fMRI Data Reveals the Number of Parallel Processes Running in the Brain

The human brain carries out many tasks at the same time, but how many? Now fMRI data has revealed just how parallel gray matter is.

The human brain is often described as a massively parallel computing machine. That raises an interesting question: just how parallel is it?

Today, we get an answer thanks to the work of Harris Georgiou at the National Kapodistrian University of Athens in Greece, who has counted the number of “CPU cores” at work in the brain as it performs simple tasks in a functional magnetic resonance imaging (fMRI) machine. The answer could help lead to computers that better match the performance of the human brain.

The brain itself consists of around 100 billion neurons that each make up to 10,000 connections with their neighbors. All of this is packed into a structure the size of a party cake and operates at a peak power of only 20 watts, a level of performance that computer scientists observe with unconcealed envy.

fMRI machines reveal this activity by measuring changes in the levels of oxygen in the blood passing through the brain. The thinking is that more active areas use more oxygen so oxygen depletion is a sign of brain activity.

Typically, fMRI machines divide the brain into three-dimensional pixels called voxels, each about five cubic millimeters in size. The complete activity of the brain at any instant can be recorded using a three-dimensional grid of 60 x 60 x 30 voxels. These measurements are repeated every second or so, usually for tasks lasting two or three minutes. The result is a dataset of around 30 million data points.

Georgiou’s work is in determining the number of independent processes at work within this vast data set. “This is not much different than trying to recover the (minimum) number of actual ‘cpu cores’ required to ‘run’ all the active cognitive tasks that are registered in the entire 3-D brain volume,” he says.

This is a difficult task given the size of the dataset. To test his signal processing technique, Georgiou began by creating a synthetic fMRI dataset made up of eight different signals with statistical characteristics similar to those at work in the brain. He then used a standard signal processing technique, called independent component analysis, to work out how many different signals were present, finding that there are indeed eight, as expected.

Next, he applied the same independent component analysis technique to real fMRI data gathered from human subjects performing two simple tasks. The first was a simple visuo-motor task in which a subject watches a screen and then has to perform a simple task depending on what appears.

In this case, the screen displays either a red or green box on the left or right side. If the box is red, the subject must indicate this with their right index finger, and if the box is green, the subject indicates this with their left index finger. This is easier when the red box appears on the right and the green box appears on the left but is more difficult when the positions are swapped. The data consisted of almost 100 trials carried out on nine healthy adults.

The second task was easier. Subjects were shown a series of images that fall into categories such as faces, houses, chairs, and so on. The task was to spot when the same object appears twice, albeit from a different angle or under different lighting conditions. This is a classic visual recognition task.

The results make for interesting reading. Although the analysis is complex, the outcome is simple to state. Georgiou says that independent component analysis reveals that about 50 independent processes are at work in human brains performing the complex visuo-motor tasks of indicating the presence of green and red boxes. However, the brain uses fewer processes when carrying out simple tasks, like visual recognition.

That’s a fascinating result that has important implications for the way computer scientists should design chips intended to mimic human performance. It implies that parallelism in the brain does not occur on the level of individual neurons but on a much higher structural and functional level, and that there are about 50 of these.

Georgiou points out that a typical voxel corresponds to roughly three million neurons, each with several thousand connections with its neighbors. However, the current state-of-the-art neuromorphic chips contain a million artificial neurons each with only 256 connections. What is clear from this work is that the parallelism that Georgiou has measured occurs on a much larger scale than this.

“This means that, in theory, an artificial equivalent of a brain-like cognitive structure may not require a massively parallel architecture at the level of single neurons, but rather a properly designed set of limited processes that run in parallel on a much lower scale,” he concludes.

Anybody thinking of designing brain-like chips might find this a useful tip.

Ref: arxiv.org/abs/1410.7100 Estimating The Intrinsic Dimension In fMRI Space Via Dataset Fractal Analysis

ORIGINAL: Tech Review

November 5, 2014

IBM Chip Processes Data Similar to the Way Your Brain Does

A chip that uses a million digital neurons and 256 million synapses may signal the beginning of a new era of more intelligent computers.

WHY IT MATTERS

Computers that can comprehend messy data such as images could revolutionize what technology can do for us.

New thinking: IBM has built a processor designed using principles at work in your brain.

A new kind of computer chip, unveiled by IBM today, takes design cues from the wrinkled outer layer of the human brain. Though it is no match for a conventional microprocessor at crunching numbers, the chip consumes significantly less power, and is vastly better suited to processing images, sound, and other sensory data.

IBM’s SyNapse chip, as it is called, processes information using a network of just over one million “neurons,” which communicate with one another using electrical spikes—as actual neurons do. The chip uses the same basic components as today’s commercial chips—silicon transistors. But its transistors are configured to mimic the behavior of both neurons and the connections—synapses—between them.

The SyNapse chip breaks with a design known as the Von Neuman architecture that has underpinned computer chips for decades. Although researchers have been experimenting with chips modeled on brains—known as neuromorphic chips—since the late 1980s, until now all have been many times less complex, and not powerful enough to be practical (see “Thinking in Silicon”). Details of the chip were published today in the journal Science.

The new chip is not yet a product, but it is powerful enough to work on real-world problems. In a demonstration at IBM’s Almaden research center, MIT Technology Review saw one recognize cars, people, and bicycles in video of a road intersection. A nearby laptop that had been programed to do the same task processed the footage 100 times slower than real time, and it consumed 100,000 times as much power as the IBM chip. IBM researchers are now experimenting with connecting multiple SyNapse chips together, and they hope to build a supercomputer using thousands.

When data is fed into a SyNapse chip it causes a stream of spikes, and its neurons react with a storm of further spikes. The just over one million neurons on the chip are organized into 4,096 identical blocks of 250, an arrangement inspired by the structure of mammalian brains, which appear to be built out of repeating circuits of 100 to 250 neurons, says Dharmendra Modha, chief scientist for brain-inspired computing at IBM. Programming the chip involves choosing which neurons are connected, and how strongly they influence one another. To recognize cars in video, for example, a programmer would work out the necessary settings on a simulated version of the chip, which would then be transferred over to the real thing.

In recent years, major breakthroughs in image analysis and speech recognition have come from using large, simulated neural networks to work on data (see “Deep Learning”). But those networks require giant clusters of conventional computers. As an example, Google’s famous neural network capable of recognizing cat and human faces required 1,000 computers with 16 processors apiece (see “Self-Taught Software”).

Although the new SyNapse chip has more transistors than most desktop processors, or any chip IBM has ever made, with over five billion, it consumes strikingly little power. When running the traffic video recognition demo, it consumed just 63 milliwatts of power. Server chips with similar numbers of transistors consume tens of watts of power—around 10,000 times more.

The efficiency of conventional computers is limited because they store data and program instructions in a block of memory that’s separate from the processor that carries out instructions. As the processor works through its instructions in a linear sequence, it has to constantly shuttle information back and forth from the memory store—a bottleneck that slows things down and wastes energy.

IBM’s new chip doesn’t have separate memory and processing blocks, because its neurons and synapses intertwine the two functions. And it doesn’t work on data in a linear sequence of operations; individual neurons simply fire when the spikes they receive from other neurons cause them to.

Horst Simon, the deputy director of Lawrence Berkeley National Lab and an expert in supercomputing, says that until now the industry has focused on tinkering with the Von Neuman approach rather than replacing it, for example by using multiple processors in parallel, or using graphics processors to speed up certain types of calculations. The new chip “may be a historic development,” he says. “The very low power consumption and scalability of this architecture are really unique.”

One downside is that IBM’s chip requires an entirely new approach to programming. Although the company announced a suite of tools geared toward writing code for its forthcoming chip last year (see “IBM Scientists Show Blueprints for Brainlike Computing”), even the best programmers find learning to work with the chip bruising, says Modha: “It’s almost always a frustrating experience.” His team is working to create a library of ready-made blocks of code to make the process easier.

Asking the industry to adopt an entirely new kind of chip and way of coding may seem audacious. But IBM may find a receptive audience because it is becoming clear that current computers won’t be able to deliver much more in the way of performance gains. “This chip is coming at the right time,” says Simon.

ORIGINAL: Tech Review

By Tom Simonite

August 7, 2014

Palm's Jeff Hawkins is building a brain-like AI. He told us why he thinks his life's work is right

Inside a big bet on future machine intelligence

Feature Jeff Hawkins has bet his reputation, fortune, and entire intellectual life on one idea: that he understands the brain well enough to create machines with an intelligence we recognize as our own.

If his bet is correct, the Palm Pilot inventor will father a new technology, one that becomes the crucible in which a general artificial intelligence is one day forged. If his bet is wrong, then Hawkins will have wasted his life. At 56 years old that might sting a little.

"I want to bring about intelligent machines, machine intelligence, accelerated greatly from where it was going to happen and I don't want to be consumed – I want to come out at the other end as a normal person with my sanity," Hawkins told The Register. "My mission, the mission of Numenta, is to be a catalyst for machine intelligence."

A catalyst, he says, staring intently at your correspondent, "is something which accelerates a reaction by a thousand or ten thousand or a million-fold, and doesn't get consumed in the process."

His goal is ambitious, to put it mildly.

Before we dig deep into Hawkins' idiosyncratic approach to artificial intelligence, it's worth outlining the state of current AI research, why his critics have a right to be skeptical of his grandiose claims, and how his approach is different to the one being touted by consumer web giants such as Google.

AI researcher Jeff Hawkins

The road to a successful, widely deployable framework for an artificial mind is littered with failed schemes, dead ends, and traps. No one has come to the end of it, yet. But while major firms like Google and Facebook, and small companies like Vicarious, are striding over well-worn paths, Hawkins believes he is taking a new approach that could take him and his colleagues at his company,Numenta, all the way.

For over a decade, Hawkins has poured his energy into amassing enough knowledge about the brain and about how to program it in software. Now, he believes he is on the cusp of a great period of invention that may yield some very powerful technology.

Some people believe in him, others doubt him, and some academics El Reg has spoken with are suspicious of his ideas.

One thing we have established is that the work to which Hawkins has dedicated his life has become an influential touchstone within the red-hot modern artificial intelligence industry. His 2004 book, On Intelligence, appears to have been read by and inspired many of the most prominent figures in AI, and the tech Numenta is creating may trounce other commercial efforts by much larger companies such as Google, Facebook, and Microsoft.

"I think Jeff is largely right in what he wrote in On Intelligence," explains Hawkins' former colleague Dileep George (now running his own AI startup, Vicarious, which recently received $40m in funding from Mark Zuckerberg, space pioneer Elon Musk, and actor-turned-VC Ashton Kutcher). "Hierarchical systems, associative memory, time and attention – I think all those ideas are correct."

One of Google's most prominent AI experts agrees: "Jeff Hawkins ... has served as inspiration to countless AI researchers, for which I give him a lot of credit," explains former Google brain king and current Stanford Professor Andrew Ng.

Some organizations have taken Hawkins' ideas and stealthily run with them, with schemes already underway at companies like IBM and federal organizations like DARPA to implement his ideas in silicon, paving the way for neuromorphic processors that process information in near–real time, develop representations of patterns, and make predictions. If successful, these chips will make Qualcomm's "neuromorphic" Zeroth processors look like toys.

He has also inspired software adaptations of his work, such as CEPT, which has built an intriguing natural language processing engine partly out of Hawkins' ideas.

How we think: time and hierarchy

Hawkins' idea is that to build systems that behave like the brain, you have to be able to 

• take in a stream of changing information, 
• recognize patterns in it without knowing anything about the input source, 
• make predictions, and 
• react accordingly. 

The only context you have for this analysis is an ability to observe how the stream of data changes over time.

Though this sounds similar to some of the data processing systems being worked on by researchers at Google, Microsoft, and Facebook, it has some subtle differences.

Part of it is heritage – Hawkins traces his ideas back to his own understanding of how our neocortex works based on a synthesis of thousands of academic papers, chats with researchers, and his own work at two of his prior tech companies, Palm and Handspring, whereas the inspiration for most other approaches are neural networks based on technology from the 80s, which itself was refined out of a 1940s paper [PDF], "A Logical Calculus of the Ideas Immanent in Nervous Activity".

"That may be the right thing to do, but it's not the way brains work and it's not the principles of intelligence and it's not going to lead to a system that can explore the world or systems that can have behavior," Hawkins tells us.

So far he has outlined the ideas for this approach in his influential On Intelligence, plus a white paper published in 2011, a set of open source algorithms called NuPIC based on his Hierarchical Temporal Memory approach, and hundreds of talks given at universities and at companies ranging from Google to small startups.

Six easy pieces and the one true algorithm

Hawkins' work has "popularized the hypothesis that much of intelligence might be due to one learning algorithm," explains Ng.

Part of why Hawkins' approach is so controversial is that rather than assembling a set of advanced software components for specific computing functions and lashing them together via ever more complex collections of software, Hawkins has dedicated his research to figuring out an implementation of a single, basic approach.

This approach stems from an observation that our brain doesn't appear to come preloaded with any specific instructions or routines, but rather is an architecture that is able to take in, process, and store an endless stream of information and develop higher-order understandings out of that.

The manifestation of Hawkins' approach is the Cortical Learning Algorithm, or CLA.

"People used to think the neocortex was divided into sensory regions and motor regions," he explains. "We know now that is not true – the whole neocortex is sensory and motor."

Ultimately, the CLA will be a single system that involves both sensory processing and motor control – brain functions that Hawkins believes must be fused together to create the possibility of consciousness. For now, most work has been done on the sensory layer, though he has recently made some breakthroughs on the motor integration as well.

To build his Cortical Learning Algorithm system, Hawkins says, he has developed six principles that define a cortical-like processor. These traits are

• "on-line learning from streaming data", 
• "hierarchy of memory regions", 
• "sequence memory", 
• "sparse distributed representations",
•  "all regions are sensory and motor", and 
• "attention".

These principles are based on his own study of the work being done by neuroscientists around the world.

Now, Hawkins says, Numenta is on the verge of a breakthrough that could see the small company birth a framework for building intelligence machines. And unlike the hysteria that greeted AI in the 70s and 80s as the defense industry pumped money into AI, this time may not be a false dawn.

"I am thrilled at the progress we're making," he told El Reg one sunny afternoon at Numenta's whiteboard-crammed offices in Redwood City, California. "It's accelerating. These things are compounding, and it feels like these things are all coming together very rapidly."

The approach Numenta has been developing is producing better and better results, he says, and the CLA is gaining broader capabilities. In the past months, Hawkins has gone through a period of fecund creativity, and has solved one of the main problems that have bedeviled his system (temporal pooling), he says. He sees 2014 as a critical year for the company.

He is confident that he has bet correctly – but it's been a hard road to get here.

That long, hard road

Hawkins' interest in the brain dates back to his childhood, as does his frustration with how it is studied.

Growing up, Hawkins spent time with his father in an old shipyard on the north shore of Long Island, inventing all manner of boats with his father, an inventor with the enthusiasm for creativity of a Dr. Seuss character. In high school, the young Hawkins developed an interest in biophysics and, as he recounts in his book On Intelligence, tried to find out more about the brain at a local library.

"My search for a satisfying brain book turned up empty. I came to realize that no one had any idea how the brain actually worked. There weren't even any bad or unproven theories; there simply were none," he wrote.

This realization sparked a lifelong passion to try to understand the grand, intricate system that makes people who they are, and to eventually model the brain and create machines built in the same manner.

Hawkins graduated from Cornell in 1979 with a Bachelor of Science in Electronic Engineering. After a stint at Intel, he applied to MIT to study artificial intelligence, but had his application rejected because he wanted to understand how brains work, rather than build artificial intelligence. After this he worked at laptop start-up GRiD Systems, but during this time "could not get my curiosity about the brain and intelligent machines out of my head," so he did a correspondence course in physiology and ultimately applied to and was accepted in the biophysics program at the University of California, Berkeley.

When Hawkins started at Berkeley in 1986, his ambition to study a theory of the brain collided with the university administration, which disagreed with his course of study. Though Berkeley was not able to give him a course of study, Hawkins spent almost two years ensconced in the school's many libraries reading as much of the literature available on neuroscience as possible.

This deep immersion in neuroscience became the lens through which Hawkins viewed the world, with his later business accomplishments – Palm, Handspring – all leading to valuable insights on how the brain works and why the brain behaves as it does.

The way Hawkins recounts his past makes it seem as if the creation of a billion-dollar business in Palm, and arguably the prototype of the modern smartphone in Handspring, was a footnote along his journey to understand the brain.

This makes more sense when viewed against what he did in 2002, when he founded the Redwood Neuroscience Institute (now a part of the University of California at Berkeley and an epicenter of cutting-edge neuroscience research in its own right), and in 2005 founded Numenta with Palm/Handspring collaborator Donna Dublinksy and cofounder Dileep George.

These decades gave Hawkins the business acumen, money, and perspective needed to make a go at crafting his foundation for machine intelligence.

Controversy

His media-savvy, confident approach appears to have stirred up some ill feeling among other academics who point out, correctly, that Hawkins hasn't published widely, nor has he invented many ideas on his own.

Numenta has also had troubles, partly due to Hawkins' idiosyncratic view on how the brain works.

In 2010, for example, Numenta cofounder Dileep George left to found his own company, Vicarious, to pick some of the more low-hanging fruit in the promising field of AI. From what we understand, this amicable separation stemmed from a difference of opinion between George and Hawkins, as George tended towards a more mathematical approach, and Hawkins to a more biological one.

Hawkins has also come in for a bit of a drubbing from the intelligentsia, with NYU psychology professor Gary Marcus dismissing Numenta's approach in a New Yorker article titled "Steamrolled by Big Data".

Other academics El Reg interviewed for this article did not want to be quoted, as they felt Hawkins' lack of peer reviewed papers combined with his entrepreneurial persona reduced the credibility of his entire approach.

Hawkins brushes off these criticisms and believes they come down to a difference of opinion between him and the AI intelligentsia.

"These are complex biological systems that were not designed by mathematical principles [that are] very difficult to formalize completely," he told us.

"This reminds me a bit of the beginning of the computer era," he said. "If you go back to the 1930s and early 40s, when people first started thinking about computers they were really interested in whether an algorithm would complete, and they were looking for mathematical completeness, a mathematical proof, that if you implemented something like an algorithm today when we build a computer, no one sits around saying "Let's look at the mathematical formalism of this computer.' It reminds me a little about that. We still have people saying 'You don't have enough math here!' There's some people that just don't like that."

Hawkins' confidence stems from the way Numenta has built its technology, which far from merely taking inspiration from the brain – as many other startups claim to do – is actively built as a digital implementation of everything Hawkins has learned about how the dense, napkin-sized sheet of cells that is our neocortex works.

"I know of no other cortical theories/models that incorporate any of the following: 

• active dendrites, 
• differences between proximal and distal dendrites, 
• synapse growth and decay, 
• potential synapses, 
• dendrite growth, 
• depolarization as a mode of prediction, 
• mini-columns, 
• multiple types of inhibition and their corresponding inhibitory neurons, 
• etcetera. 

The new temporal pooling mechanism we are working on requires metabotropic receptors in the locations they are, and are not, found. Again, I don't know of any theories that have been reduced to practice that incorporate any, let alone all of these concepts," he wrote in a post to the discussion mailing list for NuPic, an open source implementation of Numenta's CLA, in February.

Deep learning is the new shallow learning

But for all the apparent rigorousness of Hawkins' approach, during the years he has worked on the technology there has been a fundamental change in the landscape of AI development: the rise of the consumer internet giants, and with them the appearance of various cavernous stores of user data on which to train learning algorithms.

Google, for instance, was said in January of 2014 to be assembling the team required for the "Manhattan Project for AI", according to a source who spoke anonymously to online publication Re/code. But Hawkins thinks that for all its grand aims, Google's approach may be based on a flawed presumption.

The collective term for the approach pioneered by companies like Google, Microsoft, and Facebook is "Deep Learning", but Hawkins fears it may be another blind path.

"Deep learning could be the greatest thing in the world, but it's not a brain theory," he says.

Deep learning approaches, Hawkins says, encourage the industry to go about refining methods based on old technology, itself based on an oversimplified version of the neurons in a brain.

Because of the vast stores of user data available, the companies are all compelled to approach the quest of creating artificial intelligence through building machines that compute over certain types of data.

In many cases, much of the development at places like Google, Microsoft, and Facebook has revolved around vision – a dead end, according to Hawkins.

"Where the whole community got tripped up – and I'm talking fifty years tripped up – is vision," Hawkins explains. "They said, 'Your eyes are moving all the time, your head is moving, the world is moving – let us focus on a simpler problem: spatial inference in vision'. This turns out to be a very small subset of what vision is. Vision turns out to be an inference problem. What that did is they threw out the most important part of vision – you must learn first how to do time-based vision."

The acquisitions these companies have made speak to this apparent flaw.

Google, for instance, hired AI luminary and University of Toronto professor Geoff Hinton and his startup DNNresearch last year to have him apply his "Deep Belief Networks" approach to Google's AI efforts.

In a talk given at the University of Toronto last year, Hinton said he believed more advanced AI should be based on existing approaches, rather than a rethought understanding of the brain.

"The kind of neural inspiration I like is when making it more like the brain works better," Hinton said. "There's lots of people who say you ought to make it more like the brain – like Henry Markram [of the European Union's brain simulation project], for example. He says, 'Give me a billion dollars and I'll make something like the brain,' but he doesn't actually know how to make it work – he just knows how to make something more and more like the brain. That seems to me not the right approach. What we should do is stick with things that actually work and make them more like the brain, and notice when making them more like the brain is actually helpful. There's not much point in making things work worse."

Hawkins vehemently disagrees with this point, and believes that basing approaches on existing methods means Hinton and other AI researchers are not going to be able to imbue their systems with the generality needed for true machine intelligence.

Another influential Googler agrees.

"We have neuroscientists in our team so we can be biologically inspired but are not slavish to it," Google Fellow Jeff Dean (creator of MapReduce, the Google File System, and now a figure in Google's own "Brain Project" team, also known as its AI division) told us this year.

"I'm surprised by how few people believe they need to understand how the brain works to build intelligent machines," Hawkins says. "I'm disappointed by this."

Hinton's foundational technologies, for example, are Boltzmann machines - advanced "stochastic recurrent neural network" tools that try to mimic some of the characteristics of the brain, which sit at the heart of Hinton's "Deep Belief Networks" (2006).

"The neurons in a restricted Boltzmann machine are not even close [to the brain] – it's not even an approximation," Hawkins explains.

Even Google is not sure about which way to bet on how to build a mind, as illustrated by its buy of UK company "DeepMind Technologies" earlier this year.

That company's founder, Demis Hassabis, has done detailed work on fundamental neuroscience, and has built technology out of this understanding. In 2010, it was reported that he mentioned both Hawkins' Hierarchical Temporal Memory and Hinton's Deep Belief Nets when giving a talk on viable general artificial intelligence approaches.

Facebook has gone down similar paths by hiring the influential artificial intelligence academic Yann LeCun to help it "predict what a user is going to do next," among other things.

Microsoft has developed significant capabilities as well, with systems like the Siri-beater "Cortana" and various endeavors by the company's research division, MSR.

Though the techniques these various researchers employ differ, they all depend on training a dataset over a large amount of information, and then selectively retraining it as information changes.

These AI efforts are built around dealing with problems backed up by large and relatively predictable datasets. This has yielded some incredible inventions, such as

• reasonable natural language processing, 
• image detection, and 
• video tagging.

It has not and cannot, however, yield a framework for a general intelligence, as it doesn't have the necessary architecture for data 

• apprehension, 
• analysis, 
• retention, and 
• recognition 

that our own brains do, Hawkins claims.

Hawkins' focus on time is why he believes his approach will win – something that the consumer internet giants are slowly waking up to.

It's all about time

"I would say that Hawkins is focusing more on how things unfold over time, which I think is very important," Google's research director Peter Norvig told El Reg via email, "while most of the current deep learning work assumes a static representation, unchanging over time. I suspect that as we scale up the applications (i.e., from still images to video sequences, and from extracting noun-phrase entities in text to dealing with whole sentences denoting actions), that there will be more emphasis on the unfolding of dynamic processes over time."

Another former Googler concurs, with Andrew Ng telling us via email, "Hawkins' work places a huge emphasis on learning from sequences. While most deep learning researchers also think that learning from sequences is important, we just haven't figured out ways to do so that we're happy with yet."

Geoff Hinton echoes this praise. "He has great insights about the types of computation the brain must be doing," he tells us – but argues that Jeff Hawkins' actual algorithmic contributions have been "disappointing" so far.

An absolutely crucial ingredient to AI

Time "is one hundred per cent crucial" to the creation of true artificial intelligence, Hawkins tells us. "If you accept the fact intelligent machines are going to work on the principles of the neocortex, it is the entire thing, basically. The only way."

"The brain does two things: 

• it does inference, which is recognizing patterns, and 
• it does behavior, which is generating patterns or generating motor behavior," 

Hawkins explains. "Ninety-nine percent of inference is time-based – language, audition, touch – it's all time-based. You can't understand touch without moving your hand. The order in which patterns occur is very important."

Numenta's approach relies on time. Its Cortical Learning Algorithm (white paper) amounts to an engine for

• processing streams of information, 
• classifying them, 
• learning to spot differences, and 
• using time-based patterns to make predictions about the future.

As mentioned above, there are several efforts underway at companies like IBM and federal research agencies like DARPA to implement Hawkins' systems in custom processors, and these schemes all recognize the importance of Hawkins' reliance on time.

"What I found intriguing about [his approach] – time is not an afterthought. In all of these [other] things, time has been an afterthought," one source currently working on implementing Hawkins' ideas tells us.

So far, Hawkins has used his system to make predictions of diverse phenomena such as 

• hourly energy use and 
• stock trading volumes, and 
• to detect anomalies in data streams. 

Numenta's commercial product, Grok, detects anomalies in computer servers running on Amazon's cloud service.

Hawkins described to us one way to understand the power of this type of pattern recognition. "Imagine you are listening to a musician," he suggested. "After hearing her play for several days, you learn the kind of music she plays, how talented she is, how much she improvises, and how many mistakes she makes. Your brain learns her style, and then has expectations about what she will play and what it will sound like. As you continue to listen to her play, you will detect if her style changes, if the type of music she plays changes, or if she starts making more errors. The same kind of patterns exist in machine-generated data, and Grok will detect changes."

Here again the wider AI community appears to be dovetailing into Hawkins' ideas, with one of Andrew Ng's former Stanford students Honglak Lee having published a paper called "A classification-based polyphonic piano transcription approach using learned feature representations" in 2011. However, the method if implementation is different.

Obscurity through biology

Part of the reason why Hawkins' technology is not more widely known is because for current uses it is hard for it to demonstrate a vast lead over rival approaches. For all of Hawkins' belief in the tech, it is hard to demonstrate a convincing killer application for it that other approaches can't do. The point, Hawkins says, is that the CLA's internal structure gets rid of some of the stumbling blocks that exist in the future of other approaches.

Hawkins believes the CLA's implicit dependence on time means that eventually it will become the dominant approach.

"At the bottom of the [neocortex's] hierarchy are fast-changing patterns and they form sequences – some of them are predictable and some of them are not – and what the neocortex is doing is trying to understand the set of patterns here and give it a constant representation – a name for the sequence, if you will – and it forms that as the next level of the hierarchy so the next level up is more stable," Hawkins explains.

"Changing patterns lead to changing representations in the hierarchy that are more stable, and then it learns the changes in those patterns, and as you go up the hierarchy it forms more and more stable representations of the world and they also tend to be independent of your body position and your senses."

A comparison between Hawkins' Hierarchical Temporal Memory cells (right), 

a neural network neuron (center), and the brain's own neuron (left)

He believes his technology is more effective than the approaches taken by his rivals due to its use of sparse distributed representations as an input device to a storage system he terms "sequence memory".

Sequence memory refers to how information makes its way into the brain as a stream of information that comes in from both external stimuli and internal stimuli, such as signals from the broader body.

Sparse Distributed Representations (SDRs) are partially based on the work of mathematician Pentti Kanerva on "Sparse Distributed Memory" [PDF].

They refer to how the brain represents and stores information. They are designed to mimic the way our brain is believed to encode memories, which is through neuron firings across a very large area in response to inputs. To achieve this, SDRs are written, roughly, as a 2000-bit string of which perhaps two percent are active. This means that you don't need to read all active bits in an SDR to say that it is similar to another, because it merely needs to share a few of the activated bits to be considered similar, due to the sparsity.

Hawkins believes SDRs give input data inherent meaning through this representation approach.

"This means that if two vectors have 1s in the same position, they are semantically similar. Vectors can therefore be expressed in degrees of similarity rather than simply being identical or different. These large vectors can be stored accurately even using a subsampled index of, say, 10 of 2,000 bits. This makes SDR memory fault tolerant to gaps in data. SDRs also exhibit properties that reliably allow the neocortex to determine if a new input is unexpected," the company's commercial website for Grok says.

But what are the drawbacks?

So if Hawkins thinks he has the theory and is on the way to building the technology, and other companies are implementing it, then why are we even calling what he is doing a "bet"? The answer comes down to credibility.

Hawkins' idiosyncratic nature and decision to synthesize insights from two different fields – neuroscience and computer science – are his strengths, but also his drawbacks.

"No one knows how the cortex works, so there is no way to know if Jeff is on the right track or not," Dr. Terry Sejnowski, the laboratory head of the Computational Neurobiology Laboratory at the SALK Institute for Biological Studies, tells us. "To the extent that [Hawkins] incorporates new data into his models he may have a shot, and there will be a flood of data coming from the BRAIN Initiative that was announced by Obama last April."

Hawkins says that this response is typical of the academic community, and that there is enough data available to learn about the brain. You just have to look for it.

"We're not going to replicate the neocortex, we're not going to simulate the neocortex, we just need to understand how it works in sufficient detail so we can say 'A-ha!' and build things like it," Hawkins says. "There is an incredible amount of unassimilated data that exists. Fifty years of papers. Thousands of papers a year. It's unbelievable, and it's always the next set of papers that people think is going to do it. ... it's not true that you have to wait for that stuff."

The root of the problems Hawkins faces may be his approach, which stems more from biology than from mathematics. His old colleague and cofounder of Numenta, Dileep George, confirms this.

"I think Jeff is largely right in what he wrote in On Intelligence," George told us. "There are different approaches on how to bring those ideas. Jeff has an angle on it; we have a different angle on it; the rest of the community have another perspective on it."

These ideas are echoed by Google's Norvig. "Hawkins, at least in his general-public-facing-persona, seems to be more driven by duplicating what the brain does, while the deep learning researchers take some concepts from the brain, but then mostly are trying to optimize mathematical equations," he told us via email.

"I live in the middle," Hawkins explains. "Where I know the neuroscience details very very well, and I have a theoretical framework, and I bounce back and forth between these over and over again."

The future

Hawkins reckons that what he is doing today "is maybe 5 per cent of how humans learn," Hawkins says.

He believes that during the coming year he will begin work on the next major area of development for his technology: action.

For Hawkins' machines to gain independence – the ability, say, to not only recognize and classify patterns, but actively tune themselves to hunt for specific bits of information – the motor component needs to be integrated, he explains.

"What we've proven so far – I say built and tested and put into a product – is pure sensor. It's like an ear listening to sounds that doesn't have a chance to move," he tells us.

If you can add in the motor component, "an entire world opens up," he says.

"For example, I could have something like a web bot – an internet crawler. Today's web crawlers are really stupid, they're like wall-following rats. They just go up and down the length up and down the length," he says.

"If I wanted to look and understand the web, I could have a virtual system that is basically moving through cyberspace thinking about 'What is the structure here? How do I model this?' And so that's an example of a behavioral system that has no physical presence. It basically says, 'OK, I'm looking at this data, now where do I go next to look? Oh, I'm going to follow this link and do that in an intelligent way'."

By creating this technology, Hawkins hopes to dramatically accelerate the speed with which generally applicable artificial intelligence is developed and integrated into our world.

It's taken a lot to get here, and the older Hawkins gets and the more rival companies spend, the bigger the stakes get. As of 2014, he is still betting his life on the fact that he is right and they are wrong. ®

ORIGINAL: The Register

By Jack Clark,

29 Mar 2014

Can The Human Brain Project Succeed?

Image: Getty Images

An ambitious effort to build human brain simulation capability is meeting with some very human resistance. On Monday, a group of researchers sent an open letter to the European Commission protesting the management of the Human Brain Project, one of two Flagship initiatives selected last year to receive as much as €1 billion over the course of 10 years (the other award went to a far less controversy-courting project devoted to graphene).

The letter, which now has more than 450 signatories, questions the direction of the project and calls for a careful, unbiased review. Although he’s not mentioned by name in the letter, news reports cited resistance to the path chosen by project leader Henry Markram of the Swiss Federal Institute of Technology in Lausanne. One particularly polarizing change was the recent elimination of a subproject, called Cognitive Architectures, as the project made its bid for the next round of funding.

According to Markram, the fuss all comes down to differences in scientific culture. He has described the project, which aims to build six different computing platforms for use by researchers, as an attempt to build a kind of CERN for brain research, a means by which disparate disciplines and vast amounts of data can be brought together. This is a "methodological paradigm shift" for neuroscientists accustomed to individual research grants, Markram told Science, and that's what he says the letter signers are having trouble with.

But some question the main goals of the project, and whether we're actually capable of achieving them at this point. The program's Brain Simulation Platform aims to build the technology needed to reconstruct the mouse brain and eventually the human brain in a supercomputer. Part of the challenge there is technological. Markram has said that an exascale-level machine (one capable of executing 1000 or more petaflops) would be needed to "get a first draft of the human brain", and the energy requirements of such machines are daunting.

Crucially, some experts say that even if we had the computational might to simulate the brain, we're not ready to. "The main apparent goal of building the capacity to construct a larger-scale simulation of the human brain is radically premature," signatory Peter Dayan, who directs a computational neuroscience department at University College London, told the Guardian. He called the project a "waste of money" that "can't but fail from a scientific perspective". To Science, he said "the notion that we know enough about the brain to know what we should simulate is crazy, quite frankly.”

This last comment resonated with me, as it reminded me of a feature that Steve Furber of the University of Manchester wrote for IEEE Spectrum a few years ago. Furber, one of the co-founders of the mobile chip design powerhouse ARM, is now in the process of stringing a million or so of the low-power processors together to build a massively parallel computer capable of simulating 1 billion neurons, about 1% as many as are contained in the human brain.

Furber and his collaborators designed their computing architecture quite carefully in order to take into account the fact that there are still a host of open questions when it comes to basic brain operation. General-purpose computers are power-hungry and slow when it comes to brain simulation. Analog circuitry, which is also on the Human Brain Project's list, might better mimic the way neurons actually operate, but, he wrote,

“as speedy and efficient as analog circuits are, they’re not very flexible; their basic behavior is pretty much baked right into them. And that’s unfortunate, because neuroscientists still don’t know for sure which biological details are crucial to the brain’s ability to process information and which can safely be abstracted away”

The Human Brain Project's website admits that exascale computing will be hard to reach: "even in 2020, we expect that supercomputers will have no more than 200 petabytes." To make up for the shortfall, it says, "what we plan to do is build fast storage random-access storage systems next to the supercomputer, store the complete detailed model there, and then allow our multi-scale simulation software to call in a mix of detailed or simplified models (models of neurons, synapses, circuits, and brain regions) that matches the needs of the research and the available computing power. This is a pragmatic strategy that allows us to keep build ever more detailed models, while keeping our simulations to the level of detail we can support with our current supercomputers."

This does sound like a flexible approach. But, as is par for the course with any ambitious research project, particularly one that involves a great amount of synthesis of disparate fields, it's not yet clear whether it will pay off.

And any big changes in direction may take a while. Although the proposal for the second round of funding will be reviewed this year, according to Science, which reached out to the European Commission, the first review of the project itself won't begin until January 2015.

Rachel Courtland can be found on Twitter at @rcourt.

ORIGINAL: Spectrum
By Rachel Courtland
Posted 9 Jul 2014 | 17:00 GMT

DARPA Wants a Memory Prosthetic for Injured Vets—and Wants It Now

Photo: Getty Images

No one will ever fault DARPA, the Defense Department's mad science wing, for not being ambitious enough. Over the next four years, the first grantees in its Restoring Active Memory (RAM) program are expected to develop and test prosthetic memory devices that can be implanted in the human brain. 

It's hoped that such synthetic devices can help veterans with traumatic brain injuries, and other people whose natural memory function is impaired. The two teams, led by researchers Itzhak Fried at UCLA and Mike Kahana at the University of Pennsylvania, will start with the fundamentals. 

They'll look for neural signals associated with the formation and recall of memories, and they'll work on computational models to describe how neurons carry out these processes, and to determine how an artificial device can replicate them. They'll also work with partners to develop real hardware suitable for the human brain. Such devices should ultimately be capable of recording the electrical activity of neurons, processing the information, and then stimulating other neurons as needed.The RAM research derives from an engineering approach to memory that's gaining traction. (Spectrum covered the work of one of its leading proponents, Ted Berger, in the recent article The End of Disability.) If the brain is essentially a collection of circuits, the thinking goes, a memory is formed by the sequential actions of many neurons. If a person has a brain injury that knocks out some of those neurons, the whole circuit may malfunction, and the person will experience memory problems. But if electrodes can pick up the signal in the neurons upstream from the problem spot, and then convey that signal around the damage to intact neurons downstream, then the memory should function as normal.

In a press briefing yesterday, program manager Justin Sanchez said that the first human experiments will be conducted with hospitalized epilepsy patients who have electrodes implanted in their brains as they await surgery (this is done so their doctors can pinpoint the origin of their seizures). Since epilepsy patients often experience memory loss as well, Sanchez said they're a natural fit for the research. Eventually trials would include military servicemembers who suffer the aftereffects of traumatic brain injuries, and finally civilians with similar injuries. 

DARPA recently decided to beef up its research in biological technologies, spurred in part by the needs of veterans returning from Iraq and Afghanistan. But it seems likely that the agency's increased attention to programs like RAM was also prompted by the recognition that neural engineering is one of the most exciting frontiers in science, with the neural technologies advancing faster than the science that guides it.

The RAM program is part of the overarching federal BRAIN Initiative, announced with much fanfare by President Obama in 2013. With a first-year budget of $110 million parceled out to three agencies and considerable cooperation from deep-pocketed private institutions, you can expect this decade to be a brainy one.

ORIGINAL: Spectrum
By Eliza Strickland
9 Jul 2014

Neuroscientists Join the Open-Source Hardware Movement

Two MIT grad students offer up DIY brain-recording gear

Photo: Open Ephys

Graduate students Josh Siegle and Jakob Voigts were planning an ambitious series of experiments at their MIT neuroscience labs in 2011 when they ran into a problem. They needed to record complex brain signals from mice, but they couldn’t afford the right equipment: The recording systems cost upward of US $60,000 each, and they wanted at least four. So they decided to solve their dilemma by building their own gear on the cheap. And knowing that they wouldn’t be the last neuroscientists to encounter such a problem, they decided to give away their designs. Now their project, Open Ephys, is the hub of a nascent open-source hardware community for neural technology.

Siegle and Voigts weren’t knowledgeable about either circuit design or coding, but they learned as they went along. By July 2013, they were ready to manufacture 50 of their recording systems, which they gave to collaborators for beta testing. This spring they manufactured 100 improved units, which are now arriving in neuroscience labs around the world. They estimate that each system costs about $3,000 to produce.

Neuroscience has a history of hackers, Siegle says, with researchers cobbling together their own gear or customizing commercial systems to meet their particular needs. But those new tools rarely leave the labs they are built in. So scientists spend a lot of time reinventing the wheel. The goal of Open Ephys (which is short for open-source electrophysiology) is not just to distribute the tools that Siegle and Voigts have come up with so far but to encourage researchers to put resources into developing open-source tools for the benefit of the whole community. “In addition to changing the tools, we also want to change the culture,” Siegle says.

Photo: Open Ephys Open Ephys just distributed 100 of its acquisition boards to neuroscience labs around the world.

The flagship tool that Siegle and Voigts developed is an acquisition board, which makes sense of the electric signals from electrodes implanted in an animal’s brain. The board interfaces with up to eight headstages that amplify, filter, multiplex, and digitize signals from the brain, and then sends those signals to a computer for further processing. Commercial systems typically have individual ICs perform each of those four functions, but Siegle and Voigts’s system uses a single microchip for the four steps. The chip was recently developed by Intan Technologies, based in Los Angeles. “Once we realized these chips were available, it seemed kind of silly to keep buying the big systems,” Siegle says.

The president and cofounder of Intan, Reid Harrison, says that shrinking and consolidating the gear wasn’t that complicated—it mostly required initiative. “It’s such a niche market that no one else had tried to miniaturize the technology,” he says. “It’s not exactly on the scale of CPUs and cellphones, which drive most IC technology.” However, Harrison says he recognized a need for his small, multipurpose chips. Neuroscientists are always trying to fit more electrodes into an animal’s brain to record more neural activity, he says, which requires ever tinier devices with the electronics close to the electrodes. “You could put 1,000 electrodes in the brain, but you don’t want 1,000 wires on an animal that’s supposed to be mobile,” he says. The Intan chips take information from up to 64 electrodes and turn it into one digital signal, eliminating the confusion of wiring.

The major neural technology companies have designed products that incorporate Intan’s chips, but they also swear by their larger, multichip systems. Keith Stengel, the founder of Neuralynx, in Bozeman, Mont., says that in his big systems, each component is optimized for peak performance. “A lot of our customers have said that you buy a Neuralynx system for the serious work that you’re going to publish, and then you get an Open Ephys system as a second system, for grad students to start their research on,” he says.

 

Illustration: Open Ephys Open Ephys offers building instructions for this head-mounted neural implant system for mice.

Andy Gotshalk, CEO of Blackrock Microsystems, in Salt Lake City, also argues that the commercial products will continue to be the gold standard. “You’re not going to be moving into FDA clinical trials using an Open Ephys system,” he says. The commercial products come with guarantees of quality and reliability, he says, as well as intensive customer support. Gotshalk says his customers are willing to pay a premium for that backing.

Both Stengel and Gotshalk say they welcome Open Ephys to the market and think that its systems can fill a niche. They’re also willing to work with the upstart to make sure their commercial software works with the Open Ephys hardware. Harrison agrees that the community is happy to have another option to work with, and he draws a parallel to the computing industry. “The existing tools are like the PCs and the Macs of the neuroscience world, but now we also have this Linux,” Harrison says. “It’s a lot less expensive, and you can hack it yourself, but it’s not for everyone.”

ORIGINAL: IEEE Spectrum

By Eliza Strickland
Posted 11 Jun 2014

Mathematical Model Of Consciousness Proves Human Experience Cannot Be Modelled On A Computer

A new mathematical model of consciousness implies that your PC will never be conscious in the way you are

The Physics arXiv Blog in The Physics arXiv Blog

One of the most profound advances in science in recent years is the way researchers from a variety of fields are beginning to think about consciousness. Until now, the c-word was been taboo for most scientists. Any suggestion that a researchers was interested in this area would be tantamount to professional suicide.

That has begun to change thanks to a new theory of consciousness developed in the last ten years or so by Giulio Tononi, a neuroscientist at the University of Wisconsin in Madison, and others. Tononi’s key idea is that consciousness is phenomenon in which information is integrated in the brain in a way that cannot be broken down.

So each instant of consciousness integrates the smells, sounds and sights of that moment of experience. And consciousness is simply the feeling of this integrated information experience.

What makes Tononi’s ideas different from other theories of consciousness is that it can be modelled mathematically using ideas from physics and information theory. That doesn’t mean this theory is correct. But it does mean that, for the first time, neuroscientists, biologists physicists and anybody else can all reason about consciousness using the universal language of science: mathematics.

This has led to an extraordinary blossoming of ideas about consciousness. A few months ago, for example, we looked at how physicists are beginning to formulate the problem consciousness in terms of quantum mechanics and information theory.

Today, Phil Maguire at the National University of Ireland and a few pals take this mathematical description even further. These guys make some reasonable assumptions about the way information can leak out of a consciousness system and show that this implies that consciousness is not computable. In other words, consciousness cannot be modelled on a computer.

Maguire and co begin with a couple of thought experiments that demonstrate the nature of integrated information in Tononi’s theory. They start by imagining the process of identifying chocolate by its smell. For a human, the conscious experience of smelling chocolate is unified with everything else that a person has smelled (or indeed seen, touched, heard and so on).

This is entirely different from the process of automatically identifying chocolate using an electronic nose, which measures many different smells and senses chocolate when it picks out the ones that match some predefined signature.

A key point here is that it would be straightforward to access the memory in an electronic nose and edit the information about its chocolate experience. You could delete this with the press of a button.

But ask a neuroscientist to do the same for your own experience of the smell of chocolate—to somehow delete this—and he or she would be faced with an impossible task since the experience is correlated with many different parts of the brain.

Indeed, the experience will be integrated with all kinds of other experiences. “According to Tononi, the information generated by such [an electronic nose] differs from that generated by a human insofar as it is not integrated,” say Maguire and co.

This process of integration is then crucial and Maguire and co focus on the mathematical properties it must have. For instance, they point out that the process of integrating information, of combining it with many other aspects of experience, can be thought of as a kind of information compression.

This compression allows the original experience to be constructed but does not keep all of the information it originally contained.

To better understand this, they give as an analogy the sequence of numbers: 4, 6, 8, 12, 14, 18, 20, 24…. This is an infinite series defined as: odd primes plus 1. This definition does not contain all the infinite numbers but it does allow it be reproduced. It is clearly a compression of the information in the original series.

The brain, say Maguire and co, must work like this when integrating information from a conscious experience. It must allow the reconstruction of the original experience but without storing all the parts.

That leads to a problem. This kind of compression inevitably discards information. And as more information is compressed, the loss becomes greater.

But if our memories were like that cannot be like that, they would be continually haemorrhaging meaningful content. “Memory functions must be vastly non-lossy, otherwise retrieving them repeatedly would cause them to gradually decay,” say Maguire and co.

The central part of their new work is to describe the mathematical properties of a system that can store integrated information in this way but without it leaking away. And this leads them to their central proof. “The implications of this proof are that we have to abandon either the idea that people enjoy genuinely [integrated] consciousness or that brain processes can be modelled computationally,” say Maguire and co.

Since Tononi’s main assumption is that consciousness is the experience of integrated information, it is the second idea that must be abandoned: brain processes cannot be modelled computationally.

They go on to discuss this in more detail. If a person’s behaviour cannot be analysed independently from the rest of their conscious experience, it implies that something is going on in their brain that is so complex it cannot feasibly be reversed, they say.

In other words, the difference between cognition and computation is that computation is reversible whereas cognition is not. And they say that is reflected in the inability of a neuroscientist to operate and remove a particular memory of the small of chocolate.

That’s an interesting approach but it is one that is likely to be controversial. The laws of physics are computable, as far as we know. So critics might ask how the process of consciousness can take place at all if it is non-computable. Critics might even say this is akin to saying that consciousness is in some way supernatural, like magic.

But Maguire and go counter this by saying that their theory doesn’t imply that consciousness is objectively non-computable only subjectively so. In other words, a God-like observer with perfect knowledge of the brain would not consider it non-computable. But for humans, with their imperfect knowledge of the universe, it is effectively non-computable.

There is something of a card trick about this argument. In mathematics, the idea of non-computability is not observer-dependent so it seems something of a stretch to introduce it as an explanation.

What’s more, critics might point to other weaknesses in the formulation of this problem. For example, the proof that conscious experience is non-computable depends critically on the assumption that our memories are non-lossy.

But everyday experience is surely the opposite—our brains lose most of the information that we experience consciously. And the process of repeatedly accessing memories can cause them to change and degrade. Isn’t the experience of forgetting a face of a known person well documented?

Then again, critics of Maguire and co’s formulation of the problem of consciousness must not lose sight of the bigger picture—that the debate about consciousness can occur on a mathematical footing at all. That’s indicative of a sea change in this most controversial of fields.

Of course, there are important steps ahead. Perhaps the most critical is that the process of mathematical modelling must lead to hypotheses that can be experimentally tested. That’s the process by which science distinguishes between one theory and another. Without a testable hypothesis, a mathematical model is not very useful.

For example, Maguire and co could use their model to make predictions about the limits in the way information can leak from a conscious system. These limits might be testable in experiments focusing on the nature of working memory or long-term memory in humans.

That’s the next challenge for this brave new field of consciousness.

Ref: arxiv.org/abs/1405.0126 : Is Consciousness Computable? Quantifying Integrated Information Using Algorithmic Information Theory

Follow the Physics arXiv Blog on Twitter at @arxivblog, on Facebook and by hitting the Follow button below.

ORIGINAL: Medium