Cognitive Computing : basic concepts and timeline

The uncontrollable human desire to solve problems is leading us to the gates of cognitive computing. In this future closer and closer, men and machines will work together to solve the most complex problems of the world.

The ability to analyze and understand the machine data will help us solve the most threatening epidemics, developing sophisticated treatments for complex diseases and find solutions to water and pollution crisis, just to name a few examples. Let’s use the machines to investigate crimes and organize traffic.

Cognitive computing systems learn and interact naturally with people (through natural language) to extend (“augment”) which humans or machines could do alone. One of the ultimate goals is to leverage powerful blends of AI algorithms in order to successfully assist human experts to make better decisions, transposing the complexity of large volumes of data.

Watson is a great example of a breakthrough cognitive system and a truly ´birth star´ of a new era in computing, represented by such disruption.

This new era of computing is already changing expectations for how technology can help live and work in better ways. Cognitive systems can properly addresses explosive data growth, rapidly changing business conditions, and the need for more intelligent, intuitive interaction between people and technology.

In this context, Watson kind of represents a new class computational solution that leverages deep content analysis and evidence-based reasoning to accelerate and improve decisions and optimize outcomes. It does it based on a set of transformational technologies which leverage NLP, hypothesis generation, and inferential, evidence-based learning. It essentialy combines these technologies and applies massive parallel probabilistic processing techniques to change the way we solve problems.

Cognitive systems are enabling the creation of a new partnership between people and computers: we bring context, problems, values; and the system brings its ability to find patterns in massive volumes of data, learning by inference.

One of the main AI goals is to enable creation of autonomous intelligent systems.

Cognitive computing aims at creating auxiliary systems that augment and expand our cognitions.

The cognitive computing era can be seen as the third cognitive disruptive paradigm in the evolution of computing — after tabulating machines, and programmable computers.

• Tabulating machines came to prominence in the late 19th century and allowed advances such as the national census and the Social Security System in the US.
• Programmable computing, emerged in the 1940s and enabled everything from space exploration to the Internet.

Cognitive systems are different. Because the system "learns" from natural interactions with data and individuals, they’re continuously improving themselves, getting more valuable with time.

One of the key implications of the data firehose where we currently live in is that we need computing systems that we can interact with using human language (rather than programming language), and badly need computers that can dish up advice (rather than wait for commands). So much that medical doctors in health institutions, sales executives in retail, investment advisors, educators and a myriad of professionals will have technology at their fingers (installed on smartphones, tablets, or desktops) that can derive and grasp insights from the pile of information being gathered to help us learn about the world we live in, making sense of it and obtaining advice on how to navigate it.

As J.Kelly puts it, “this is the most significant paradigm shift in the history of computing”.

The so-called history of Cognitive Computing actually got its start roughly sixty years ago, by the time when Alan Turing published its now famous Test (back then, a validation of intelligent behavior by a computing machine) and Claude Shannon published a detailed analysis of chess playing as search. AI was just getting started — MIT´s John McCarthy (the inventor of LISP) was the first to use the name “artificial intelligence” during the second Dartmouth Conference.

During the 50s the first “smart” programs were developed, such as Samuel´s checkers player and Newell/Shaw/Simon´s “logic theorist”. McCarthy, together with Marvin Minsky, later came to found the MIT AI lab. It was around this time that fascination with machine translation began, when Masterman and colleagues at the University of Cambridge designed basic semantic nets for machine translation (R.Quillian later effectively demonstrated semantic nets in 1966 at the Carnegie Institute).

The early 60s saw Ray Solomonoff lay the foundations of a mathematical theory of AI; he introduced the universal Bayesian methods for inductive inference and prediction. By 1963, Feigenbaum and Feldman were publishing the first set of AI articles , in the now famous “Computers and Thought” journal. A student named Danny Bobrow at MIT showed that computers could be programmed to understand natural language to grasp natural language well enough to solve algebra word problems correctly. Before the advent of the first generation of database systems capable tracking a substantial amount of (structured) data to support the NASA missions to the Moon (project Apolo), J.Weizenbaum put together the famous Eliza, an interactive computer program that could carry on a dialogue (in English language) on any topic. Worth noting that around 1966, a series of negative reports on machine translation would do much to kill work in NLP for several years, due to the complexity and lack of maturity in the techniques available at the time. Before the 60s were over, J.Moses demonstrated the power of symbolic reasoning and the first good chess playing programs (such as MacHack) came to being. It was also in 1968 that a program called Snob (created by Wallace and Boulton) effected clustering work (unsupervised classification, in machine learning parlance) using Bayesian minimum message length criterion (the first mathematical realization of Occam razor principle!)

The 70s started with Minsky and Papert “Perceptrons” , recognizing the limits of basic neural network structures. However, at around same time, Robinson and Walker also established an influential NLP group at SRI think thank, sparking significant (productive) research in this arena. Colmerauer developed Prolog in 1972 and only two years after T.Shortliffe´s PhD dissertation on the MYCIN program demonstrated a practical rule-based approach to medical diagnoses (even under uncertainty), influencing expert systems development for years to come.

In 1975, Minsky published his widely read and influential article on Frames, as a representation of knowledge — where several ideas about schemas and semantic links were brought together.

The 80s saw the first machine learning algorithms for language processing be introduced. UCLA´s Judea Pearl is credited for bringing probability and decision theory into AI by this time. Temporal events were formalized through Interval Calculus, pioneered by J. Allen in 1983. By the mid-80s neural enteworks became widely used with backpropagation algorithms.

Early evidences of successful reinforcement learning (very important in contemporary machine learning systems) date back to the early 90s — Gary Tesauro´s backgammon computer program demonstrated that reinforcement learning was powerful enough to enable creation of competitive gammon software. In general, the 90s witnessed relevant advances in all areas of AI, with significant demonstrations in machine learning, intelligent tutoring, case-based reasoning, multi-agent planning, scheduling, uncertain reasoning, data mining, natural language understanding and translation, and to some degree machine vision. One highlight was semantic classification and probabilistic parsing being combined to enable derivation of rules and their probabilities. Also of note is advent of TAKMI, a tool to capture and use knowledge embedded in text files (applied to contact-centers).

The 2000´s are the moment in time when QA (question/answer) systems came to prominence. Watson itself is the result of sophisticated work conducted by IBM to further develop and enhance Piquant (Practical Intelligent Question Answering Technology) adapting for the Jeopardy challenge.

After the mid-2000 we’ve been watching substantial progress in a broad set of fields of AI, taking the advances from the 90s way further and giving birth to disruptive innovations — things such as the first program to effectively solve checkers (2007), the self-driving car (2009), computer generated articles (2010), smartphone-based assistants (2010/11) and a plethora of advances in robotics.

In 2011 Watson famously defeated the two best human contenders of Jeopardy! Game tv show (Jennings and Rutter), kickstarting the Cognitive Computing era. Over recent years we are also witnessing significant developments in deep learning techniques, which grant computing systems with ability to recognize images and video with high degree of precision.

short history of cognitive computing (image credit: Deloitte)