On NeuroSymbolic AI

New to me, worth understanding in useful contexts ...

Neurosymbolic AI, By Don Monroe

Communications of the ACM, October 2022, Vol. 65 No. 10, Pages 11-13   10.1145/3554918

The ongoing revolution in artificial intelligence (AI)—in image recognition, natural language processing and translation, and much more—has been driven by neural networks, specifically many-layer versions known as deep learning. These systems have well-known weaknesses, but their capability continues to grow, even as they demand ever more data and energy. At the same time, other critical applications need much more than just powerful pattern recognition, and deep learning does not provide the sorts of performance guarantees that are customary in computer science.

To address these issues, some researchers favor combining neural networks with older tools for artificial intelligence. In particular, neurosymbolic AI incorporates the long-studied symbolic representation of objects and their relationships. A combination could be assembled in many different ways, but so far, no single vision is dominant.

The complementary capabilities of such systems are frequently likened to psychologist Daniel Kahneman's human "System 1" which, like neural networks, makes rapid, heuristic decisions, and the more rigorous and methodical "System 2." "The field is growing really quickly, and there's a lot of excitement," said Swarat Chaudhuri of the University of Texas at Austin. Even though "Neural networks are going to become ubiquitous, even more than they are today," he said, "not all of computer science is going to become replaced by deep learning."

A Long History

In the early years of artificial intelligence, researchers had high hopes for symbolic rules, such as simple if-then rules and higher-order logical statements. Although some experts, such as Doug Lenat at Cycorp, still hold hopes for this strategy to impart common sense to AI, the collection of rules needed is widely regarded as Unpractically large. "If you try to encode all human knowledge manually, we know that's not possible. That has been tried and failed," said Asim Munawar, a program director of neurosymbolic AI at IBM.

Neural networks also fell short of their aspirations in the 1980s and '90s, and artificial intelligence entered a long "winter" of reduced interest and funding. This situation changed a decade ago, however, largely due to the availability of enormous datasets for training, and massive computer power. Recent architectural innovations, notably attention and transformers, have driven further advances, such as the uncannily plausible text generation by OpenAI's large language model, GPT-3.

Deep learning does surprisingly well at generalizing, for reasons that are only partly understood. Despite impressive successes on average, however, these systems still make some odd errors when presented with novel examples that do not fit patterns they infer from the training data. Errors also can be created using maliciously altered data, sometimes in ways essentially imperceptible to people.

In addition, racial, gender, and other biases in the training data can be unintentionally enshrined by neural networks. Thus, for ethical and safety reasons, users often expect an explanation of how the networks came to a conclusion in medical, financial, legal, and military applications.

In spite of widespread concerns, these problems are not actually "limitations of deep learning systems," said Yann LeCun, chief AI scientist and a vice president at Meta, of the widely used supervised learning paradigm. LeCun, who shared the 2018 ACM A.M. Turing Award with fellow deep learning pioneers Geoffrey Hinton and Yoshua Bengio, believes that if users adopt "self-supervised learning, things that are not trained for a given task but are trained generically, a lot of those problems will essentially disappear." (LeCun regards explainability as a "non-issue.")  ... ' 

More Need for Common Sense

Good piece in Wired on the need for 'common sense'  reasoning to bring human Intelligence to AI.  Even a mention of CYC, and the years of work it has used to include human ideas.    But I have not heard any cases where it, or it successor Lucid, has suddenly sprung out in human common sense insight.   Some say that CYC just produced a very big, very brittle knowledge graph.   It just found lots and lots of patterns, sometimes useful, often not.     Its often not about the fact that patterns are there, but how humans can assemble them in useful ways.

How to Teach Artificial Intelligence Some Common Sense
We’ve spent years feeding neural nets vast amounts of data, teaching them to think like human brains. They’re crazy-smart, but they have absolutely no common sense. What if we’ve been doing it all wrong?

Five years ago, the coders at DeepMind, a London-based artificial intelligence company, watched excitedly as an AI taught itself to play a classic arcade game. They’d used the hot technique of the day, deep learning, on a seemingly whimsical task: mastering Breakout,1 the Atari game in which you bounce a ball at a wall of bricks, trying to make each one vanish.

1 Steve Jobs was working at Atari when he was commissioned to create 1976’s Breakout, a job no other engineer wanted. He roped his friend Steve Wozniak, then at Hewlett-­Packard, into helping him.

Deep learning is self-education for machines; you feed an AI huge amounts of data, and eventually it begins to discern patterns all by itself. In this case, the data was the activity on the screen—blocky pixels representing the bricks, the ball, and the player’s paddle. The DeepMind AI, a so-called neural network made up of layered algorithms, wasn’t programmed with any knowledge about how Breakout works, its rules, its goals, or even how to play it. The coders just let the neural net examine the results of each action, each bounce of the ball. Where would it lead?

To some very impressive skills, it turns out. During the first few games, the AI flailed around. But after playing a few hundred times, it had begun accurately bouncing the ball. By the 600th game, the neural net was using a more expert move employed by human Breakout players, chipping through an entire column of bricks and setting the ball bouncing merrily along the top of the wall.  ... " 

Lucid AI and Cyc: Seeking Common Sense Reasoning

In support of reading Rebooting AI, been exploring where the implementation of CYC has taken place.    We saw it demonstrated in the 80s.   Followed it here since.    There are some mentions of applications like Glaxo and Cleveland Clinic in the WP article https://en.wikipedia.org/wiki/Cyc  But not much  mention of its use by Lucid.ai.   Is it possible to deliver a complete and usable common sense knowledge base, even for domain areas?  Where now is deep understanding?  Quotes below are interesting, but dated.  Our own use of machine learning before it was sexy keeps me a student.

An AI with 30 Years’ Worth of Knowledge Finally Goes to Work

An effort to encode the world’s knowledge in a huge database has sometimes seemed impractical, but those behind the technology say it is finally ready.
by Will Knight  Mar 14, 2016 Technology Review

Having spent the past 31 years memorizing an astonishing collection of general knowledge, the artificial-intelligence engine created by Doug Lenat is finally ready to go to work.

Lenat’s creation is Cyc, a knowledge base of semantic information designed to give computers some understanding of how things work in the real world.

Cyc has been given many thousands of facts, including lots of information that you wouldn’t find in an encyclopedia because it seems self-evident. It knows, for example, that that Sir Isaac Newton is a famous historical figure who is no longer alive. But more important, Cyc also understands that if you let go of an apple it will fall to the ground; that an apple is not bigger than a person; and that a person cannot throw an apple into space.

And now, after years of work, Lenat’s system is being commercialized by a company called Lucid.

“Part of the reason is the doneness of Cyc,” explains Lenat, who left his post as a professor at Stanford to start the project in late 1984. “Not that there’s nothing else to do,” he says. But he notes that most of what is left to be added is relevant to a specific area of expertise, such as finance or oncology.  ... "

NELL and Machine Learning

In a re-examination of 'machine learning', the broader definition that includes but is not restricted to deep learning.   I recalled our look the CMU effort called NELL,  Never Ending Language Learner'.  At the time it was insufficiently oriented to our needs, any information about the actual use of its learning?  Here is the intro to it on the Carnegie Mellon site.

NELL: The Computer that Learns,   Professor Tom Mitchell

Tom Mitchell's two daughters are grown but watching his newest 'baby' learn to read is an unprecedented achievement.

Professor Mitchell leads the team that developed the Never-Ending Language Learner – NELL – a computer system that, over time, is teaching itself to read and understand the web.
"I've been interested for many, many years in how machines learn because I'm also interested in how humans learn," explained Mitchell, who heads Carnegie Mellon's Machine Learning department – the first and only department of its kind in the world. "NELL comes naturally out of that. The current machine-learning algorithms are very different in style than how you and I learn. They analyze a single data set, output an answer, and then you turn them off. That's not like us at all! The idea of NELL is to capture a style more like the on-going learning of humans."

Understanding language – the way humans do – depends on both context and background knowledge gained over time. So NELL scans the web – attempting to "read" hundreds of millions of web pages on a fact-finding mission.

For example, the repeated combination of a phrase like "New York City Marathon" in combination with other words has taught NELL to learn that it's a "race" and a "sports event." ...  "

See also:
Project website:  http://rtw.ml.cmu.edu/rtw/
On twitter:           https://twitter.com/cmunell

CYC: Logical Reasoning with a Knowledge Base

My recent post prompted me take a a look at the CYC site again, first time in a long time, here is an excerpt as to what they are about.    Further looking to see what the granularity of their knowledge base looks like, and how it can be leveraged to approximate common sense.    Also recalling our interactions, and why it did not work for us,  because common sense depends upon particular context too.   Is that being better considered better now?   Also how is machine learning, mentioned below, integrated into a knowledge bases?   There is still great need for an architecture that links common sense to a particular knowledge domain.

Cyc:  Logical Reasoning with the World's Largest Knowledge Base

Home of Smarter Solutions

The Cyc software combines an unparalleled common sense knowledge base with powerful inference engines and natural language interfaces to deliver human-like understanding and transparent explanations of its output and reasoning.  Cyc applications can stand alone or work in concert with pattern matching AI tools, such as Machine Learning, to deliver truly differentiated value.

ENTERPRISE SOLUTIONS:

Differentiated applications for your enterprise that deliver value only available with Cyc’s vast knowledge base and powerful inference engines.

CYC PLATFORM:

Develop smarter applications using Cyc’s Knowledge Application Development Environment. The Cyc knowledge base and inference engine are available for commercial application developers with support if necessary.

ABOUT US:

Cycorp is a leading provider of Artificial Intelligence technologies bringing a new level of common sense reasoning and understanding to a wide variety of software applications.   ..... "

Also about their work with Accenture.

Machine Common Sense

Been a challenge for along time. At DARPA and elsewhere.   One of our first challenges during the first AI revolution.     Putting humans in the loop is the easiest current solution, assuming you can leverage the common sense they have.     We engaged briefly with CYC on this, they were working with the US government then, busily encoding all human common sense, where is that now?

DARPA Embraces ‘Common Sense’ Approach to AI  By George Leopold in Datanami

The Pentagon’s top research agency is focusing its considerable AI efforts on the interim stage of machine intelligence between “narrow” and “general” AI.

The Defense Advanced Research Projects Agency (DARPA), which announced a multi-year $2 billion “AI Next” “campaign last month, is tightening its focus on teaching machine “common sense” reasoning. That capability remains beyond the reach of current AI constructs, and the research agency said it hopes to launch a “third wave” of AI technology that is adaptable while shedding light on the mystery of how machines learn.

Common sense reasoning is defined as “the basic ability to perceive, understand, and judge things that are shared by nearly all people and can be reasonably expected of nearly all people without need for debate.”

AI experts note the gap between AI inference and the ability to design systems that can draw directly on the rules of inference to achieve common sense reasoning. “Articulating and encoding this obscure-but-pervasive capability is no easy feat,” DARPA program managers note.

The lack of machine common sense is also seen as among the biggest barriers to advancing beyond narrow to general AI applications. The DARPA effort, dubbed “Machine Common Sense,” seeks to move beyond current frameworks that the agency considers “brittle” and lacking in semantic understanding. .... "

Never-Ending Learning Continues at CMU

Was reminded of this effort.  During our early work with AI, based on very large and messy rule bases, we tried to introduce learning capabilities.   I note this has been going on at CMU since 2010.  How might this connect with other kinds of applications like CYC?   Ultimately such efforts will build the foundation of broad AI.  Here an update, examining further and will now stay in touch.

Though technical, some useful discussion of the underlying motivations.

Here the site for the project, which includes a simple statement of purpose.

A paper of depth with statements of progress:  Never-Ending Learning

By T. Mitchell, W. Cohen, E. Hruschka, P. Talukdar, B. Yang, J. Betteridge, A. Carlson, B. Dalvi, M. Gardner, B. Kisiel, J. Krishnamurthy, N. Lao, K. Mazaitis, T. Mohamed, N. Nakashole, E. Platanios, A. Ritter, M. Samadi, B. Settles, R. Wang, D. Wijaya, A. Gupta, X. Chen, A. Saparov, M. Greaves, J. Welling 

Communications of the ACM, Vol. 61 No. 5, Pages 103-115
10.1145/3191513

Whereas people learn many different types of knowledge from diverse experiences over many years, and become better learners over time, most current machine learning systems are much more narrow, learning just a single function or data model based on statistical analysis of a single data set. We suggest that people learn better than computers precisely because of this difference, and we suggest a key direction for machine learning research is to develop software architectures that enable intelligent agents to also learn many types of knowledge, continuously over many years, and to become better learners over time. In this paper we define more precisely this never-ending learning paradigm for machine learning, and we present one case study: the Never-Ending Language Learner (NELL), which achieves a number of the desired properties of a never-ending learner. NELL has been learning to read the Web 24hrs/day since January 2010, and so far has acquired a knowledge base with 120mn diverse, confidence-weighted beliefs (e.g., servedWith(tea,biscuits)), while learning thousands of interrelated functions that continually improve its reading competence over time. NELL has also learned to reason over its knowledge base to infer new beliefs it has not yet read from those it has, and NELL is inventing new relational predicates to extend the ontology it uses to represent beliefs. We describe the design of NELL, experimental results illustrating its behavior, and discuss both its successes and shortcomings as a case study in never-ending learning. NELL can be tracked online at http://rtw.ml.cmu.edu, and followed on Twitter at @CMUNELL.  ... "

Assistants and Common Sense

Its actually not too often that assistants speak gibberish, they more often just admit to not knowing what was asked, when a human would understand readily.  That's usually better to diminish risk.  See my previous posts on Common Sense reasoning, which we worked on in the enterprise.  The idea of a challenge is good, it will at least scope the problem in contextual and current terms.   See the examples at the link below.  Also the notes on what 'fundamental limitations' are. 

AI assistants don’t have the common sense to avoid talking gibberish
A new test could prove that when it comes to language, today’s best AI systems are fundamentally limited.   by Will Knight in Technology Review

Siri and Alexa are clearly far from perfect, but there is hope that steady progress in machine learning will turn them into articulate helpers before long. A new test, however, may help show that a fundamentally different approach is required for AI systems to actually master language.

Developed by researchers at the Allen Institute for AI (Ai2), a nonprofit based in Seattle, the Arc Reasoning Challenge (ARC) will pose elementary-school-level multiple-choice science questions. Each question will require some understanding of how the world works.  .... " 

Reasoning with Common Sense

Can we teach machines common sense?  Which points back to our own look at the work Doug Lenat was doing with the Cyc Project. We wanted to create a means by which an agent could understand the basic terminology and goals of an organization like a company, and reason about it with the right language and direction.  That did not happen,  but the goal is still there.   And Cycorp has been active for 34 years, and produced a stable software release just three months ago.  We talked to them back then, but it didn't meet our needs. And every time we make some progress, like our broad use of assistants, now in many homes, we see the need again.

See the Cyc Wikipedia Article.  and

The Limits of Explainability
Academics, economists, and AI researchers often undervalue the role of intuition in science. Here's why they're wrong.
by Joi Ito in Wired ....

Paul Allen Wants to Teach Machines Common Sense 
By Cade Metz in the NYT. .... "

From Natural Language to AI

Reminds me of some of our own investigations of building more perceptive AIs and bots.  We still don't have a workable language model of our world.  Limited, focused models do exist,  and are called Ontologies.

The Two Paths from Natural Language Processing to Artificial Intelligence  By Jonathan Mugan
Robot psychologist. Co-founder and CEO at DeepGrammar, a startup specializing in NLP and deep learning. Author of The Curiosity Cycle.

Why isn’t Siri smarter? AI has accelerated in recent years, especially with deep learning, but current chatbots are an embarrassment. Computers still can’t read or converse intelligently. Their deficiency is disappointing because we want to interact with our world using natural language, and we want computers to read all of those documents out there so they can retrieve the best ones, answer our questions, and summarize what is new.

To understand our language, computers need to know our world. They need to be able to answer questions like “Why does it only rain outside?” and “If a book is on a table, and you push the table, what happens?” .... " 

Toward a Layer of Common Sense

Admittedly dated, but been reexamining the 'common sense' ontology Cyc again recently, this article reexamines the premise and progress. Circa 2005. I recall thinking that we might be able to build our own ontologies on top of this one. Give them a layer of badly needed common sense.   Has anyone done that for enterprise applications?  Pointers to experience?

AI and Human Values

Just this week looked back again at the work by Cyc to understand human common sense.    But that is only part of the challenge.   We will need to address values as AI advances.  Can values be readily defined in something like an ontology?

Digital Genies
A.I. researcher Stuart Russell discusses the uncertain work of helping artificial intelligence understand human values.  By Jacob Brogan .... " 

Detecting Funny

The emergence of a New Yorker crowdsourcing system that tries to determine, with interaction and algorithm, what is funny.  In CNet.  Seems this could also be linked to Cyc and perhaps Lucid, to determine the subtlety of our knowledge about our world.    You can join in the crowd sourcing process here 

Lucid and Cyc for Pattern Recognition

Description of Lucid, which combines pattern recognition and learned AI using Cyc,

The Next Step 
Artificial Intelligence Has Been Waiting For
Over 800 AIs can crunch your data. Only one can understand it. ... 

No way shown that I can yet test it.  Can someone give me access?  We tested Cyc in the 90s for potential enterprise use.

See also their blog and more information.

( Historical note.  It came to mind I had heard of Lucid before.  Looked them up and they had existed as a Lisp based AI company we worked with before they morphed into an OOP approach.  Then apparently had connections to early Amazon.    https://en.wikipedia.org/wiki/Lucid_Inc.  )

Finally Leveraging 30 Years of Knowledge in AI?

In the Enterprise we talked to Cyc for potential application in the enterprise, but at the time there were no clear applications.   Cyc is an attempt to record and make useful common knowledge that can be part of AI. Ultimately fundamental to create generally useful virtual assistants.

In Technology Review:
Doug Lenat’s creation is Cyc, a knowledge base of semantic information designed to give computers some understanding of how things work in the real world.

Cyc has been given many thousands of facts, including lots of information that you wouldn’t find in an encyclopedia because it seems self-evident. It knows, for example, that that Sir Isaac Newton is a famous historical figure who is no longer alive. But more important, Cyc also understands that if you let go of an apple it will fall to the ground; that an apple is not bigger than a person; and that a person cannot throw an apple into space.

And now, after years of work, Lenat’s system is being commercialized by a company called Lucid. .... 
" 

Common Sense Reasoning in Cognitive AI

Excellent piece in the ACM.  You need membership to read the full article, but even the abstract introduction makes the point very well.  To intelligently reason about the real world you need to have at hand a large amount of common sense.  And have the ability to reason using that common sense to make your way around a complex world.  What we call 'common sense' is a lot deeper than you think.

That has not changed very much since the last time we rode the AI hype curve.   We have found ways to make better intelligent choices in new and larger and changing contexts. Very useful.  But we have not solved the complete intelligence problem.  See CYC tag for one historic approach to address, but not yet solve this.

MIT ConceptNet5 Stores Knowledge

Via Jim Spohrer, some Notes from the recent AAAI meetings:

" .. I ran into Henry Lieberman (MIT) at AAAI15 conference on Artificial Intelligence/Robotics/CognitiveScience (AI/Robotics/CogSci). Henry suggested that I check out MIT's ConceptNet5

ConceptNet5 contains over 10 millions assertions (facts) about the world.  An example fact is "An apple is a fruit" ... Here are some fun facts about MIT's ConceptNet5: (1) over 10 million assertions (facts)  (2) over 7 million English assertions (facts) (3) 50 relationships (4) nearly 3 million English concepts  If you want to know more, check out my blog post here:
http://service-science.info/archives/3701    

Further .... ConceptNet includes Cyc's upper ontology and some items from DBpedia as well.    ... "  

Some of my response: I wrote more about our investigation of the Cyclopedia work here. 

Also for a while, Cycorp was posting snippets of knowledge on Twitter. Which I thought was a clever idea. But they have not done so since March 2013. More about that effort here, and a link to the tweets to date.  With IBM's collaboration with Twitter this might be an interesting effort to reexamine.

NEIL Learns from Images

I was involved with a startup system that used image recognition and categorization called Photrade several years ago.  There I got an appreciation of how difficult this problem was.   Also how closely it matched to the process of 'learning' we had mimicked using AI methods in the enterprise years before.  What we did was very primitive compared to what is underway at Carnegie Mellon University with The Never Ending Image Learner (NEIL) program.  This also sounds like it should be directly connected to the work at Cyc we also explored.  Will continue to watch.

In  BBC Technology: 

" ... The work is being funded by the US Department of Defense's Office of Naval Research and Google. Since July, the NEIL program has looked at three million images. As a result it has managed to identify 1,500 objects in half a million images and 1,200 scenes in hundreds of thousands of images as well as making 2,500 associations ... The team working on the project hopes that NEIL will learn relationships between different items without being taught. Computer programs can already identify and label objects using computer vision, which models what humans can see using hardware and software, but the researchers hope that NEIL can bring extra analysis to the data. ... " 

Cyc, the Enterprise and Wikipedia

Recent research had me take a look back on the Cyc project, which we had examined during our AI explorations.   We talked to Lenat's Cycorp. The system promised to encode common sense knowledge relationships and make it available for natural language search and reasoning. It is a semantic analysis of knowledge.

Although it looked very promising, we were unable at the time to integrate it into our rule based reasoning effort.  If our expert systems had seen broader use,  a next step use would have been to include corporate enterprise knowledge using Cyc.

By overlaying the Cyc structure on our internal enterprise Wikipedia?  What might this have looked like?    An examination of the Wikipedia page in Cyc. I  found the following reference  " ... Cyclopedia is being developed; it superimposes Cyc keywords on pages taken from Wikipedia pages. ... "   A paper referenced talks about the project.  A followup paper.  And a sample page is shown.   Nothing about the project seems to exist after 2012.

Not sure of the current state of this work.  Does anyone know?  References would be appreciated.

CYC Corp

Been a long time since I looked there, or visited them, Doug Lenat's CYC group.  I was reminded by a mention in a book ...  worth a reexamination. Early pioneer in AI that we followed.