Posts Tagged ‘neural networks’
Video: A Robot That Can Figure Out New Tasks Based On the Ones It Knows
In other words, the robot learns as it goes. This ability is imparted by an algorithmic technology called the Self-Organizing Incremental Neural Network (SOINN) developed by the researchers to give their ‘bot more mental dexterity. To borrow an example from the video below, the robot can fill up a glass of water from a bottle via pre-programmed instructions. But if halfway through the task its overseer asks it to chill the water, the robot will actually stop and think about the next steps.
Figuring that it can’t grab an ice cube until it empties one of its two hands, it then reasons that the water bottle is more expendable than the glass of water and sets the bottle down. It then grabs the ice and drops the cube into the glass. Task completed, no extra programming necessary.
All that might seem anticlimactic, but it’s it’s pretty impressive from an AI perspective. The ability to adapt to new situations and learn from non-programmed past experiences is one that robotic systems sorely lack. Coupled with the ability to go online and ask other robots--robots that are also learning on the fly--how to perform a certain task, a global network of such robots could quickly assemble a vast body of know-how from which to draw, making every robot in the network smarter.
[]
With Electrical Stimulation to the Spinal Cord, Paralyzed Man Walks Again
The research team, led by Dr. Susan Harkema of the University of Louisville, Ky., stressed that the treatment is not a cure for paralysis and that it worked with just one patient in one trial. But researchers not involved in the study say it is promising — one UK doctor it was “mind-blowing.”
The findings appear to show that the legs and spinal cord, not the brain, are in control of movement. That means interruption of messages from the brain may not preclude paralyzed patients from walking again — they would just need new electrical signals to stimulate the spinal cord.
Summers appeared in various media outlets Friday to discuss the research.
Weeks after winning the College World Series with Oregon State University in 2006, Summers was hit by a drunk driver, suffering spinal cord damage that paralyzed him from the chest down. Neuroscientists implanted 16 electrodes in his spine, and sent electrical impulses to his lower spinal cord, mimicking the signals normally sent by the brain to initiate movement. Summers was suspended over a treadmill while the signals were transmitted to his spine. Writing in the British medical journal , researchers say the spinal cord’s own neural network, combined with sensory information from his legs, is able to to control muscle and joint movement.
Summers trained for two years with a treadmill and physical therapists moving his legs to help him stand and walk.
V. Reggie Edgerton of the David Geffen School of Medicine at UCLA said sensory information is sent via neural networks in the legs directly to the spinal cord. The sensory feedback allows Summers to balance himself, bear his own weight and take steps over various speeds and directions, Edgerton said in a .
In a statement, Summers said the treatment has changed his life.
“For someone who for four years was unable to even move a toe, to have the freedom and ability to stand on my own is the most amazing feeling,” he said.
He was left with some sensation below the chest, so it’s not clear whether the treatment would work for spinal cord injury patients who experience no sensation. What’s more, Summers was an athlete in excellent physical condition before his injury, which could have helped his rehabilitation.
Still, his doctors hope that someday, patients with spinal cord injuries could use a portable electrical stimulation unit to move independently once again.
The work was funded by the National Institutes of Health and the Christopher & Dana Reeve Foundation.
[via ]
Computer Scientists Induce Schizophrenia in a Neural Network, Causing it to Make Ridiculous Claims
Computer scientists at the University of Texas-Austin built a neural network called DISCERN, which is able to learn natural language. The humans taught it a series of simple stories, teaching it to store information as relationships between words and sentences — much the same way a person would learn a story.
Then they started again, but cranked up DISCERN’s rate of learning — so it was assimilating words at a faster rate, and it was not ignoring as much noise in the data.
Some mental health experts believe schizophrenics cannot forget or ignore as much stimuli as they should, which makes it difficult to synthesize and process the correct information. This “hyperlearning” phenomenon causes schizophrenics to lose connections among individual stories, losing the distinction between reality and illusion. Dopamine is a key factor in the process of understanding and differentiating experiences.
Telling the computer to “forget less” was akin to flooding the system with dopamine, confounding its ability to discern relationships between words, sentences and events, according to a .
“DISCERN began putting itself at the center of fantastical, delusional stories that incorporated elements from other stories it had been told to recall,” according to the news release. In one answer, it claimed responsibility for a terrorist bombing.
The experiment doesn’t prove the hyperlearning hypothesis, but it does lend it additional credence, according to the researchers, who published their crazed computer findings in the journal Biological Psychiatry. It also shows that neural networks can be a useful analogue for the information-processing centers of the brain, according to graduate student Uli Grasemann, who participated in the research.
“We have so much more control over neural networks than we could ever have over human subjects,” he said. “The hope is that this kind of modeling will help clinical research.”
[via ]
Fruit Flies’ Neural Networks Solve Distributed Computing Problem Better Than Humans
It’s not the first time we’ve seen an insect solve a problem that plagues computer scientists — — but the fruit fly discovery does one better, leading to an algorithm that can be used to develop more efficient computer and wireless networks.
Distributed computing involves several processors working in concert to solve a problem. Some are chosen as leaders, collecting data from the other processors and passing it along. Organizing these networks into efficient processor-leader groups is one of the biggest challenges in computing — but millions of cells in a fly’s nervous system do it automatically, organizing themselves so that a small number of cells serve as leaders. It is much better than anything humans have come up with, scientists say: “It is such a simple and intuitive solution, I can’t believe we did not think of this 25 years ago,” according to co-author Noga Alon, a mathematician and computer scientist at Tel Aviv University and the Institute for Advanced Study in Princeton.
Fruit fly bristles, which are used for feeling and hearing, develop as nerve cells self-select to become leaders. The cells send chemical signals to their neighboring cells, ensuring that those cells cannot become leaders, too. Using fluorescence microscopy, the researchers watched an entire network form in about three hours.
They developed an algorithm based on the cells’ self-selection approach, and say it’s particularly effective for adaptive networks where the number and position of each node is not certain, according to . That could include environmental monitoring sensors, robot swarms and more.
The research is published today in the journal Science.
[]
IBM Researchers Create the Most Detailed Brain Map Yet
A significant stride towards reverse-engineering the darn thing
Focusing on a long-distance network connecting 383 brain regions and 6,602 long-distance connections that function like highways to connect disparate regions of the brain. Shorter, more localized connections were found to carry signals within regions.
But most importantly, they found what they describe in a paper published in PNAS as a "tightly integrated core" that might be they key to cognition in higher-thinking biological creatures. That core might be what gives us consciousness (we won't get into the philosophical implications there). Further, the core isn't located in one, or even two regions. The researchers found it stretches through the premotor cortex, prefrontal cortex, temporal lobe, thalamus, visual cortex and a handful of other regions.
Another surprising find: the prefrontal cortex, though at the front of the brain, might actually serve as its central information hub that distributes information throughout the brain.
The study included mapping of four times as many regions and three times the number of connections than the largest previous attempt. Those findings could finally help researchers mimic the brain -- which, even in this seemingly advanced era, is something of a mystery to us. That in turn could lead to network architecture and computer chips that process and move information as quickly and seamlessly as our brains do.
[]
Neural Network Simulator Models Blood Platelet Response During Heart Attack
We often think of our blood as specifically tasked with carrying oxygen to our brains and other organs, but it's also a living fluid, changing up its duties in response to various stimuli. To better understand -- and anticipate -- one aspect of this complicated biology, researchers have trained a neural network computer to in the blood react to complicated conditions like those experienced during heart attack or stroke.
The University of Pennsylvania's robotic automated system works by measuring the reaction of platelets to agonists -- chemicals that bind to platelets to initiate a cellular response -- introduced in pairs. The process tags each platelet with 34,000 data points acquired during those evaluations, with each duo of agonists leaving its unique fingerprint.
By cross-referencing the responses to the different pairs of agonists, the system can extrapolate the intracellular signaling responses of the platelets to far more complex combinations of up to six different agonists at a time, even at varying dosages and in situations where the agonists are applied at different times.
In tests, the model predicted platelet responses with a high degree of accuracy, even differentiating between several different donors, opening the door to complex, patient-specific platelet modeling that could be critically helpful during a heart attack or stroke. Researchers hope to refine the system such that it is computationally fast enough to calculate and model a patient's complex response to agonists present during a heart attack or stroke even as the event is unfolding.
Even better, the system could lead to a powerful diagnostic tool that predicts cardiovascular disease before a patient ends up in the emergency room.
[]
Neural Networks Designed to ‘See’ are Quite Good at ‘Hearing’ As Well
The network is composed of three "thinking" layers stacked one atop the other, with the first taking in the raw data and the third outputting a genre. Drawing from a database that spanned 10 musical genres, the machine went to work. Within each layer, each neuron only hears a snippet of the song about 23 milliseconds long. But each node overlaps the one next to it by half, so in total each node really gets to hear about two seconds worth of audio.
The algorithms employed by the network needed only that amount of time to process and identify the genres of songs from the database. However, when turned loose on songs not included in the library that it learned on, it didn't perform well at all. Which tells us a few things.
For one, to work universally the network needs to be trained on a more universally representative library, as there are more than 10 genres in the entire universe of music. But more importantly, as points out, this neural network shows that a device designed for one function -- this particular neural network was inspired by the visual cortex of a cat -- can be re-wired to do something different (in this case, to hear).
Similar networks based on auditory cortexes have been rewired for vision, so it would appear these kinds of neural networks are quite flexible in their functions. As such, it seems they could potentially be applied to all sorts of perceptual tasks in artificial intelligence systems, the possibilities of which have only begun to be explored.
[]