Artificial Intelligence: Job Killer

Michael Feldman

---

A recent article in the New York Times points out that sophisticated data analytics software is doing the kinds of jobs once reserved for highly paid specialists. Specifically, it talks about data mining-type software applied to document discovery for lawsuits. In this realm, these applications are taking the place of expensive teams of lawyers and paralegals.
Basically it works by performing deep analysis of text to find documents pertinent to the case at hand. It's not just a dumb keyword search; the software is smart enough to find relevant text even in the absence of specific terms. One application was able to analyze 1.5 million documents for less than $100,000 -- a fraction of the cost of a legal team, and performed in a fraction of the time.
Mike Lynch, founder of Autonomy (a UK-based e-discovery company), thinks this will lead to a shrinking legal workforce in the years ahead. From the article:

He estimated that the shift from manual document discovery to e-discovery would lead to a manpower reduction in which one lawyer would suffice for work that once required 500 and that the newest generation of software, which can detect duplicates and find clusters of important documents on a particular topic, could cut the head count by another 50 percent.

Such software can also be used to connect chains of events mined from a variety of sources: e-mail, instant messages, telephone calls, and so on. Used in this manner, it can be used to sift out digital anomalies to track various types of criminal behavior. Criminals, of course, are one workforce we'd like to reduce.  But what about the detectives that used to perform this kind of work?
The broader point the NYT article illuminates is that software like this actually targets mid-level white collar jobs, rather than low-end labor jobs we usually think of as threatened by computer automation. According to David Autor, an economics professor at MIT, this is leading to a "hollowing out" of the US economy. While he doesn't think technology like this is driving unemployment per se, he believes the job mix will inevitably change, and not necessarily for the better.
It's the post-Watson era. Get used to it.

The past, present and future of cancer

Leading cancer researchers reflected on past achievements and prospects for the future of cancer treatment during a special MIT symposium on Wednesday titled “Conquering Cancer through the Convergence of Science and Engineering.”

The event, one of six academic symposia taking place as part of MIT’s 150th anniversary, focused on the Institute’s role in studying the disease over the past 36 years since the founding of MIT’s Center for Cancer Research.

During that time, MIT scientists have made critical discoveries that resulted in new cancer drugs such as Gleevec and Herceptin. The center has since become the David H. Koch Institute for Integrative Cancer Research, which now includes a mix of biologists, who are trying to unravel what goes wrong inside cancer cells, and engineers, who are working on turning basic science discoveries into real-world treatments and diagnostics for cancer patients.

That “convergence” of life sciences and engineering is key to making progress in the fight against cancer, said Institute Professor Phillip Sharp, a member of the Koch Institute. “We need that convergence because we are facing a major demographic challenge in cancer as well as a number of other chronic diseases” that typically affect older people, such as Alzheimer’s, Sharp said.

In opening the symposium, MIT President Susan Hockfield said that MIT has “the right team, in the right place, at the right moment in history” to help defeat cancer.

“It’s in the DNA of MIT to solve problems,” said Tyler Jacks, director of the Koch Institute. “I’m very optimistic and very encouraged about what this generation of cancer researchers at MIT will do to overcome this most challenging problem.”

Past and present

In the past few decades, a great deal of progress has been made in understanding cancer, said Nancy Hopkins, the Amgen, Inc. Professor of Biology and Koch Institute member, who spoke as part of the first panel discussion, on major milestones in cancer research.

In the early 1970s, before President Richard Nixon declared the “War on Cancer,” “we really knew nothing about human cells and what controls their division,” Hopkins recalled. Critical discoveries by molecular biologists, including MIT’s Robert Weinberg, revealed that cancer is usually caused by genetic mutations within cells.

The discovery of those potentially cancerous genes, including HER2 (often mutated in breast cancer), has lead to the development of new drugs that cause fewer side effects in healthy cells. While that is a major success story, many other significant discoveries have failed to make an impact in patient treatment, Hopkins said.

“The discoveries we have made are not being exploited as effectively as they could be,” Hopkins said. “That’s where we need the engineers. They’re problem-solvers.”

Institute Professor Robert Langer described his experiences as one of the rare engineers to pursue a career in biomedical research during the 1970s. After he finished his doctoral degree in chemical engineering in 1974, “I got four job offers from Exxon alone,” plus offers from several other oil companies. But Langer had decided he wanted to do something that would more directly help people, and ended up getting a postdoctoral position in the lab of Judah Folkman, the scientist who pioneered the idea of killing tumors by cutting off their blood supplies.

In Folkman’s lab, Langer started working on drug-delivering particles made from polymers, which are now widely used to deliver drugs in a controlled fashion.

Langer and other engineers in the Koch Institute are now working on ways to create even better drug-delivery particles. Sangeeta Bhatia, the Wilson Professor of Health Sciences and Technology and Electrical Engineering and Computer Science, described an ongoing project in her lab to create iron oxide nanoparticles that can be tagged with small protein fragments that bind specifically to tumor cells. Such particles could help overcome one major drawback to most chemotherapy: Only about 1 percent of the drug administered reaches the tumor.

“If we could simply take these poisonous drugs more directly to the tumors, it would increase their effectiveness and decrease side effects,” Bhatia said.

Other Koch engineers are working on new imaging agents, tiny implantable sensors, cancer vaccines and computational modeling of cancer cells, among other projects.

Personalized medicine

Many of the targeted drugs now in use came about through serendipitous discoveries, said Daniel Haber, director of the Massachusetts General Hospital Cancer Center, during a panel on personalized cancer care. Now, he said, a more systematic approach is needed. He described a new effort underway at MGH to test potential drugs on 1,000 different tumor cell lines, to find out which tumor types respond best to each drug.

At MIT, Koch Institute members Michael Hemann and Michael Yaffe have shown that patient response to cancer drugs that damage DNA can be predicted by testing for the status of two genes — p53, a tumor suppressor, and ATM, a gene that helps regulate p53.

Their research suggests that such drugs should be used only in patients whose tumors have mutations in both genes or neither gene — a finding that underscores the importance of understanding the genetic makeup of patients’ tumors before beginning treatment. It also suggests that current drugs could be made much more effective by combining them in the right ways.

“The therapies of the future may not be new therapies,” Hemann said. “They may be existing therapies used significantly better.”

The sequencing of the human genome should also help achieve the goal of personalized cancer treatment, said Eric Lander, director of the Broad Institute and co-chair of the President’s Council of Advisors on Science and Technology, who spoke during a panel on biology, technology and medical applications. Already, the sequencing of the human genome has allowed researchers to discover far more cancer-causing genes. In 2000, before the sequence was completed, scientists knew of about 80 genes that could cause solid tumors, but by 2010, 240 were known.

Building on the human genome project, the National Cancer Institute has launched the Cancer Genome Atlas Project, which is sequencing the genomes of thousands of human tumors, comparing them to each other and to non-cancerous genomes. “By looking at many tumors at one time, you can begin to pick out common patterns,” Lander said.

He envisions that once cancer scientists have a more complete understanding of which genes can cause cancer, and the functions of those genes, patient treatment will become much more effective. “Doctors of the future will be able to pick out drugs based on that information,” he said.

Is government ready for the semantic Web?

So far it's been slow going, but an interagency XML project could boost law enforcement, health care efforts

• By John Moore
• Mar 18, 2011

When IBM’s Watson recently trounced the two most successful "Jeopardy!" players of all time, the supercomputer was relying in part on an emerging field of computerized language processing known as semantic technology.
In addition to being able to work out the answer to questions for Watson such as what fruit trees provide flavor to Sakura cheese, semantic technology is capable of providing answers to questions that might interest government agencies and other groups that historically have had problems identifying patterns or probable sequences in oceans of data. 
The idea is to help machines understand the context of a piece of information and how it relates to other bits of content. As such, it has the potential to improve search engines and enable computer systems to more readily exchange data in ways that could be useful to agencies involved in a wide range of pursuits, including homeland security and health care.
While semantic technology has mostly been an academic exercise in recent years, it is now finding a greater role in a practical-minded government project called the National Information Exchange Model (NIEM).

NIEM pursues intergovernment information exchange standards with the goal of helping agencies more readily circulate suspicious activity reports or issue Amber Alerts, for example. The goal is to create bridges, or exchanges, between otherwise isolated applications and data stores.
The building of those exchanges calls for a common understanding of the data changing hands. The richer detail of semantic descriptions makes for more precise matches when systems seek to consume data from other systems. Agreement on semantics also promotes reuse; common definitions let agencies recycle exchanges.
Semantics in government IT
Today, NIEM offers a degree of semantic support. But some observers believe the interoperability effort will take a deeper dive into semantic technology. They view NIEM as a vehicle that could potentially make semantics a mainstream component of government IT.
“Semantically, there is a huge opportunity with NIEM,” said Peter Doolan, vice president and chief technology officer at Oracle Public Sector, which is working on tools for NIEM. “NIEM is a forcing function for the broader adoption of the deeper semantic technology that we have talked about for some time.”

As more agencies adopt NIEM, the impetus for incorporating semantics will grow. NIEM launched in 2005 with the Justice and Homeland Security departments as the principal backers. Last year, the Health and Human Services Department joined Justice and DHS as co-partners. State and local governments, particularly in law enforcement, have taken to NIEM as well. And in a move that underscores that trend, the National Association of State Chief Information Officers last month joined the NIEM executive steering committee.
“NIEM adoption is going at a furious pace,” said Mark Soley, chairman and CEO of the Object Management Group (OMG), which has been working with NIEM. “As it gets adoption, they are going to need a way to translate information that is currently in other formats. That is when you need semantic descriptions.”
NIEM’s leadership says the program is prepared for greater use of semantics. “The NIEM program stands ready to respond to the overall NIEM community regarding a broader adoption of semantic technologies,” said DHS officials who responded to questions via e-mail.
Support for semantics
NIEM is based on XML. The project grew out of the Global Justice XML Data Model (GJXDM), a guide for information exchange in the justice and public safety sectors. Although XML serves as a foundational technology for data interoperability, it is not necessarily viewed as semantic.
However, John Wandelt, principal research scientist at the Georgia Tech Research Institute (GTRI) and division chief of that organization’s Information Exchange and Architecture Division, said semantic capability has been part of NIEM since its inception. GTRI serves as the technical architect and lead developer for GJXDM and NIEM.
“From the very early days, the community has pushed for strong semantics,” he said. Wandelt pointed to XML schema, which describes the data to be shared in an exchange. “Some say schema doesn’t carry semantics,” he said. "But the way we do XML schema in NIEM, it does carry semantics.”
NIEM’s Naming and Design Rules help programmers layer an “incremental set of semantics on top of base XML,” Wandelt said. For example, a group of XML programmers tasked to build a data model of their family trees would depict relationships between parents, siblings, and grandparents. But those ties would be implied and based entirely on an individual programmer’s way of modeling.
NIEM’s design rules, on the other hand, provide a consistent set of instructions for describing connections among entities. Wandelt said those roles make relationships explicit, thereby boosting semantic understanding.
NIEM also uses Resource Description Framework (RDF), an important underpinning of the Semantic Web, which has been slowly making its way into government IT (see sidebar).
RDF aims to describe data in a way the helps machines better understand relationships.

see the full article here: http://gcn.com/articles/2011/03/21/niem-and-semantic-web.aspx

Collective Intelligence Outsmarts Artificial Intelligence

When computers first started  to infringe on everyday life, science fiction authors and society in general had high expectations for "intelligent" systems. Isaac Asimov's "I, Robot" series from the 1940s portrayed robots with completely human intelligence and personality, and, in the 1968 movie "2001: A Space Odyssey," the onboard computer HAL (Heuristically programmed ALgorithmic computer) had a sufficiently human personality to suffer a paranoid break and attempt to murder the crew!
While the computer revolution has generally outstripped almost all expectations for the role of computers in society, in the area of artificial intelligence (AI), the predictions have, in fact, outstripped our achievements. Attempts to build truly intelligent systems have been generally disappointing.
Fully replicating human intelligence would require a comprehensive theory of consciousness which we unfortunately lack. Therefore, AI has generally attempted to focus on simulating intelligent behavior, rather than intelligence itself. In the algorithmic approach, programmers labor to construct sophisticated programs that emulate a specific intelligent behavior, such as voice recognition. In the other traditional approach - expert systems - a database of facts is collected, and logical routines applied to perform analysis and deduction. Expert systems have had some success in medical and other diagnostic applications, such as systems performance management.
Each of these approaches has shown success in limited scenarios, but neither achieves the sort of broadly intelligent system promised in the early days of computing. Attempts to emulate more human-like cognitive or learning systems-using technologies such as the neural nets, fuzzy logic, and genetic algorithms-have only slightly improved the intelligence of everyday software applications.
Most of us experience the limitations of artificial intelligence every day. Spell-checkers in applications such as Microsoft Word do an amazingly poor job of applying context to language correction. As a result, sentences such as, "Eye have a spelling checker, it came with my pea sea," pass through the Microsoft spelling and grammar checker without a hitch. While the Microsoft software can recognize spelling mistakes in individual words, it cannot understand the meaning of the sentence as a whole, and the result is a long way from intelligent judgment. 
Collective intelligence offers a powerful alternative to traditional artificial intelligence paradigms. Collective intelligence leverages the inputs of large numbers of individuals to create solutions that traditional approaches cannot achieve. Although the term "collective intelligence" is not widely recognized, most of us experience the results of collective intelligence every day. For instance, Google uses collective intelligence when auto-correcting search inputs. Google has a large enough database of search terms to be able to automatically detect when you make an error and correct that error on-the-fly.  Consequently, Google is more than able to determine that "pea sea" is almost certainly meant to be "PC."
Collective intelligence not only allows for superior spelling and grammar correction, but also is used in an increasingly wide variety of contexts, including spam detection, diagnostic systems, retail recommendations, predictive analytics, and many other fields. Increasingly, organizations find that it is more effective to apply brute force algorithms to masses of data generated by thousands of users, than to attempt to explicitly create sophisticated algorithmic models. 
The ability of collective intelligence to solve otherwise intractable business and scientific problems is one of the driving forces behind the "big data" evolution. Organizations are increasingly realizing that the key to better decision making is not better programs but granular crowd-sourced data sets.
Collective intelligence is merely one of the techniques used to endow computer systems with more apparent intelligence and to better solve real world problems - it's not in any way a replacement for the human brain. However, in an increasingly wide range of applications, collective intelligence is clearly outsmarting traditional artificial intelligence approaches.

Artificial intelligence has just got smarter

Rajeev Srinivasan

The American TV quiz showJeopardy! has been running for over 25 years. Contestants are given clues in categories ranging from serious subjects such as World War II, to more frivolous topics like rock musicians. They then have to come up with a question in the format: “Who is…”, or “what is…” based on the clues. The clues are not straightforward and factual — a computer with a large database can crack such statements quickly — but oblique. They are full of puns, obscure relationships, jokes, allusions and so on that only a human being steeped in that culture will recognise. In that sense, the clues are not ‘context-free’ as computer languages are (or for that matter, classical Paninian Sanskrit): you must know quite a bit of cultural context to decode them. This is infernally hard for computers, and a challenge that artificial intelligence (AI) researchers have been struggling with for decades — the holy grail of ‘natural language processing’. There have been several false starts in AI, and enthusiasm has waxed and waned, but the iconic promise of computers that can converse (such as the talking computer HAL in 2001: A Space Odyssey) has remained elusive. This is why it is exciting news that a new IBM program (dubbed ‘Watson’ after the founder of the company), built specifically to play Jeopardy, defeated two of the world’s best human players in a special edition of the show on February 16th. There was some quiet satisfaction among the techie crowd that the day may yet arrive when intelligent robots can respond to conversational queries. Watson runs on a cluster of ninety Linux-based IBM servers, and has the horsepower to process 500 gigabytes of data (the equivalent of a million books) per second — which is necessary to arrive at an answer in no more than 3 seconds; that is the time human champions need to press the buzzer that would give them the right to answer the question. Ray Kurzweil, an AI pioneer and futurist, suggests this level of computing power will be available in a desktop PC in about a decade. Watson’s accomplishments are qualitatively different from those of its predecessor, Deep Blue, which defeated world chess champion Garry Kasparov in 1977. In many ways, chess, with its precise rules, is much easier for computers than the loose and unstructured Jeopardy! game. Thus, Watson is much more complex than Deep Blue, which stored the standard chess openings, and did a brute-force analysis of every possible outcome a few moves into the future. The interesting question though, is, what does all this mean for humans? The nightmare possibility is that we have reached that tipping point where humans will become redundant. That of course was the precise problem that 2001: A Space Odyssey’s HAL had - it felt the humans on board its spaceship were likely to cause the mission to fail; therefore it methodically set about eliminating them. Much the same dystopic vision haunts us in other science-fiction films: for instance the omniscient Skynet in The Terminator series or the maya-sustaining machines in The Matrix. Berkeley philosopher John Searle, writing in the Wall Street Journal, gives us some comfort. According to him, Watson is merely a symbol-manipulating engine, and it does not have superior intelligence; nor is it ‘thinking’. It merely crunches symbols, i.e. syntax, with no concept of meaning, i.e. semantics. “Symbols are not meanings,” he concludes, “Watson did not understand the questions, or its answers… nor that it won — because it doesn’t understand anything.” Even without becoming our overlords, Watson and its descendents may cause displacement. They will cause a number of jobs to disappear, just as voice recognition is affecting the transcription industry. Former head-fund manager Andy Kessler suggests in the WSJ that there are several types of workers, but basically ‘creators’ and ‘servers’; only the former are safe. Technology such as Watson will, he says, not only disrupt retail workers (eg. travel agents), bureaucrats, stockbrokers and customer support staff, but also legal and medical professionals. The latter may find applications like a doctor’s or lawyer’s assistant increasingly cutting into their job content. Thus the arrival of Watson-like artificial intelligences may cause serious disruption in the workforce, although it is not likely that they will be ordering us around any day soon. At least not yet. Humanity may be more resilient than we thought.

Robots Achieve Self-Awareness, May Also Develop ‘Mental Problems’

Artificial intelligence has taken a big leap forward: two roboticists (Lipson and Zagal), working at the University of Chile, Santiago, have created what they claim is the first robot to possess “metacognition” — a form of self-awareness which involves the ability to observe ones’ own thought processes and thus alter one’s behavior accordingly.
The starfish-like robot (which has but four legs) accomplished this mind-like feat by first possessing two brains, similar to how humans possess two brain hemispheres (left and right*). This provided the key to the automaton’s adaptability within a dynamic, and unpredictable, environment.
The double bot brain was engineered such that one ‘controller’ (i.e., one brain) was “rewarded” for pursuing blue dots of  light moving in random circular patterns, and avoiding running into moving red dots. The second brain, meanwhile, modeled how well the first brain did in achieving its goal.
But then, to determine if the bot had adaptive self-awareness, the researchers reversed the rules (red dots pursued, blue dots avoided) of the first brain’s mission. The second brain was able to adapt to this change by filtering sensory data to make red dots seem blue and blue dots seem red; the robot, in effect, reflected on its own “thoughts” about the world and modified its behavior (in the second brain), fairly rapidly, to reflect the new reality.
This achievement represents a significant advancement over earlier successes with AI machines in which a robot was able to model its own body plan and movements in its computer brain, make “guesses” as to which of its randomly selected body-plan models was responsible for the correct behavior (movement), and then eliminate all the unsuccessful models, thus exhibiting an “analogue” form of natural selection (see Bongard, Zykov, Lipson, 2006). **

TOPIO, a humanoid robot, played ping pong at Tokyo International Robot Exhibition (IREX) 2009.

The team is already moving beyond this apparent meta-cognition stage and is attempting to enabled a robot to develop what’s known as a ‘theory of mind’ – the ability to “know” and predict what another person (or robot) is thinking. In an early experiment, the team had one robot observe another robot moving in a semi-erratic manner (in a spiral pattern) in the direction of a light source. After a short while, the observer bot was able to predict the other’s movement so well that it was able to “lay a trap” for it.
Lipson believes this to be a form of “mind reading”. However, a critic might argue that this is more movement reading, than mind, and that it remains to be proven that the observer bot has any understanding of the other’s “mind”. A behavior (such as the second bot trapping the first) might simulate some form of awareness of another’s thought process, but can we say for sure that this is what is really happening?
One idea that might lend credence to this claim is if the observer bot had a language capacity that allowed it to express its awareness, or ‘theory of mind’. Nearly two decades ago, pioneering cognitive biologists Maturana and Varela posited “Language is the sin qua non of that experience called mind.”
And, achieving such a “languaging” capacity in not out of the question; a few years ago, a team of European roboticists created a community of robots that not only learned language, but soon learned to invent new words and to share these new words with the other robots in the community (see: Luc Steels, of the University of Brussels/SONY Computer Science Laboratory in Paris).
It is conceivable that a similarly equipped robot — also possessing the two-brain structure of Lipson’s robots — could observe itself thinking about thinking, and express this awareness through its own (meta) language. Hopefully, we will be able to understand what it is trying to express when and if it does.

A Pick and Place robot in a factory. You've come a long way droidy.

In a recent SciAm article on this topic, Lipson stated:
“Our holy grail is to give machines the same kind of self-awareness capabilities that humans have”
One other question that remains, then: Will the robot develop a more complex simulation/awareness of itself, and the world, as it learns and interacts with the world, as we do?
The four-legged, robot also exhibited another curious behavior: when one of its legs was removed (so that it had to relearn to walk) , it seemed to show signs of what is known as phantom limb syndrome, the sensation ta one still has a limb though it is in fact missing (this is common in people who have lost limbs in war or accidents). In humans, this syndrome represent a form of mental aberration or neurosis (perhaps even an hallucination). A robot acting in this way — holding a false notion of itself — may give scientists and AI engineers a glimpse into robot mental illness.
A robot with a mental illness or neurosis? Yes, this seem entirely likely given the following three theorems:
1] Neurosis is accompanied (and is perhaps a function of) acute self-awareness; the more self-aware, the more potentially neurotic one becomes.
2} Robots with advanced heuristics (enabled by multiple brains, self-simulators and sensor inputs) will inevitably develop advanced self-awareness, thus the greater potential for 1] above.
3] There is an ancient, magickal maxim: Like begets like. The creator is in the created (in Biblical terms: “God made man in his own image.”

What would Freud say about this form of attachment?

Mayhaps the ‘Age of Spiritual Machines‘ could become an ‘Age of Neurotic Machines‘ (or Psychotic Machines, depending on your view of humans), too. So then, f this is be the fate of  I, Robot, let’s do our droid druggs a favor and engineer a robo-shrink, or, at least, a good self-help program…and a love for Beethoven.