With NXS Core, you get access to a wide range of low-level algorithms and ease-of-use features that help you develop and program AI solutions more efficiently. NXS Core offers modules for neural network-based learning, (deep) reinforcement learning with prediction and control algorithms, k-means clustering and natural language analysis.
NXS includes a number of modular tools that allow developers to create artificial neural networks of different types and sizes and later integrate them into custom applications. The software library gives you access to a comprehensive set of algorithms to create, train, validate and test supervised networks. The many algorithms and available options allow you to create fully customized networks that can be fine-tuned to work on a huge variety of different problems.
- Flexibly design networks that are ideally suited to your problem. The NXS ANNs are feedforward, multilayer perceptron (MLP) artificial neural networks that allow for arbitrary network shapes and support an unlimited number of perceptrons and (hidden) layers.
- Get access to a comprehensive set of functions and algorithms. NXS offers nonlinear activation functions, robust backpropagation algorithms for supervised learning, an extensive set of initialization functions and support for metrics and Stochastic Gradient Descent (SGD) optimization.
- Flexible options for learning rates. Learning rates can be determined by either a fixed rate or an adaptive method.
- Build extremely efficient deep networks. The NXS ANN algorithms use minimal computational and memory resources, even for deep networks.
- Create complex networks more easily with the ANN factory. The factory notifies you if the network is not valid or if the predicted computation time is exceptionally high. It includes built-in helper functions to easily persist trained networks to the filesystem for later application in live production environments.
- Get comprehensive support for training data. NXS includes tools to read arbitrarily structured training data in CSV format, offers on-the-fly preprocessing (one-hot encoding, square augmented and custom transform) and includes a reader for IDX (image) datasets.
- Execute arbitrary code at iteration or epoch level with the integrated callback feature. This makes it possible to, for example, monitor learning progress using built-in error functions or run debugging operations during the computationally intensive network training phase.
Reinforcement learning algorithms learn by interacting with their environment. When the algorithm performs correctly, it is rewarded; when it performs incorrectly, it is penalized. In order to maximize rewards over the long-term, the algorithm must be able to reason about the long-term consequences of its actions vs. the short-term benefit. This makes it ideal for problems involving planning and scheduling, for example.
- Access a broad set of reinforcement learning algorithms for a wide array of real-world problems. Both predictor and control algorithms are included, which cover most needs in fields like robotic process automation, big data analysis and other areas where reinforcement learning can be highly beneficial.
- Easily define the problem to be solved. Blackzendo NXS gives you a simple, streamlined means of defining the problems that need to be solved, which makes it far easier to implement reinforcement learning in your AI applications.
- Minimize use of computational resources and memory. Heavily optimized reinforcement learning algorithms reduce consumption of both memory and computational resources during the learning process and during application in a trained system.
k-means clustering is commonly used in data mining. The goal of this algorithm is to find groups in unlabeled data sets, with the number of groups represented by the variable “k”. You can either manually set a number of clusters, or have the algorithm analyze the data set to determine an appropriate number.
- Optimized for speed and minimal memory consumption.
- Determine K by either manually setting a number of clusters or choosing from a selection of algorithms to determine “k” automatically.
NXS Core offers a comprehensive suite of NLP features that allow for realistic human-computer interaction and natural language analytics.
NLP Conversational Features:
- Easy-to-use knowledge base in AIML 2.0 format
- Robust, complex pattern-matching engine for natural-sounding responses
- Includes schema validation for your knowledge base files
- Pattern matching engine includes a built-in shadowing check
- Input/output matching can be either simple (direct matching of user input to IVE reply) or complex (matching multi-sentence input to dynamically generated multi-sentence output)
- A single instance of a conversational IVE can be used simultaneously by multiple users
- Synonyms can be matched using sets for improved language comprehension
- Key-value pairs can be defined using maps (e.g. linking “Bern” to “Switzerland” in a map titled ‘Capitals’)
- Maps can handle both static values and dynamically generated values
- Multiple patterns can be mapped to a single reply using references
- A single pattern can contain multiple wildcards
- Wildcards can be assigned varying priority, allowing you to precisely control pattern matching
- Multiple wildcards in a pattern can be accessed by index
- The IVE can access the conversation history (e.g. what the user said two messages before) for improved continuity
- Replies can be randomly selected from a list. Random lists can also be nested
- The system support emoticons
- User data can be persisted, allowing the creation of a user profile over time
- You can define topics which the IVE can switch between to guide the conversation
- Conditionals can be used to react to user input or dynamic variables. Conditionals can also be nested
- Custom tags can be used for client-side formatting. For example: sentence splitting (creating a new line on the client), choices (buttons, drop-down lists, etc.)
- The IVE’s emotional simulation can be fully integrated into chat responses
- User input can be analyzed for emotion/sentiment, which will influence the IVE’s mood
- As the IVE’s mood changes, it can adapt its tone and conversational style accordingly (if you allow it to do so)
- You can assign personal details such as astrology sign, birthday, etc. to the IVE to enrich its small-talk
- Integrate external systems can be integrated by triggering calls to external asynchronous services or functions directly from the knowledge base
- You can integrate dynamic data into the knowledge base using a user interface
NLP Analytics Features:
- Valence-aware sentiment reasoning. Our sentiment reasoner is specifically attuned to sentiments expressed in social media and also works well on texts from other domains. More specifically, it is able to correctly understand and handle:
- typical negations (e.g., “not good”)
- use of contractions as negations (e.g., “wasn’t very good”)
- conventional use of punctuation to signal increased sentiment intensity (e.g., “Good!!!”)
- conventional use of word-shape to signal emphasis (e.g., using ALL CAPS for words/phrases)
- use of degree modifiers to alter sentiment intensity (e.g., intensity boosters such as “very” and intensity dampeners such as “kind of”)
- many sentiment-laden slang words (e.g., ‘sux’)
- many sentiment-laden slang words as modifiers such as ‘uber’ or ‘friggin’ or ‘kinda’
- many sentiment-laden emoticons such as 🙂 and 😀
- sentiment-laden initialisms and acronyms (e.g., ‘lol’)
- Bias detection. The ‘Bias Statement Detector’ (BSD) computationally detects and quantifies the level of bias in sentence-level texts. Common linguistic and structural cues that are indicative of biased language are incorporated, including sentiment analysis, subjectivity analysis, modality (expressed certainty), the use of factive verbs, hedge phrases, and many other features.