Python library that calls C++. The library can be included by other languages (e.g. Java).
- Framework is scalable and can run on any CPU and GPU
Disadvantages
- Decision making/prediction requires the framework to process the input data through multiple nodes, which is time consuming
- There are few pre-built AI models
Apache Mahout
Mahout is a distributed linear algebra framework and a mathematically expressive Scala DSL that allows users to implement their own algorithms. Apache Spark serves as the preferred distributed back-end.
Characteristics
- Focuses on creating scalable machine learning algorithms
- Implements common machine learning techniques such as: Recommendation, Classification, Clustering
- Propagation through mathematically highly expressive Scala DSL
- Support for various distributed back-ends
Advantages
- Clustering support
- Use of Java libraries for computing operations (thus more performant)
Disadvantages
- Ability to integrate Python libraries not so well
- Spark MLib has faster arithmetic operations
Spark MLib
MLlib is Apache Spark’s scalable machine learning library with APIs for Java, Scala, Python, and R. The project originated at the University of California at Berkeley in 2014 and was released as an open source project under the Apache license. The Spark architecture consists of several components, some of which are interdependent: Core, SQL, Streaming, MLlib. MLlib and SparkML are the function libraries that contain typical ML algorithms and enable distribution.
Characteristics
- High performance
- Runs on different architectures
Advantages
- Performing iterative calculations and thus suitable for processing large amounts of data
- Support of various programming languages
Disadvantages
- No support for real-time processing
- Problem with small file and no file management system
- Less available algorithms
Touch
Torch is an open-source scientific framework with extensive support for machine learning algorithms that puts GPUs at the forefront. Torch is easy to use and efficient thanks to the LuaJIT scripting language and the underlying C/CUDA implementation.
Characteristics
- Torch contains a large ecosystem of community-driven packages in machine learning, computer vision, signal processing, parallel processing, image, video, audio, and networking, and builds on the Lua community
- Core of Torch are the neural network and optimization libraries with high flexibility in implementing complex neural network topologies
- Neural network graphs can be created and parallelized across CPUs and GPUs
Advantages
- High flexibility in terms of languages and integrations
- High speed and efficiency of GPU utilization
- Provision of models to train the data (pre-built models)
Disadvantages
- Complex documentation
- Few examples (tutorials, examples) for quick use of the framework
- Lua as a programming language less common than Python
Amazon Machine Learning on AWS
Is a big player in the AI community and provides support for the development of self-learning tools for commercial applications. Models can be created, analyzed, trained and evaluated. The vendor provides a wide range of AI tools for the development and use of AI models.
Characteristics
- Customizable AI tools for different application purposes and user roles (newbie, data scientist or developer)
- High data security (data encrypted)
- Comprehensive data analysis and understanding tools
- Wide range of integrations to various data sources
Advantages
- Use Amazon AI framework APIs instead of complex code development
- Suitable for commercial and professional AI applications
Disadvantages
- Limited flexibility, as the use of custom learning algorithms is not possible and only the available framework algorithms can be used
- Limited data visualization
- No option for on-premise (operation only in the cloud)
Caffe
The library was developed by Berkeley AI Research (BAIR) and contains a variety of algorithms and deep learning architectures for classification and cluster analysis of image data. Convolutional neural networks (CNNs) are also supported by the library. The primary programming interface is Python and MATLAB. In order to use Caffe in a distributed manner, the library has been integrated into Apache Spark (see caffeonspark).
Characteristics
- Models are written in plain text schemas
- Supported by large open source community
Advantages
- Modeling of CNN (Convolutional Neural Networks) is supported
- High efficiency in the calculation of numerical tasks
Disadvantages
- Caffe must be developed using mid- or low-level APIs, which limits the configurability of the workflow model and restricts most development time to a C++ environment
- Less suitable for complex data (exception for visual processing of images)
Theano
Theano was a Python library (published by the University of Montreal) for efficiently defining, optimizing, and evaluating mathematical expressions involving multidimensional arrays. The library is now continued as aesara. This library is also based on Python.
Characteristics
- Focus on general tools and methods for calculating mathematical expressions
- No prefabricated models
- Calculation instructions are implemented in C++ or CUDA code
Advantages
- Efficiently optimized for CPU and GPU
- Support for all data-intensive applications
Disadvantages
- Theano is obsolete and no longer maintained. Successor is aesara
- Used mainly for academic purposes
Microsoft Cognitive Toolkit (CNTK)
The deprecated open-source toolkit supports distributed deep learning and is based on neural networks. It describes neural networks as a series of computational steps over a directed graph. It is suitable for creating, training and evaluating neural networks with large data sets. Models and computations can be distributed across different workstations and GPUs. Toolkit supports the programming languages C#, BrainScript, Python and Java for evaluating the models.
Characteristics
- Optimized for efficiency, scalability and speed
- Additional features such as supervised learning models
- High and low level APIs
Advantages
- Training of neural networks is fast and efficient, because learning process can be distributed to different servers
- Support for Generative Adversarial Networks (GANs), Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs)
Disadvantages
- Toolkit is outdated and is no longer being developed (last update on GitHub in 2019)
- No mobile ARM support
Accord.net
The framework was a .NET machine learning framework combined with audio and image processing libraries written entirely in C#. It is particularly suitable for developing neural networks for audio and image processing. Accord.net is released under the Gnu Lesser Public License.
Characteristics
- Wide range of sample models and data sets
- Mature framework with a history of over 10 years
- Accord.NET project has been discontinued
Advantages
- Accord.NET is extensively documented and efficiently manages numerically intensive calculations and visualizations
- Implementation of algorithms and signal processing can be easily performed
Disadvantages
- Accord.NET project was archived after approx. 15 years and is no longer being developed further
- Performance lower to other better known frameworks
Conclusion
The list of frameworks should give you a short overview. Each of these frameworks has different characteristics and is sometimes more and sometimes less suitable for certain tasks. The respective use case is decisive in order to be able to select a suitable framework.
Do you want to learn more about the subject as well as which AI frameworks are suitable for your enterprise chatbots? Learn more here.