Do low neural networks ‘see’ like we and we do?

Computer components designed to commend images act surprisingly like neurons in a brain, researchers during a University of Washington School of Medicine report.

“We found that low neural networks used for noticing images indeed have singular units within them that respond in remarkably identical ways as do neurons within a visible complement of a brain,” pronounced lead author Dean A Pospisil, a neuroscience doctoral claimant in a laboratory of Wyeth Bair.

Like a V4 neuron, a synthetic neuron responded many greatly to objects with a pointy points (top row) and slightest to dull objects (bottom row). Image credit: Dean Pospisil / University of Washington Department of Biological Structure.

Bair and Anitha Pasupathy, both UW associate professors of biological structure, co-authored the paper with Pospisil. It seemed in a journal eLife.

Artificial (computer-based) neural networks are done adult of layers of “neurons” that accept submit signals from higher-level neurons, routine those inputs regulating mathematical algorithms, and send their outputs on to a neurons in a subsequent level, that repeat a process.

If a network’s final outlay is scold — or proceed a scold answer — feedback is sent behind by a complement to strengthen a algorithms that gave a scold response and stop those that were incorrect.

Although these networks are distant reduction formidable than a brain, they can be lerned to solve a accumulation of questions, such as how to win during chess. But only how these networks “learn” and how closely they impersonate a problem-solving processes of a mind is unclear.

In their study, a UW researchers complicated an synthetic neural network that is modeled on a structure in a mind called the ventral visible stream.  In a brain, neurons within this structure routine signals from a eye. As these signals pierce from neuron to neuron by a ventral visible stream, sold neurons respond to gradually some-more formidable elements of an image. First they respond to rags of dim and light, and afterwards after to elements such as edges and shapes, until finally an intent in an picture is categorized, such as being that of a cat or a car. That information is sent on for estimate in other mind areas.

In their study, a researchers focused on a sold mark in a ventral visible tide called V4.  Neurons in this area specialize in noticing a bounds of objects. Many are privately tuned to respond to bounds that have a bend and are oriented in a sold direction.

To know either nodes in a computerized neural network behaved similarly to neurons formerly available in area V4 in a macaque’s brain, a researchers presented accurately a same visible stimuli from a macaque experiments to singular nodes in a network.

They found that sold nodes indeed behaved like sold V4 neurons, responding to a same specific shapes while ignoring others.

“It seemed like a function of singular units of a neural network were concentration on a function we saw in singular neurons in V4,” Pospisil said.

For neuroscientists who study the brain, a commentary advise that synthetic neural networks might be useful models to understand how a mind works, he added, and, for mechanism scientists, insight into how neural networks solve problems.

Source: University of Washington


Comment this news or article