Tag Archives: memory
“We cannot be conscious of what we are not conscious of.” – Julian Jaynes, The Origin of Consciousness in the Breakdown of the Bicameral Mind
Unlike the director leads you to believe, the protagonist of Ex Machina, Andrew Garland’s 2015 masterpiece, isn’t Caleb, a young programmer tasked with evaluating machine consciousness. Rather, it’s his target Ava, a breathtaking humanoid AI with a seemingly child-like naïveté and an enigmatic mind.
Like most cerebral movies, Ex Machina leaves the conclusion up to the viewer: was Ava actually conscious? In doing so, it also cleverly avoids a thorny question that has challenged most AI-centric movies to date: what is consciousness, and can machines have it?
Hollywood producers aren’t the only people stumped. As machine intelligence barrels forward at breakneck speed—not only exceeding human performance on games such as DOTA and Go, but doing so without the need for human expertise—the question has once more entered the scientific mainstream.
Are machines on the verge of consciousness?
This week, in a review published in the prestigious journal Science, cognitive scientists Drs. Stanislas Dehaene, Hakwan Lau and Sid Kouider of the Collège de France, University of California, Los Angeles and PSL Research University, respectively, argue: not yet, but there is a clear path forward.
The reason? Consciousness is “resolutely computational,” the authors say, in that it results from specific types of information processing, made possible by the hardware of the brain.
There is no magic juice, no extra spark—in fact, an experiential component (“what is it like to be conscious?”) isn’t even necessary to implement consciousness.
If consciousness results purely from the computations within our three-pound organ, then endowing machines with a similar quality is just a matter of translating biology to code.
Much like the way current powerful machine learning techniques heavily borrow from neurobiology, the authors write, we may be able to achieve artificial consciousness by studying the structures in our own brains that generate consciousness and implementing those insights as computer algorithms.
From Brain to Bot
Without doubt, the field of AI has greatly benefited from insights into our own minds, both in form and function.
For example, deep neural networks, the architecture of algorithms that underlie AlphaGo’s breathtaking sweep against its human competitors, are loosely based on the multi-layered biological neural networks that our brain cells self-organize into.
Reinforcement learning, a type of “training” that teaches AIs to learn from millions of examples, has roots in a centuries-old technique familiar to anyone with a dog: if it moves toward the right response (or result), give a reward; otherwise ask it to try again.
In this sense, translating the architecture of human consciousness to machines seems like a no-brainer towards artificial consciousness. There’s just one big problem.
“Nobody in AI is working on building conscious machines because we just have nothing to go on. We just don’t have a clue about what to do,” said Dr. Stuart Russell, the author of Artificial Intelligence: A Modern Approach in a 2015 interview with Science.
The hard part, long before we can consider coding machine consciousness, is figuring out what consciousness actually is.
To Dehaene and colleagues, consciousness is a multilayered construct with two “dimensions:” C1, the information readily in mind, and C2, the ability to obtain and monitor information about oneself. Both are essential to consciousness, but one can exist without the other.
Say you’re driving a car and the low fuel light comes on. Here, the perception of the fuel-tank light is C1—a mental representation that we can play with: we notice it, act upon it (refill the gas tank) and recall and speak about it at a later date (“I ran out of gas in the boonies!”).
“The first meaning we want to separate (from consciousness) is the notion of global availability,” explains Dehaene in an interview with Science. When you’re conscious of a word, your whole brain is aware of it, in a sense that you can use the information across modalities, he adds.
But C1 is not just a “mental sketchpad.” It represents an entire architecture that allows the brain to draw multiple modalities of information from our senses or from memories of related events, for example.
Unlike subconscious processing, which often relies on specific “modules” competent at a defined set of tasks, C1 is a global workspace that allows the brain to integrate information, decide on an action, and follow through until the end.
Like The Hunger Games, what we call “conscious” is whatever representation, at one point in time, wins the competition to access this mental workspace. The winners are shared among different brain computation circuits and are kept in the spotlight for the duration of decision-making to guide behavior.
Because of these features, C1 consciousness is highly stable and global—all related brain circuits are triggered, the authors explain.
For a complex machine such as an intelligent car, C1 is a first step towards addressing an impending problem, such as a low fuel light. In this example, the light itself is a type of subconscious signal: when it flashes, all of the other processes in the machine remain uninformed, and the car—even if equipped with state-of-the-art visual processing networks—passes by gas stations without hesitation.
With C1 in place, the fuel tank would alert the car computer (allowing the light to enter the car’s “conscious mind”), which in turn checks the built-in GPS to search for the next gas station.
“We think in a machine this would translate into a system that takes information out of whatever processing module it’s encapsulated in, and make it available to any of the other processing modules so they can use the information,” says Dehaene. “It’s a first sense of consciousness.”
In a way, C1 reflects the mind’s capacity to access outside information. C2 goes introspective.
The authors define the second facet of consciousness, C2, as “meta-cognition:” reflecting on whether you know or perceive something, or whether you just made an error (“I think I may have filled my tank at the last gas station, but I forgot to keep a receipt to make sure”). This dimension reflects the link between consciousness and sense of self.
C2 is the level of consciousness that allows you to feel more or less confident about a decision when making a choice. In computational terms, it’s an algorithm that spews out the probability that a decision (or computation) is correct, even if it’s often experienced as a “gut feeling.”
C2 also has its claws in memory and curiosity. These self-monitoring algorithms allow us to know what we know or don’t know—so-called “meta-memory,” responsible for that feeling of having something at the tip of your tongue. Monitoring what we know (or don’t know) is particularly important for children, says Dehaene.
“Young children absolutely need to monitor what they know in order to…inquire and become curious and learn more,” he explains.
The two aspects of consciousness synergize to our benefit: C1 pulls relevant information into our mental workspace (while discarding other “probable” ideas or solutions), while C2 helps with long-term reflection on whether the conscious thought led to a helpful response.
Going back to the low fuel light example, C1 allows the car to solve the problem in the moment—these algorithms globalize the information, so that the car becomes aware of the problem.
But to solve the problem, the car would need a “catalog of its cognitive abilities”—a self-awareness of what resources it has readily available, for example, a GPS map of gas stations.
“A car with this sort of self-knowledge is what we call having C2,” says Dehaene. Because the signal is globally available and because it’s being monitored in a way that the machine is looking at itself, the car would care about the low gas light and behave like humans do—lower fuel consumption and find a gas station.
“Most present-day machine learning systems are devoid of any self-monitoring,” the authors note.
But their theory seems to be on the right track. The few examples whereby a self-monitoring system was implemented—either within the structure of the algorithm or as a separate network—the AI has generated “internal models that are meta-cognitive in nature, making it possible for an agent to develop a (limited, implicit, practical) understanding of itself.”
Towards conscious machines
Would a machine endowed with C1 and C2 behave as if it were conscious? Very likely: a smartcar would “know” that it’s seeing something, express confidence in it, report it to others, and find the best solutions for problems. If its self-monitoring mechanisms break down, it may also suffer “hallucinations” or even experience visual illusions similar to humans.
Thanks to C1 it would be able to use the information it has and use it flexibly, and because of C2 it would know the limit of what it knows, says Dehaene. “I think (the machine) would be conscious,” and not just merely appearing so to humans.
If you’re left with a feeling that consciousness is far more than global information sharing and self-monitoring, you’re not alone.
“Such a purely functional definition of consciousness may leave some readers unsatisfied,” the authors acknowledge.
“But we’re trying to take a radical stance, maybe simplifying the problem. Consciousness is a functional property, and when we keep adding functions to machines, at some point these properties will characterize what we mean by consciousness,” Dehaene concludes.
Image Credit: agsandrew / Shutterstock.com Continue reading
The brain has long inspired the design of computers and their software. Now Intel has become the latest tech company to decide that mimicking the brain’s hardware could be the next stage in the evolution of computing.
On Monday the company unveiled an experimental “neuromorphic” chip called Loihi. Neuromorphic chips are microprocessors whose architecture is configured to mimic the biological brain’s network of neurons and the connections between them called synapses.
While neural networks—the in vogue approach to artificial intelligence and machine learning—are also inspired by the brain and use layers of virtual neurons, they are still implemented on conventional silicon hardware such as CPUs and GPUs.
The main benefit of mimicking the architecture of the brain on a physical chip, say neuromorphic computing’s proponents, is energy efficiency—the human brain runs on roughly 20 watts. The “neurons” in neuromorphic chips carry out the role of both processor and memory which removes the need to shuttle data back and forth between separate units, which is how traditional chips work. Each neuron also only needs to be powered while it’s firing.
At present, most machine learning is done in data centers due to the massive energy and computing requirements. Creating chips that capture some of nature’s efficiency could allow AI to be run directly on devices like smartphones, cars, and robots.
This is exactly the kind of application Michael Mayberry, managing director of Intel’s research arm, touts in a blog post announcing Loihi. He talks about CCTV cameras that can run image recognition to identify missing persons or traffic lights that can track traffic flow to optimize timing and keep vehicles moving.
There’s still a long way to go before that happens though. According to Wired, so far Intel has only been working with prototypes, and the first full-size version of the chip won’t be built until November.
Once complete, it will feature 130,000 neurons and 130 million synaptic connections split between 128 computing cores. The device will be 1,000 times more energy-efficient than standard approaches, according to Mayberry, but more impressive are claims the chip will be capable of continuous learning.
Intel’s newly launched self-learning neuromorphic chip.
Normally deep learning works by training a neural network on giant datasets to create a model that can then be applied to new data. The Loihi chip will combine training and inference on the same chip, which will allow it to learn on the fly, constantly updating its models and adapting to changing circumstances without having to be deliberately re-trained.
A select group of universities and research institutions will be the first to get their hands on the new chip in the first half of 2018, but Mayberry said it could be years before it’s commercially available. Whether commercialization happens at all may largely depend on whether early adopters can get the hardware to solve any practically useful problems.
So far neuromorphic computing has struggled to gain traction outside the research community. IBM released a neuromorphic chip called TrueNorth in 2014, but the device has yet to showcase any commercially useful applications.
Lee Gomes summarizes the hurdles facing neuromorphic computing excellently in IEEE Spectrum. One is that deep learning can run on very simple, low-precision hardware that can be optimized to use very little power, which suggests complicated new architectures may struggle to find purchase.
It’s also not easy to transfer deep learning approaches developed on conventional chips over to neuromorphic hardware, and even Intel Labs chief scientist Narayan Srinivasa admitted to Forbes Loihi wouldn’t work well with some deep learning models.
Finally, there’s considerable competition in the quest to develop new computer architectures specialized for machine learning. GPU vendors Nvidia and AMD have pivoted to take advantage of this newfound market and companies like Google and Microsoft are developing their own in-house solutions.
Intel, for its part, isn’t putting all its eggs in one basket. Last year it bought two companies building chips for specialized machine learning—Movidius and Nervana—and this was followed up with the $15 billion purchase of self-driving car chip- and camera-maker Mobileye.
And while the jury is still out on neuromorphic computing, it makes sense for a company eager to position itself as the AI chipmaker of the future to have its fingers in as many pies as possible. There are a growing number of voices suggesting that despite its undoubted power, deep learning alone will not allow us to imbue machines with the kind of adaptable, general intelligence humans possess.
What new approaches will get us there are hard to predict, but it’s entirely possible they will only work on hardware that closely mimics the one device we already know is capable of supporting this kind of intelligence—the human brain.
Image Credit: Intel Continue reading
Quantum computers could give the machine learning algorithms at the heart of modern artificial intelligence a dramatic speed up, but how far off are we? An international group of researchers has outlined the barriers that still need to be overcome.
This year has seen a surge of interest in quantum computing, driven in part by Google’s announcement that it will demonstrate “quantum supremacy” by the end of 2017. That means solving a problem beyond the capabilities of normal computers, which the company predicts will take 49 qubits—the quantum computing equivalent of bits.
As impressive as such a feat would be, the demonstration is likely to be on an esoteric problem that stacks the odds heavily in the quantum processor’s favor, and getting quantum computers to carry out practically useful calculations will take a lot more work.
But these devices hold great promise for solving problems in fields as diverse as cryptography or weather forecasting. One application people are particularly excited about is whether they could be used to supercharge the machine learning algorithms already transforming the modern world.
The potential is summarized in a recent review paper in the journal Nature written by a group of experts from the emerging field of quantum machine learning.
“Classical machine learning methods such as deep neural networks frequently have the feature that they can both recognize statistical patterns in data and produce data that possess the same statistical patterns: they recognize the patterns that they produce,” they write.
“This observation suggests the following hope. If small quantum information processors can produce statistical patterns that are computationally difficult for a classical computer to produce, then perhaps they can also recognize patterns that are equally difficult to recognize classically.”
Because of the way quantum computers work—taking advantage of strange quantum mechanical effects like entanglement and superposition—algorithms running on them should in principle be able to solve problems much faster than the best known classical algorithms, a phenomenon known as quantum speedup.
Designing these algorithms is tricky work, but the authors of the review note that there has been significant progress in recent years. They highlight multiple quantum algorithms exhibiting quantum speedup that could act as subroutines, or building blocks, for quantum machine learning programs.
We still don’t have the hardware to implement these algorithms, but according to the researchers the challenge is a technical one and clear paths to overcoming them exist. More challenging, they say, are four fundamental conceptual problems that could limit the applicability of quantum machine learning.
The first two are the input and output problems. Quantum computers, unsurprisingly, deal with quantum data, but the majority of the problems humans want to solve relate to the classical world. Translating significant amounts of classical data into the quantum systems can take so much time it can cancel out the benefits of the faster processing speeds, and the same is true of reading out the solution at the end.
The input problem could be mitigated to some extent by the development of quantum random access memory (qRAM)—the equivalent to RAM in a conventional computer used to provide the machine with quick access to its working memory. A qRAM can be configured to store classical data but allow the quantum computers to access all that information simultaneously as a superposition, which is required for a variety of quantum algorithms. But the authors note this is still a considerable engineering challenge and may not be sustainable for big data problems.
Closely related to the input/output problem is the costing problem. At present, the authors say very little is known about how many gates—or operations—a quantum machine learning algorithm will require to solve a given problem when operated on real-world devices. It’s expected that on highly complex problems they will offer considerable improvements over classical computers, but it’s not clear how big problems have to be before this becomes apparent.
Finally, whether or when these advantages kick in may be hard to prove, something the authors call the benchmarking problem. Claiming that a quantum algorithm can outperform any classical machine learning approach requires extensive testing against these other techniques that may not be feasible.
They suggest that this could be sidestepped by lowering the standards quantum machine learning algorithms are currently held to. This makes sense, as it doesn’t really matter whether an algorithm is intrinsically faster than all possible classical ones, as long as it’s faster than all the existing ones.
Another way of avoiding some of these problems is to apply these techniques directly to quantum data, the actual states generated by quantum systems and processes. The authors say this is probably the most promising near-term application for quantum machine learning and has the added benefit that any insights can be fed back into the design of better hardware.
“This would enable a virtuous cycle of innovation similar to that which occurred in classical computing, wherein each generation of processors is then leveraged to design the next-generation processors,” they conclude.
Image Credit: archy13 / Shutterstock.com Continue reading