Tag Archives: AlphaGo

Brain stuff: quantum entanglement and a multi-dimensional universe

I have two brain news bits, one about neural networks and quantum entanglement and another about how the brain operates on more than three dimensions.

Quantum entanglement and neural networks

A June 13, 2017 news item on phys.org describes how machine learning can be used to solve problems in physics (Note: Links have been removed),

Machine learning, the field that’s driving a revolution in artificial intelligence, has cemented its role in modern technology. Its tools and techniques have led to rapid improvements in everything from self-driving cars and speech recognition to the digital mastery of an ancient board game.

Now, physicists are beginning to use machine learning tools to tackle a different kind of problem, one at the heart of quantum physics. In a paper published recently in Physical Review X, researchers from JQI [Joint Quantum Institute] and the Condensed Matter Theory Center (CMTC) at the University of Maryland showed that certain neural networks—abstract webs that pass information from node to node like neurons in the brain—can succinctly describe wide swathes of quantum systems.

An artist’s rendering of a neural network with two layers. At the top is a real quantum system, like atoms in an optical lattice. Below is a network of hidden neurons that capture their interactions (Credit: E. Edwards/JQI)

A June 12, 2017 JQI news release by Chris Cesare, which originated the news item, describes how neural networks can represent quantum entanglement,

Dongling Deng, a JQI Postdoctoral Fellow who is a member of CMTC and the paper’s first author, says that researchers who use computers to study quantum systems might benefit from the simple descriptions that neural networks provide. “If we want to numerically tackle some quantum problem,” Deng says, “we first need to find an efficient representation.”

On paper and, more importantly, on computers, physicists have many ways of representing quantum systems. Typically these representations comprise lists of numbers describing the likelihood that a system will be found in different quantum states. But it becomes difficult to extract properties or predictions from a digital description as the number of quantum particles grows, and the prevailing wisdom has been that entanglement—an exotic quantum connection between particles—plays a key role in thwarting simple representations.

The neural networks used by Deng and his collaborators—CMTC Director and JQI Fellow Sankar Das Sarma and Fudan University physicist and former JQI Postdoctoral Fellow Xiaopeng Li—can efficiently represent quantum systems that harbor lots of entanglement, a surprising improvement over prior methods.

What’s more, the new results go beyond mere representation. “This research is unique in that it does not just provide an efficient representation of highly entangled quantum states,” Das Sarma says. “It is a new way of solving intractable, interacting quantum many-body problems that uses machine learning tools to find exact solutions.”

Neural networks and their accompanying learning techniques powered AlphaGo, the computer program that beat some of the world’s best Go players last year (link is external) (and the top player this year (link is external)). The news excited Deng, an avid fan of the board game. Last year, around the same time as AlphaGo’s triumphs, a paper appeared that introduced the idea of using neural networks to represent quantum states (link is external), although it gave no indication of exactly how wide the tool’s reach might be. “We immediately recognized that this should be a very important paper,” Deng says, “so we put all our energy and time into studying the problem more.”

The result was a more complete account of the capabilities of certain neural networks to represent quantum states. In particular, the team studied neural networks that use two distinct groups of neurons. The first group, called the visible neurons, represents real quantum particles, like atoms in an optical lattice or ions in a chain. To account for interactions between particles, the researchers employed a second group of neurons—the hidden neurons—which link up with visible neurons. These links capture the physical interactions between real particles, and as long as the number of connections stays relatively small, the neural network description remains simple.

Specifying a number for each connection and mathematically forgetting the hidden neurons can produce a compact representation of many interesting quantum states, including states with topological characteristics and some with surprising amounts of entanglement.

Beyond its potential as a tool in numerical simulations, the new framework allowed Deng and collaborators to prove some mathematical facts about the families of quantum states represented by neural networks. For instance, neural networks with only short-range interactions—those in which each hidden neuron is only connected to a small cluster of visible neurons—have a strict limit on their total entanglement. This technical result, known as an area law, is a research pursuit of many condensed matter physicists.

These neural networks can’t capture everything, though. “They are a very restricted regime,” Deng says, adding that they don’t offer an efficient universal representation. If they did, they could be used to simulate a quantum computer with an ordinary computer, something physicists and computer scientists think is very unlikely. Still, the collection of states that they do represent efficiently, and the overlap of that collection with other representation methods, is an open problem that Deng says is ripe for further exploration.

Here’s a link to and a citation for the paper,

Quantum Entanglement in Neural Network States by Dong-Ling Deng, Xiaopeng Li, and S. Das Sarma. Phys. Rev. X 7, 021021 – Published 11 May 2017

This paper is open access.

Blue Brain and the multidimensional universe

Blue Brain is a Swiss government brain research initiative which officially came to life in 2006 although the initial agreement between the École Politechnique Fédérale de Lausanne (EPFL) and IBM was signed in 2005 (according to the project’s Timeline page). Moving on, the project’s latest research reveals something astounding (from a June 12, 2017 Frontiers Publishing press release on EurekAlert),

For most people, it is a stretch of the imagination to understand the world in four dimensions but a new study has discovered structures in the brain with up to eleven dimensions – ground-breaking work that is beginning to reveal the brain’s deepest architectural secrets.

Using algebraic topology in a way that it has never been used before in neuroscience, a team from the Blue Brain Project has uncovered a universe of multi-dimensional geometrical structures and spaces within the networks of the brain.

The research, published today in Frontiers in Computational Neuroscience, shows that these structures arise when a group of neurons forms a clique: each neuron connects to every other neuron in the group in a very specific way that generates a precise geometric object. The more neurons there are in a clique, the higher the dimension of the geometric object.

“We found a world that we had never imagined,” says neuroscientist Henry Markram, director of Blue Brain Project and professor at the EPFL in Lausanne, Switzerland, “there are tens of millions of these objects even in a small speck of the brain, up through seven dimensions. In some networks, we even found structures with up to eleven dimensions.”

Markram suggests this may explain why it has been so hard to understand the brain. “The mathematics usually applied to study networks cannot detect the high-dimensional structures and spaces that we now see clearly.”

If 4D worlds stretch our imagination, worlds with 5, 6 or more dimensions are too complex for most of us to comprehend. This is where algebraic topology comes in: a branch of mathematics that can describe systems with any number of dimensions. The mathematicians who brought algebraic topology to the study of brain networks in the Blue Brain Project were Kathryn Hess from EPFL and Ran Levi from Aberdeen University.

“Algebraic topology is like a telescope and microscope at the same time. It can zoom into networks to find hidden structures – the trees in the forest – and see the empty spaces – the clearings – all at the same time,” explains Hess.

In 2015, Blue Brain published the first digital copy of a piece of the neocortex – the most evolved part of the brain and the seat of our sensations, actions, and consciousness. In this latest research, using algebraic topology, multiple tests were performed on the virtual brain tissue to show that the multi-dimensional brain structures discovered could never be produced by chance. Experiments were then performed on real brain tissue in the Blue Brain’s wet lab in Lausanne confirming that the earlier discoveries in the virtual tissue are biologically relevant and also suggesting that the brain constantly rewires during development to build a network with as many high-dimensional structures as possible.

When the researchers presented the virtual brain tissue with a stimulus, cliques of progressively higher dimensions assembled momentarily to enclose high-dimensional holes, that the researchers refer to as cavities. “The appearance of high-dimensional cavities when the brain is processing information means that the neurons in the network react to stimuli in an extremely organized manner,” says Levi. “It is as if the brain reacts to a stimulus by building then razing a tower of multi-dimensional blocks, starting with rods (1D), then planks (2D), then cubes (3D), and then more complex geometries with 4D, 5D, etc. The progression of activity through the brain resembles a multi-dimensional sandcastle that materializes out of the sand and then disintegrates.”

The big question these researchers are asking now is whether the intricacy of tasks we can perform depends on the complexity of the multi-dimensional “sandcastles” the brain can build. Neuroscience has also been struggling to find where the brain stores its memories. “They may be ‘hiding’ in high-dimensional cavities,” Markram speculates.

###

About Blue Brain

The aim of the Blue Brain Project, a Swiss brain initiative founded and directed by Professor Henry Markram, is to build accurate, biologically detailed digital reconstructions and simulations of the rodent brain, and ultimately, the human brain. The supercomputer-based reconstructions and simulations built by Blue Brain offer a radically new approach for understanding the multilevel structure and function of the brain. http://bluebrain.epfl.ch

About Frontiers

Frontiers is a leading community-driven open-access publisher. By taking publishing entirely online, we drive innovation with new technologies to make peer review more efficient and transparent. We provide impact metrics for articles and researchers, and merge open access publishing with a research network platform – Loop – to catalyse research dissemination, and popularize research to the public, including children. Our goal is to increase the reach and impact of research articles and their authors. Frontiers has received the ALPSP Gold Award for Innovation in Publishing in 2014. http://www.frontiersin.org.

Here’s a link to and a citation for the paper,

Cliques of Neurons Bound into Cavities Provide a Missing Link between Structure and Function by Michael W. Reimann, Max Nolte, Martina Scolamiero, Katharine Turner, Rodrigo Perin, Giuseppe Chindemi, Paweł Dłotko, Ran Levi, Kathryn Hess, and Henry Markram. Front. Comput. Neurosci., 12 June 2017 | https://doi.org/10.3389/fncom.2017.00048

This paper is open access.

Does understanding your pet mean understanding artificial intelligence better?

Heather Roff’s take on artificial intelligence features an approach I haven’t seen before. From her March 30, 2017 essay for The Conversation (h/t March 31, 2017 news item on phys.org),

It turns out, though, that we already have a concept we can use when we think about AI: It’s how we think about animals. As a former animal trainer (albeit briefly) who now studies how people use AI, I know that animals and animal training can teach us quite a lot about how we ought to think about, approach and interact with artificial intelligence, both now and in the future.

Using animal analogies can help regular people understand many of the complex aspects of artificial intelligence. It can also help us think about how best to teach these systems new skills and, perhaps most importantly, how we can properly conceive of their limitations, even as we celebrate AI’s new possibilities.
Looking at constraints

As AI expert Maggie Boden explains, “Artificial intelligence seeks to make computers do the sorts of things that minds can do.” AI researchers are working on teaching computers to reason, perceive, plan, move and make associations. AI can see patterns in large data sets, predict the likelihood of an event occurring, plan a route, manage a person’s meeting schedule and even play war-game scenarios.

Many of these capabilities are, in themselves, unsurprising: Of course a robot can roll around a space and not collide with anything. But somehow AI seems more magical when the computer starts to put these skills together to accomplish tasks.

Thinking of AI as a trainable animal isn’t just useful for explaining it to the general public. It is also helpful for the researchers and engineers building the technology. If an AI scholar is trying to teach a system a new skill, thinking of the process from the perspective of an animal trainer could help identify potential problems or complications.

For instance, if I try to train my dog to sit, and every time I say “sit” the buzzer to the oven goes off, then my dog will begin to associate sitting not only with my command, but also with the sound of the oven’s buzzer. In essence, the buzzer becomes another signal telling the dog to sit, which is called an “accidental reinforcement.” If we look for accidental reinforcements or signals in AI systems that are not working properly, then we’ll know better not only what’s going wrong, but also what specific retraining will be most effective.

This requires us to understand what messages we are giving during AI training, as well as what the AI might be observing in the surrounding environment. The oven buzzer is a simple example; in the real world it will be far more complicated.

Before we welcome our AI overlords and hand over our lives and jobs to robots, we ought to pause and think about the kind of intelligences we are creating. …

Source: pixabay.com

It’s just last year (2016) that an AI system beat a human Go master player. Here’s how a March 17, 2016 article by John Russell for TechCrunch described the feat (Note: Links have been removed),

Much was written of an historic moment for artificial intelligence last week when a Google-developed AI beat one of the planet’s most sophisticated players of Go, an East Asia strategy game renowned for its deep thinking and strategy.

Go is viewed as one of the ultimate tests for an AI given the sheer possibilities on hand. “There are 1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 possible positions [in the game] — that’s more than the number of atoms in the universe, and more than a googol times larger than chess,” Google said earlier this year.

If you missed the series — which AlphaGo, the AI, won 4-1 — or were unsure of exactly why it was so significant, Google summed the general importance up in a post this week.

Far from just being a game, Demis Hassabis, CEO and Co-Founder of DeepMind — the Google-owned company behind AlphaGo — said the AI’s development is proof that it can be used to solve problems in ways that humans may be not be accustomed or able to do:

We’ve learned two important things from this experience. First, this test bodes well for AI’s potential in solving other problems. AlphaGo has the ability to look “globally” across a board—and find solutions that humans either have been trained not to play or would not consider. This has huge potential for using AlphaGo-like technology to find solutions that humans don’t necessarily see in other areas.

I find Roff’s thesis intriguing and is likely applicable to the short-term but in the longer term and in light of the attempts to  create devices that mimic neural plasticity and neuromorphic engineering  I don’t find her thesis convincing.

Deep learning and some history from the Swiss National Science Foundation (SNSF)

A June 27, 2016 news item on phys.org provides a measured analysis of deep learning and its current state of development (from a Swiss perspective),

In March 2016, the world Go champion Lee Sedol lost 1-4 against the artificial intelligence AlphaGo. For many, this was yet another defeat for humanity at the hands of the machines. Indeed, the success of the AlphaGo software was forged in an area of artificial intelligence that has seen huge progress over the last decade. Deep learning, as it’s called, uses artificial neural networks to process algorithmic calculations. This software architecture therefore mimics biological neural networks.

Much of the progress in deep learning is thanks to the work of Jürgen Schmidhuber, director of the IDSIA (Istituto Dalle Molle di Studi sull’Intelligenza Artificiale) which is located in the suburbs of Lugano. The IDSIA doctoral student Shane Legg and a group of former colleagues went on to found DeepMind, the startup acquired by Google in early 2014 for USD 500 million. The DeepMind algorithms eventually wound up in AlphaGo.

“Schmidhuber is one of the best at deep learning,” says Boi Faltings of the EPFL Artificial Intelligence Lab. “He never let go of the need to keep working at it.” According to Stéphane Marchand-Maillet of the University of Geneva computing department, “he’s been in the race since the very beginning.”

A June 27, 2016 SNSF news release (first published as a story in Horizons no. 109 June 2016) by Fabien Goubet, which originated the news item, goes on to provide a brief history,

The real strength of deep learning is structural recognition, and winning at Go is just an illustration of this, albeit a rather resounding one. Elsewhere, and for some years now, we have seen it applied to an entire spectrum of areas, such as visual and vocal recognition, online translation tools and smartphone personal assistants. One underlying principle of machine learning is that algorithms must first be trained using copious examples. Naturally, this has been helped by the deluge of user-generated content spawned by smartphones and web 2.0, stretching from Facebook photo comments to official translations published on the Internet. By feeding a machine thousands of accurately tagged images of cats, for example, it learns first to recognise those cats and later any image of a cat, including those it hasn’t been fed.

Deep learning isn’t new; it just needed modern computers to come of age. As far back as the early 1950s, biologists tried to lay out formal principles to explain the working of the brain’s cells. In 1956, the psychologist Frank Rosenblatt of the New York State Aeronautical Laboratory published a numerical model based on these concepts, thereby creating the very first artificial neural network. Once integrated into a calculator, it learned to recognise rudimentary images.

“This network only contained eight neurones organised in a single layer. It could only recognise simple characters”, says Claude Touzet of the Adaptive and Integrative Neuroscience Laboratory of Aix-Marseille University. “It wasn’t until 1985 that we saw the second generation of artificial neural networks featuring multiple layers and much greater performance”. This breakthrough was made simultaneously by three researchers: Yann LeCun in Paris, Geoffrey Hinton in Toronto and Terrence Sejnowski in Baltimore.

Byte-size learning

In multilayer networks, each layer learns to recognise the precise visual characteristics of a shape. The deeper the layer, the more abstract the characteristics. With cat photos, the first layer analyses pixel colour, and the following layer recognises the general form of the cat. This structural design can support calculations being made upon thousands of layers, and it was this aspect of the architecture that gave rise to the name ‘deep learning’.

Marchand-Maillet explains: “Each artificial neurone is assigned an input value, which it computes using a mathematical function, only firing if the output exceeds a pre-defined threshold”. In this way, it reproduces the behaviour of real neurones, which only fire and transmit information when the input signal (the potential difference across the entire neural circuit) reaches a certain level. In the artificial model, the results of a single layer are weighted, added up and then sent as the input signal to the following layer, which processes that input using different functions, and so on and so forth.

For example, if a system is trained with great quantities of photos of apples and watermelons, it will progressively learn to distinguish them on the basis of diameter, says Marchand-Maillet. If it cannot decide (e.g., when processing a picture of a tiny watermelon), the subsequent layers take over by analysing the colours or textures of the fruit in the photo, and so on. In this way, every step in the process further refines the assessment.

Video games to the rescue

For decades, the frontier of computing held back more complex applications, even at the cutting edge. Industry walked away, and deep learning only survived thanks to the video games sector, which eventually began producing graphics chips, or GPUs, with an unprecedented power at accessible prices: up to 6 teraflops (i.e., 6 trillion calculations per second) for a few hundred dollars. “There’s no doubt that it was this calculating power that laid the ground for the quantum leap in deep learning”, says Touzet. GPUs are also very good at parallel calculations, a useful function for executing the innumerable simultaneous operations required by neural networks.
Although image analysis is getting great results, things are more complicated for sequential data objects such as natural spoken language and video footage. This has formed part of Schmidhuber’s work since 1989, and his response has been to develop recurrent neural networks in which neurones communicate with each other in loops, feeding processed data back into the initial layers.

Such sequential data analysis is highly dependent on context and precursory data. In Lugano, networks have been instructed to memorise the order of a chain of events. Long Short Term Memory (LSTM) networks can distinguish ‘boat’ from ‘float’ by recalling the sound that preceded ‘oat’ (i.e., either ‘b’ or ‘fl’). “Recurrent neural networks are more powerful than other approaches such as the Hidden Markov models”, says Schmidhuber, who also notes that Google Voice integrated LSTMs in 2015. “With looped networks, the number of layers is potentially infinite”, says Faltings [?].

For Schmidhuber, deep learning is just one aspect of artificial intelligence; the real thing will lead to “the most important change in the history of our civilisation”. But Marchand-Maillet sees deep learning as “a bit of hype, leading us to believe that artificial intelligence can learn anything provided there’s data. But it’s still an open question as to whether deep learning can really be applied to every last domain”.

It’s nice to get an historical perspective and eye-opening to realize that scientists have been working on these concepts since the 1950s.