Examples of Potential Applications
-
Defense
Imagine a multimodal model that can understand the text content of a military OPORD (operational order) and the associated maps that go with it. The OPORD conforms to a standard format and the maps have NATO-standard icons and other detailed markings indicating unit movements, areas of operation, supply routes, etc.
Such a system could:
Generate tactical scenarios for training in VBS4
Suggest possible enemy COAs (courses of action)
Predict potential civilian damage and casualties
Automatically incorporate new intelligence updates and detect new opportunities or threats
-
Education
We can apply large language models in several ways to further online education. They allow us to automate the assessment of free-text student answers.
We can:
Score answers
Explain the scoring
Suggest how to improve the answers
Provide detailed highlighting of text areas that are exemplary or need improvement
Generate counterexamples to incorrect parts of an answer
-
Business
Chatbots can explain technologies, products, and services, but what they say must be reliable. Here, we can apply a synthesis of the generative language capabilities of neural approaches, coupled with the precision (e.g., constraint-based reasoning) and knowledge (e.g., knowledge graphs) of symbolic approaches.
Chatbot-based dialogue agents can:
Answer questions
Instruct customers on product use
Assist in troubleshooting, while explaining their reasoning
Core Technologies
-
Natural Language Processing (NLP) and Natural Language Understanding (NLU)
Natural language processing (NLP) treats text as data, but does not necessarily build an “understanding” (shared model) of a situation. In contrast, natural language understanding (NLU) goes further: It creates a formal representation of the situation that can be queried, combined with other knowledge, and used to infer new facts.
With NLP we can:
Build natural language interfaces for planning or control. Constraint reasoning systems and automated planners allow detection and rejection of
Infeasible interpretations or plans
Hallucinated knowledge
With NLU we can:
Analyze documents that support formal reasoning, e.g., creating
Argumentation or other discourse relations that show how sentences relate to each other
A formal sematic knowledge representation, such as BABEL output, which represents WordNet word senses (synsets) for verbs and nouns in each sentence
-
Multimodal Reasoning
Multimodal reasoning incorporates two or more modalities of reasoning, most commonly natural language processing and computer vision. We can combine visual interpretations (e.g., figures and diagrams) with textual interpretations. Images can be created from textual descriptions. Text can be created describing images. Further, a single joint understanding of text complemented by diagrams can be leveraged.
-
Knowledge Graphs and Symbolic Reasoning
Knowledge graphs provide highly-precise human-curated sources of knowledge that can augment large language models to assist in entity disambiguation, data augmentation, and filtering / reranking generated results.
More generally, deep learning lacks explainability and the ability to make reliably correct deductive inferences. Symbolic reasoning approaches complement deep learning approaches by adding automated planning, explicit knowledge representation, and provably correct reasoning.
Advanced Research and Applications
-
Neuro-symbolic Architectures
Both neural and symbolic approaches have complementary strengths and weaknesses. Neurosymbolic hybrids can combine the best of these two approaches.
Here are some examples:
A blackboard control architecture that can incorporate both neural perceptive modules with symbolic automated planning and explanation models.
A neural architecture that can translate text into BabelNet, a knowledge representation for documents. Symbolic reasoning can then make additional inferences from background knowledge. Q/A can proceed from the augmented knowledge representation.
-
Cognitive Agent Architectures
Most AI applications, both neural or symbolic, are narrow AI and not AGI (artificial general intelligence). Narrow AI is domain-specific: It is typically designed to do one highly specific task very well, such as to classify images into 1 of 100 categories using ImageNet for training. Narrow AI in symbolic AI was typified by expert systems, such as MYCIN, which could diagnose causing severe infections such as bacteremia or meningitis.
In contrast, AGI requires handling multiple problem-solving domains. Cognitive agent architectures, such as ACT-R, SOAR, ICARUS, and PRODIGY have been developed for cross-domain problem-solving. Most of these are based on early Newell and Simon work on GPS, but then extended to handle to chunking or reinforcement learning. These cognitive agent architectures also differ in emphasis, with some, like ACT-R, placing high importance on modeling human cognitive capabilities. Others, such as Ron Sun's CLARION, investigate neuro-symbolic approaches to cognitive architectures.
ChatGPT and GPT 4 provide an important step forward in AGI, yet still remain limited in their ability to learn on the fly, to tackle problems in domains with limited training data, and in their ability to invent new knowledge itself. They are useful in generating text, code, and image data to greatly accelerate productivity, but still require checks to rule out hallucination, i.e., errors in their results, and to ensure instructions are carried out as intended. For domains where mistakes can be critical, we still need higher reliability, additional planning capabilities, an ability to explain reasoning, and an ability to learn from human advice.
Our services.
-
Dream it.
It all begins with an idea. Maybe you have an idea about how you can apply AI to your business that you want to explore. You want to know what the state of the art is, and would like to know what you would need to implement your idea. PostModern AI can provide expert consulting to help you assess your risks and to assess practical details you would need to realize your vision.
-
Build it.
Build it and they will come. We can help you build a proof of concept prototype for your idea. Your stakeholders can more easily see the value of your AI approach if they can see how it works for some typical examples, especially when they can interact with the demo.
-
Grow it.
Let us help you patent your idea and start your own business if you like, or get started in your current company. Either way, we can help you get started and act as consultants to help the project lift off and then keep it on track as it grows and conditions change over time.