Cognitive Software Group — Loose lips sink AI ships
In February, a couple of weeks ago, Dr. Roger Schank castigated IBM. Accusing IBM of “fraud”, Dr Schank asserted “they are not doing “cognitive computing” no matter how many times they say they are”.
Dr Schank has been CEO of Socratic Arts since 2002, is a prolific publisher of articles and books, and formerly holding positions including Professor of Computer Science and Psychology at Yale University and Director of the Yale Artificial Intelligence Project (1974–1989), a visiting professor at the University of Paris VII, Assistant Professor of Linguistics and Computer Science at Stanford University (1968–1973), and research fellow at the Institute for Semantics and Cognition in Switzerland.
What attracted Dr Schank’s ire was a proclamation from an IBM Vice President of Marketing, Ann Rubin, that IBM’s Watson AI platform could “outthink” human brains in areas where finding insights and connections can be difficult due to the abundance of data.
“You can outthink cancer, outthink risk, outthink doubt, outthink competitors if you embrace this idea of cognitive computing,” she apparently said.
In 2018, IBM’s Australian advertising agency ran a campaign called “Outthink Melanoma” that appeared to be based on using Machine Learning pattern recognition techniques to identify melanoma, rather than AI or “cognitive” anything.
Clearly attempting to out-Musk the wildly optimistic predictions of wires embedded in a pig’s brain and interacting with a “connected” microchip being the cusp of the great “Singularity”, Ms Rubin may soon be tempted to predict a colony of one thousand baby Watson’s colonising Mars by Christmas 2026. Nevertheless, she runs IBM’s corporate advertising program and can be forgiven for doing what Marketing people are supposed to do.
Computers that “think”??? Oh well, OK then.
Soon after the COVID-19 outbreak, we approached a University epidemiologist to offer help with our Artificial Intelligence know-how and our “cognitiveAI”, platform. After a couple of chats with an eminent Professor and epidemiologist, the challenge we set ourselves was to build a system that could read, understand and remember the content of 70,000 coronavirus research papers stored and distributed by the “Semantic Scholar” medical research database and search engine.
We had learned that researchers look to prior research for clues when pursuing some new hypothesis. Clearly no human researcher can read 70,000 research papers and remember everything in them; just reading them would take a year reading 192 papers every day. Nor could a human remember the content of 70,000 research papers and then ‘join the dots’ between all that content to find clues to propose or support a new hypothesis.
Over the next five months we allocated around 100 person days to see what we could do with an artificial intelligence system, built from scratch.
Keyword search and semantic search provided by common search engines are not enough for this task. We targeted cognitive search and storage, where the computer compared what it read with knowledge it already had. We needed it to compare each research paper and rank it based on similarity to others that had been written about the same topics. If the researcher is trying to find the 25 papers that most closely examine the transmission characteristics of mutations in the spike protein of the COVID-19 coronavirus, that is what we need to provide. A human cannot find the 25 most relevant papers in 70,000; it’s simply beyond our capability.
For the nerds amongst us, a quick overview of the technical stuff….
“Betsy”, as we call our coronavirus AI system, was provided with 3000 coronavirus research papers that we downloaded from Semantic Scholar. Betsy reads each sentence in each paper and uses Natural Language Processing, Neural Networks, Semantic Computing and proprietary algorithms to extract “relevant clauses and words”, and the relationships between them, what we call “facts”. Epidemiologists annotate some of the extracted clauses and words with a concept to assist with training the neural network to associate phrases and words with concepts, adding even more facts. The facts are represented in “RDF triples”, in subject-predicate-object form.
A large number of RDF triples can therefore be generated from each research paper that highlight key concepts and related phrases/words being discussed in the paper. The facts (RDF triples) are stored in a “Semantic Graph” to provide a very rich description of the content of each research paper.
A putative ontology, also stored in the Semantic Graph, can be semi-automatically generated from the triples to “join the dots” within and between papers. This process can be performed on 700,000 research papers as easily as it can be performed on 700 and provides a rich description of each research paper and relationships between them encoded for a computer to understand, by itself.
The putative ontology can be used to identify external ontologies that match or partly match the putative ontology and these can be integrated to add additional concepts to the semantic graph.
Semantic search techniques such as word vectors are obviously inferior to the above richly informed search and store technique that is much more likely to match documents and is therefore a significant leap forward for concept-rich research such as medical research.
Graph databases, despite their name, are nothing like the graphs we are familiar with. Rather, they are multi-dimensional data structures that allow for substantially more sophisticated descriptions of relationships between items recorded in the data structure. Even better and unlike traditional databases, they are coded according to a data model that a computer understands, by itself.
A Semantic Graph provides the richest descriptions of relationships, making it particularly well-suited to the high demands of artificial intelligence, and it models the facts extracted from each research paper and find relationships with facts in other research papers.
When searching a Semantic Graph, ranking can be applied to relationships to highlight the papers of highest relevance. Indeed, when Betsy if extracting phrases and words from a research paper to record new facts, she interacts with the Semantic Graph to consistently improve her extractions.
In this way, a Semantic Graph possibly mimics the semantic memory occurring in the human brain’s neocortex, where neuroscientists suggest that semantic memory is “a type of long-term memory involving the capacity to recall words, concepts, or numbers, which is essential for the use and understanding of language”.
SPARQL queries can quickly find related papers to correspond with a researcher’s domain of interest and be represented graphically using Visualizers. Betsy also supports queries expressed in natural language for query and answer of the Semantic Graph, maybe mimicking a human at a very basic conversational level, but with a massive amount of well-ordered data at Betsy’s immediate disposal.
In other similarities between computers and human brains
1. a bus is a communication system that transfers data between components inside a computer, and covers all related hardware and software components. In the human brain and nervous system, “data” is moved around the brain and the body by 100 billion specialized cells called “neurons”. The interaction of the nervous system, the brain neurons, and the Hippocampus is thought to be of utmost importance in memory formation.
2. knowledge is encoded in human brain memory cells by a combination of cell biology, biochemistry and electrical pulses, but we don’t know how. Obviously that encoding is independent of spoken language so that you learn in one language and the learning is transportable to another language for multilingual people, but we don’t know what the mechanism is for that in human brains. Interestingly, “neuroplastic brain changes”, including increased grey matter density, have been found in people with skills in more than one language, from children and young adults through to the elderly. In a computer’s semantic graph, knowledge is encoded in the RDF-triple data model and ontologies that is also independent of the data format that it may be acquired from or published in. In both cases some kind of data transformation occurs.
3. Computer reasoning can be applied to the triples to infer relationships between concepts and facts in different research papers. Here the computer activity of the processors and RAM possibly mimics the working memory that “occurs in the prefrontal cortex of the brain and is a type of short-term memory that facilitates planning, comprehension, reasoning, and problem-solving. The prefrontal cortex is the most recent addition to the mammalian brain and has often been connected or related to intelligence and learning of humans”. Human experts can then consider, accept or ignore the computer-generated inferences in a human-machine partnership of discovery.
Enough now nerds!
In plain non-technical language, what is indisputable is that a computer can, at thousands of times the speed of a human
- read a digital document or a database at astonishing speed compared to a human
- extract relevant phrases, words, and concepts (with some human training)
- remember everything it reads and extracts
- using semantic computing techniques, establish and record relationships in the extracted information
- using computer reasoning techniques, infer relationships for humans to consider.
In the above, the computer out-paces any human researcher; a human simply cannot compete with the speed of a computer for some important tasks.
But “thinking” Ms Rubin? A computer can out-think a human? Pigs might fly to Mars as well. Maybe if you are in IBM’s marketing department, but not if you are a human neurologist, or any other type of neuroscience expert, or Dr Schank.
While intelligence and thinking are linked to the poorly understood power of human reasoning, computers are currently limited to Semantic Graphs (that link “facts” to represent “knowledge”) and logical reasoning. Both are tightly bound by mathematics, are very useful, but also currently limited. In contrast, Humans have no such limitations. Adult humans have very much superior reasoning based on experience, intuition, imagination, and emotion, for example. (Try telling that to a teenager.)
Just two years ago many “experts” referred to “Machine Learning” and “AI” interchangeably, implying that Machine Learning enabled computer intelligence. Today, we distinguish between the two.
The not-so-good news is that
- Machine Learning techniques such as Neural Networks are clever rather than intelligent, limited to pattern recognition and largely dependent on the human intelligence involved in making and supervising them
- Artificial Intelligence techniques such as Semantic Graphs and computer reasoning are still at a very basic stage of mimicking human intelligence.
The good news is that
- these resource hungry techniques are better supported by faster infrastructure solutions, such as solid-state memory and cloud computing
- The rate of understanding Semantic Graphs and computer reasoning is now increasing quickly, and getting much greater attention now than it has in the last twenty years. As a result, the potential of expanding the functionality of AI techniques such as Semantic Graphs and RDF is growing rapidly.
Artificial Intelligence is not a technique; it is a collection of techniques and as the functionality of each one is extended it will complement and extend the others.
So, is Artificial Intelligence equivalent to or close to Human intelligence? Definitely not; superficial resemblance of something is not the same as equaling it.
Is the combination of Semantic Graphs and computer reasoning equivalent to or close to human thinking? Definitely not.
Can computers outperform humans in finding hidden “insights and connections where there is an abundance of data”? Absolutely, astonishingly so.
Is Artificial Intelligence capable of substantial support to human thinking? Most definitely.
Is Watson capable of “cognitive computing”, implying that they are thoughtful? It’s really a matter of semantics!
Dr Schank’s article can be found here.
Cognitive Computing explained here.
Author: Mark Bradley is the founder and sales chief of Cognitive Software Group, the leading cognitive computing company in Australia. www.cognitivesoftware.com