In Too Deep: Deep Learning vs. Deep Reasoning

A great way to evaluate a team’s prospects is to look at a depth chart describing the athletes available to fill each position. Can we do the same with AIs? Fortunately, there are multiple tools and techniques people can use to improve the quality, reliability and trustworthiness of AI-based results. In today’s post I will compare and contrast Deep Learning in Generative AI and Deep Reasoning in Cognitive AI. Today these two approaches are mutually exclusive though appearances may be deceptive. I’ll describe why these distinctions are important and conclude with a perspective on prospects for safe Artificial General Intelligence or AGI.

___________________________________________________________________________________________________

Deep Learning

Deep learning is the technology Generative AI uses to construct Large Language Models (LLMs). Using multi-layered artificial neural networks that mimic the visual cortex in mammalian brains they cyclically ingest massive datasets to learn complex patterns. Such systems can create realistic content such as text, images, and audio. They don’t analyze information – they learn how information is structured then predict probabilities based on new inputs formatted as English Language Prompts.

Artificial Neural Networks are modeled after brain synapses, using multiple “hidden” layers between the input and output to extract deep, abstract features from data. From vast amounts of information, deep learning models learn by recognizing patterns and modeling the statistical probabilities of how words, pixels, or symbols are interrelated. With massive repetition, models trained on this data serve as prediction engines generating new content piece-by-piece by predicting what elements should logically come next.

Deep Learning Architectures include:

• Transformers form the backbone of LLMs like ChatGPT. They focus on capturing the context and relationships between words in a sequence to generate fluent, human-like text.
• Generative Adversarial Networks (GANs) are used heavily in image generation using two neural networks that operate in tandem: 1) a generator that creates new images, and 2) a discriminator that tries to tell if the images are fake or real, continuously improving the quality of the output.
• Diffusion Models are used for image generators by gradually adding noise to an image and then teaching the deep learning model to reverse the process to learn how to generate an image from random static

Deep Reasoning

Unlike Generative AI’s implicit knowledge based on patterns, cognitive AI stores concept and context knowledge explicitly as data in graphs or files. Concept learning in cognitive AI is a system’s ability to learn and understand abstract ideas or categories from remembering and categorizing examples. This is analogous to how humans learn. This involves identifying patterns, relationships, and features within data to form a mental representation of a concept, which can then be used to classify new, unseen examples. It’s a fundamental aspect of cognitive AI, enabling machines to go beyond simple pattern recognition and develop more sophisticated understanding of the world. Techniques include:

• Learning from examples is an approach where Machines are presented with labeled data (supervised learning) or unlabeled data (unsupervised learning) to identify patterns and features associated with a specific concept. They then extract characteristics or features of a concept using Abstraction, rather than just memorizing individual examples.
• Generalization is a form of classification or categorization that moves from specific examples to new, unseen instances with shared characteristics to recognize the same or similar concepts in different contexts and variations.
• Categorization forms concept taxonomies or hierarchies that enable the system to group similar objects or ideas into meaningful domains and categories.

Producing Results

Generative AI neither understands nor considers – it just predicts probabilities. Cognitive AI uses reasoning and inference on explicitly learned concepts to interpret new situations or data. Cognitive AI systems use various human-like computational models to simulate concept learning reasoning. These models may include neural networks, but more commonly use Bayesian models, and symbolic reasoning systems. Some cognitive models process multisensory information (like images and sounds) and text-based representations to build more comprehensive and nuanced concept representations. A semantic control system may be used to coordinate and integrate different types of concept representations to improve consistency and coherence.

Some example uses of Cognitive AI systems include:

• Object Recognition where a system learns to recognize different types of objects or organisms by being shown pictures and identifying features like shape, color, and size.
• Natural Language Understanding where a system learns to understand the meaning of words and sentences by analyzing text and associating words with concepts.
• Problem Solving and Analytics where a system uses learned concepts to solve problems by applying appropriate strategies and actions.

Benefits of Concept Learning in AI:

• Improved Accuracy: By learning concepts, AI systems can make more accurate predictions and classifications.
• Increased Adaptability: Concept learning enables AI systems to adapt to new situations and data more effectively.
• Enhanced Reasoning and Decision Making: Learned concepts allow AI systems to reason and make decisions in a more human-like way.
• No Hallucinations and Improved Auditability: Trust is not a problem because answers are based on knowledge with referenceable sources – not patterns in black boxes.

In essence, concept learning in cognitive AI is about enabling machines to understand the world in a way that is more similar to human cognition, allowing them to go beyond simple pattern recognition and develop more sophisticated understanding of concepts and categories.

Toward AGI

Both Generative and Cognitive AIs, as well as other AI learning and reasoning models will continually grow together toward AGI. Concept learning and results can be curated and managed more easily to reduce bias so alignment is much easier to manage with Cognitive AI. We have seen over and over again how the Black Boxes endemic to Generative AI impair alignment and require complex and costly add-ons to protect users and companies from damaging false information and hallucinations. More startups are providing such add-ons to help plug this gap and the results are beginning to look promising.

To avoid a world where deep fakes and simple errors require constant vigilance, or to adapt to the status quo that already demands careful review and fact-checking, consumers and companies are looking for trustworthy solutions to automatically validate AI results and provide greater comfort so users can take advantage of the amazing new capabilities popping up every day in this transition period between the information age and the age of knowledge.