Jeremy Howard: Deep Learning, AI and the Future of Jobs

Jeremy Howard, CEO of Enlitic, is exploring these capabilities for medical applications. He was an early adopter of neural-network and big-data methodologies in the 1990s. As the president & chief scientist of Kaggle, a platform for data science competitions, he witnessed the rise of an algorithmic method called “deep learning”.

In a recent interview he describes that in 2012 deep neural networks started becoming good at things that previously only humans were able to do, particularly at understanding the content of images. Image recognition may sound like a fairly niche application, but when you think about it, it is actually critical. Computers before were blind. Today they are now more accurate and much faster at recognising objects in pictures than humans are.

He explains the difference between humans and machines is that once you create a specific module, you don’t have to have every machine learn it. We can download these networks between machines, but we can’t download knowledge between brains. This is a huge benefit of deep-learning machines that we refer to as “transfer learning”. The only thing holding back he states the growth of machine learning is 1) data access, and 2) the ability to do logic.

Deep learning is at the point that in many areas it is already exceeding human learning and intelligence. With deep learning there’s a mathematical proof that it can model anything that can be modelled as long as it has enough computing capacity and data to learn it. Instead of being a physical engine, it is an intellectual engine. There is now, in fact, a deep-learning network for building deep-learning networks! And, indeed, the networks that they come up with are, in fact, much better than those that humans have created.

Summing up he states: “Human-learning theory really is all about the power of creating additional connectivity in the brain, using things like mnemonics and context to use stuff we already learned to help us learn other things. It’s coloured my understanding of the transfer learning that we talked about. For me, the most exciting work in deep learning right now is this ability to transfer knowledge across networks and have networks that are continually getting better, not just at one particular thing but at new things as well.”

Howard, Bostrom, Hugo de Garis, and Elon Musk all agree that this future endeavor will have consequences for humanity and that we need to think through our sociological, ethical, and economic reasoning about this new world that is emerging more quickly than we at first imagined. That AI will mimic us and surpass us in aspects of machine learning through such techniques and algorithms as Deep Learning is only a beginning. For as this systems begin to self-learn and have greater access to data and the logics of access and reasoning we do not know where it will lead. Below I gather a few of the youtube documentaries from these gentlemen.

Whereas Howard explores the advances in war, medicine, home robotics other aspects of the economic sphere that will allow that will drive this process, he reminds us that the sociological impact to humans themselves may be enormous. Bostrom believes we need to set up the initial conditions of such algorithms to produce and ethical and value intensive process otherwise the AI’s of our near future may be asymmetrical and amoral to human needs and desires, while also becoming eventually our rivals and competitors and see us as future threats to their own well being and advancement. de Garis and Musk take that theme of threat to its logical conclusion and warn us of where we are going. I’ll leave it to the viewer to make up their own mind. I’m working with one such scenario in my cartoon noir version of how an AI multiplicity might effect freedom, change, and exit from its human controllers and dominion. Where that will lead is anyone’s guess…

Deep Learning

For more information on Deep Learning: http://deeplearning.net/tutorial/

Deep Learning unlike simulations of the Brain is about enforce and adapt rules: the ability to take the right decisions, according to some criterion (e.g. survival and reproduction, for most animals). To take better decisions requires knowledge, in a form that is operational, i.e., can be used to interpret sensory data and use that information to take decisions.

Computers already possess some intelligence thanks to all the programs that humans have crafted and which allow them to “do things” that we consider useful (and that is basically what we mean for a computer to take the right decisions). But there are many tasks which animals and humans are able to do rather easily but remain out of reach of computers, at the beginning of the 21st century. Many of these tasks fall under the label of Artificial Intelligence, and include many perception and control tasks. Why is it that we have failed to write programs for these tasks? It is mostly because we do not know explicitly (formally) how to do these tasks, even though our brain (coupled with a body) can do them. Doing those tasks involve knowledge that is currently implicit, but we have information about those tasks through data and examples (e.g. observations of what a human would do given a particular request or input). How do we get machines to acquire that kind of intelligence? Using data and examples to build operational knowledge is what learning is about.

Deep learning is part of a broader family of machine learning methods based on learning representations of data. An observation (e.g., an image) can be represented in many ways such as a vector of intensity values per pixel, or in a more abstract way as a set of edges, regions of particular shape, etc.. Some representations make it easier to learn tasks (e.g., face recognition or facial expression recognition) from examples. One of the promises of deep learning is replacing handcrafted features with efficient algorithms for unsupervised or semi-supervised feature learning and hierarchical feature extraction.

Research in this area attempts to make better representations and create models to learn these representations from large-scale unlabeled data. Some of the representations are inspired by advances in neuroscience and are loosely based on interpretation of information processing and communication patterns in a nervous system, such as neural coding which attempts to define a relationship between the stimulus and the neuronal responses and the relationship among the electrical activity of the neurons in the brain.

Various deep learning architectures such as deep neural networks, convolutional deep neural networks, deep belief networks and recurrent neural networks have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. – from Wiki

As Geoff Hinton recognizes realistically, deep learning is only part of the larger challenge of building intelligent machines. Such techniques lack ways of representing causal relationships (such as between diseases and their symptoms), and are likely to face challenges in acquiring abstract ideas like “sibling” or “identical to.” They have no obvious ways of performing logical inferences, and they are also still a long way from integrating abstract knowledge, such as information about what objects are, what they are for, and how they are typically used. The most powerful A.I. systems, like Watson, the machine that beat humans in “Jeopardy,” use techniques like deep learning as just one element in a very complicated ensemble of techniques, ranging from the statistical technique of Bayesian inference to deductive reasoning.

As Nature reports early skeptics have now become believers: “But winning over computer-vision scientists would take more: they wanted to see gains on standardized tests. Malik remembers that Hinton asked him: “You’re a sceptic. What would convince you?” Malik replied that a victory in the internationally renowned ImageNet competition might do the trick.” More:

In that competition, teams train computer programs on a data set of about 1 million images that have each been manually labelled with a category. After training, the programs are tested by getting them to suggest labels for similar images that they have never seen before. They are given five guesses for each test image; if the right answer is not one of those five, the test counts as an error. Past winners had typically erred about 25% of the time. In 2012, Hinton’s lab entered the first ever competitor to use deep learning. It had an error rate of just 15%.

“Deep learning stomped on everything else,” says LeCun, who was not part of that team. The win landed Hinton a part-time job at Google, and the company used the program to update its Google+ photo-search software in May 2013.

Malik was won over. “In science you have to be swayed by empirical evidence, and this was clear evidence,” he says. Since then, he has adapted the technique to beat the record in another visual-recognition competition. Many others have followed: in 2013, all entrants to the ImageNet competition used deep learning.

Such advance Deep Learning techniques are now being used in various scientific industries: In 2012, the pharmaceutical company Merck offered a prize to whoever could beat its best programs for helping to predict useful drug candidates. The task was to trawl through database entries on more than 30,000 small molecules, each of which had thousands of numerical chemical-property descriptors, and to try to predict how each one acted on 15 different target molecules. Dahl and his colleagues won $22,000 with a deep-learning system. “We improved on Merck’s baseline by about 15%,” he says.

Ray Kurzweil is Hedging his bets by using a combination of both Quantum Computing on non-AI related as well as Deep Learning: Ray Kurzweil to pursue various ways for computers to learn from experience — using techniques including but not limited to deep learning. Last May, Google acquired a quantum computer made by D-Wave in Burnaby, Canada. This computer holds promise for non-AI tasks such as difficult mathematical computations — although it could, theoretically, be applied to deep learning.

D-Wave and Quantum computing

One facet of this technology may be the convergence toward Quantum Computing using the Deep Learning algorithms. We are only at the beginning of this stage of advanced quantum computing and many problems are still in front of it. Yet, a quantum computer built by D-Wave is already being used by NASA, Google, and other corporations and governmental agencies.

D-Wave, the company that built the thing, calls it the world’s first quantum computer, a seminal creation that foretells the future of mathematical calculation. But many of the world’s experts see it quite differently, arguing the D-Wave machine is something other than the computing holy grail the scientific community has sought since the mid-1980s.

According to D-Wave, the machine contains 512 superconducting circuits, each a tiny loop of flowing current. These are cooled to almost absolute zero, the company says, so they enter a quantum state where the current flows both clockwise and counterclockwise at the same time. When you feed the machine a task, it uses a set of algorithms to map a calculation across these qubits — and then execute that calculation. Basically, this involves determining the probability that a given set of circuits will emerge in a particular pattern when the temperature inside the system is raised.

Rather than simulating the brain it uses other forms and algorithms. As Wired reports:

What they can say for sure is that the system doesn’t use simulated annealing, which is essentially a means of searching for a mathematical solution. According to Lidar, simulated annealing is akin to looking for the lowest possible point in a vast landscape.

“We call it an energy landscape,” he says. “There is a solution hiding somewhere in that landscape, and you can imagine that solution is hiding at the lowest point on the surface. You’re trying to find that lowest point.” This is done by randomly traveling across the landscape, moving down “hills” and back up them, until you locate the deepest valley.

As Harmut Neven of Google says: “We believe quantum computing may help solve some of the most challenging computer science problems, particularly in machine learning,” said a Google director of engineering, wrote in a blog post. “Machine learning is all about building better models of the world to make more accurate predictions.”

The NSA is using such q-computing by prepping quantum-resistant algorithms to head off crypto-apocalypse. As Dan Goodwin says at the moment, quantum computers are believed to be little more than a theoretical phenomenon. Consider, for instance, that the biggest number factored to date using Shor’s algorithm is just 21. But a significant percentage of computer scientists say practical quantum computing is only a matter of time, and once that happens (anywhere in the next 10 to 50 years, most of them forecast), public-key crypto systems that form the bedrock of most modern data protection will be trivial to break. Such a doomsday scenario would jeopardize not only all transactions and records going forward, but it would also allow attackers to decrypt more than half a century’s worth of old communications, assuming someone took the time to collect and store the encrypted data.

“What the NSA is saying in this release is that they are really worried about quantum computers right now, worried enough that they are planning a big effort to switch all of the public-key crypto used by anyone who interacts with classified data to post-quantum algorithms,” Nadia Heninger, an assistant professor of computer and information science at the University of Pennsylvania, wrote in an e-mail to Ars. “This will have a big impact for the security industry, since companies that produce products that they want to sell to the government will need to implement these algorithms, and they are likely to be deployed widely as a result.”

There is even a D-Wave blog to further your information needs on this area.

None of these are state of the art, but the convergence of all these various technologies along with the exponential curve of advancement is bringing about certain forms of algorithmic machinic intelligence in various areas of advance industry applications. Whether it will ever mimic humans is of little concern to the industrial and digital corporations. What is important is that this convergence could be used for command and control of vast global resources and populations through decisional processes that are far beyond our own capacity. This isn’t fiction even if the anthropomorphic imprint of human personality onto these systems is fantasy. One needs to separate our fears from what is truly happening in this machine learning in its exponential growth in industry, science, and governmental systems.

The economics angle

As Forbes reports Shivon Zilis, an investor at BloombergBETA in San Francisco, put together the graphic below to show what she calls the Machine Intelligence Landscape. The fund specifically focuses on “companies that change the world of work,” so these sorts of automation are a large area of concern. Zilis explains, “I created this landscape to start to put startups into context. I’m a thesis-oriented investor and it’s much easier to identify crowded areas and see white space once the landscape has some sort of taxonomy.”

As the article suggests what is striking in this landscape is how filled-in it is. At the top are core technologies that power the applications below. Big American companies like Google, IBM, Microsoft, Facebook and China’s Baidu are well-represented in the core technologies themselves. These companies, particularly Google, are also the prime bidders for core startups as well. Many of the companies that describe themselves as engaging in artificial intelligence, deep learning or machine learning have some claim to general algorithms that work across multiple types of applications. Others specialize in the areas of natural language processing, prediction, image recognition and speech recognition.

Mobility is the new game in town: And what most concerns the big tech companies from Apple to Google to Microsoft and IBM? Yep, mobile, and as Zilis points out, “Winning mobile will require lots of machine intelligence.” Siri and Google Now are responses to the need for highly contextual voice interaction in mobile. Visual search like Amazon’s FireFly involves location-based pattern recognition to create a pleasing experience. The reason for the current great enthusiasm for deep learning is that these kinds of problems can be solved now in minutes or days instead of years.

Other companies like Numenta offer an instructive contrast according to Anthony Wing Cosner to the enthusiasm for deep learning. Although deep learning uses neural networks, Hawkins claims that the Numenta approach is significantly more brain-like. The difference is that Numenta’s method uses Hierarchical Temporal Memory (HTM), which can natively discern patterns in time, as well as computational space.

But against those who think machine intelligence will never simulate the Brain. They may be right, but machine intelligence does not need to resemble the human brain at all. As Jeremy Howard said in a recent AMA on Reddit, “The more interesting question is: what can machines do? Not ‘are they truly intelligent?’ Machine ‘intelligence’ is different enough from human intelligence that I don’t think it is a terribly useful analogy.” What neuroscience and cognitive psychology do inform is an understanding of what kinds of tasks are performed by which systems and circuits of systems in the brain.

So this distinction between human and machine intelligence is central to many current notions regarding what an AI might become in the future. As Kosner concludes “Machine intelligence in general and deep learning in particular will have a significant impact on what happens in technology in the coming years. Large tech companies with vast data holdings will be particularly motivated to extract value from all of this data now that there appears to be a scalable way to do so. On the other end of the spectrum, app developers will be encouraged to ramp up the input factories they entice people to place on their smartphones now that the output has a clear value. Machine learning and data science talent will continue to move from academia to big companies like Google for access to all that data.”

TedX Talks and other Investigations

Two youtube.com TedX Talks relate this in detail and his worries about our social systems as machine intelligence overtakes many areas of the knowledge industry that humans thought would be secure in our economic world:

---

Nick Bostrom explores What happens when our computers get smarter than we are?  

An artificial intellect (or “artilect”), according to Dr. Hugo de Garis, is a computer intelligence superior to that of humans in one or more spheres of knowledge together with an implicit will to use the intelligence. Artilects are the concern of artificial intelligence specialists (or “intelligists”) like de Garis, who speculates that human society may soon have to face the question of whether and how we can restrain artificial intelligence from making decisions inimical to humans.

Dr. de Garis assumes that within one or two generations, we will have computers that are more sophisticated than human brains with the ability to experimentally evolve their intelligence into something much beyond what humans might contemplate or understand. de Garis wonders whether such machines would consider human beings important enough to preserve. He speculates that society will soon need to face the question of whether we should permit artilects to be built. He foresees two factions arising: the Cosmists, who argue that they should be built, and the Terras, believing that they should not. The Cosmists might believe that artilects would probably want to leave our planet to seek intelligences elsewhere in the universe. The Terras believe that it would be too dangerous for the human race to allow artilects to be developed.

More information: Artilect World. Below a History Channel Docu:

Bill Gates, Elon Musk, Steven Hawkings on “Summoning the Demon”:

“Hope we’re not the biological bootloader for digital SuperIntelligence. Unfortunately, that is increasingly probable…”

---

Current literature and non-fiction

• The Relativistic Brain: How it works and why it cannot be simulated by a Turing machine – provides a critique of Simulation Theories of AI by Ronald Cicurel and Miguel Nicolelis refuting the possibility that any Turing machine will ever succeed in such a simulation.
• Superintelligence: Paths, Dangers, Strategies by Nick Bostrom – He asks the question: Will it be possible to construct a seed AI or otherwise to engineer initial conditions so as to make an intelligence explosion survivable?
• Our Final Invention: Artificial Intelligence and the End of the Human Era by James Barat He asks Can we coexist with beings whose intelligence dwarfs our own? And will they allow us to?
• The Artificial Intelligence Revolution: Will Artificial Intelligence Serve Us Or Replace Us? by Louis A. Del Monte. Del Monte and other AI experts predict that AI capabilities will develop into SAMs with abilities far beyond what human beings can even fathom.
• Turing’s Cathedral: The Origins of the Digital Universe by George Dyson Great background on the history of the sciences and technologies of information processing
• Ex Machina by Alex Garland. The novel “script format” from which the movie was made.