February 16, 2016

As machines get smarter, evidence grows that they learn like us

It's a long read (5 pages) but if you are interested in the history and future of neural networks and machine intelligence it's most definitely worth your time. https://www.simonsfoundation.org/quanta/20130723-as-machines-get-smarter-evidence-they-learn-like-us/ - You can find a collection of interesting bits from the article below but you might want to check it out in full.

Studies suggest that computer models called neural networks may learn to recognize patterns in data using the same algorithms as the human brain.

One of the most promising of these algorithms, the Boltzmann machine, bears the name of 19th century Austrian physicist Ludwig Boltzmann, who developed the branch of physics dealing with large numbers of particles, known as statistical mechanics. Boltzmann discovered an equation giving the probability of a gas of molecules having a particular energy when it reaches equilibrium. Replace molecules with neurons, and the Boltzmann machine, as it fires, converges on exactly the same equation.

Each virtual synapse tracks both sets of statistics. If the neurons it connects fire in close sequence more frequently when driven by data than when they are firing randomly, the weight of the synapse is increased by an amount proportional to the difference. But if two neurons more often fire together during random firing than data-driven firing, the synapse connecting them is too thick and consequently is weakened.

Neural networks have recently hit their stride thanks to Hinton’s layer-by-layer training regimen, the use of high-speed computer chips called graphical processing units, and an explosive rise in the number of images and recorded speech available to be used for training. The networks can now correctly recognize about 88 percent of the words spoken in normal, human, English-language conversations, compared with about 96 percent for an average human listener. They can identify cars and thousands of other objects in images with similar accuracy and in the past three years have come to dominate machine learning competitions.

Adult brains are less malleable than juvenile ones, much as a Boltzmann machine trained with 100,000 car images won’t change much upon seeing another: Its synapses already have the correct weights to categorize a car. And yet, learning never ends. New information can still be integrated into the structure of both brains and Boltzmann machines.

studies of brain activity during sleep have provided some of the first direct evidence that the brain employs a Boltzmann-like learning algorithm in order to integrate new information and memories into its structure. Neuroscientists have long known that sleep plays an important role in memory consolidation, helping to integrate newly learned information. In 1995, Hinton and colleagues proposed that sleep serves the same function as the baseline component of the algorithm, the rate of neural activity in the absence of input.

The easiest way for the brain to run the Boltzmann algorithm, he said, is to switch from beefing synapses up during the day to whittling them down during the night. Giulio Tononi, head of the Center for Sleep and Consciousness at the University of Wisconsin-Madison, has found that gene expression inside synapses changes in a way that supports this hypothesis: Genes involved in synaptic growth are more active during the day, and those involved in synaptic pruning are more active during sleep.

A Boltzmann-like algorithm may be only one of many that the brain employs to tweak its synapses. In the 1990s, several independent groups developed a theoretical model of how the visual system efficiently encodes the flood of information striking the retina. The theory held that a process similar to image compression called “sparse coding” took place in the lowest layers of the visual cortex, making later stages of the visual system more efficient.

The model’s predictions are gradually passing more and more stringent experimental tests. In a paper published in PLOS Computational Biology in May, computational neuroscientists in the United Kingdom and Australia found that when neural networks using an algorithm for sparse coding called Products of Experts, invented by Hinton in 2002, are exposed to the same abnormal visual data as live cats (for example, the cats and neural networks both see only striped images), their neurons develop almost exactly the same abnormalities.

The human brain, of course, remains much more complicated than any of the models; it is larger, denser, more efficient, more interconnected, has more complex neurons — and juggles several algorithms simultaneously. Olshausen has estimated that we understand only 15 percent of the activity in the visual cortex. Although the models are making progress, neuroscience is still “a bit like physics before Newton,” he said. Still, he is confident that the process of building on these algorithms may one day explain the ultimate riddle of the brain — how sensory data gets transformed into a subjective awareness of reality.

No comments:

Post a Comment