Some Artificial Neural Networks Learn Language Like Humans

A UC Berkeley study just showed that some artificial neural networks can learn language just like humans, which raises suspicions about AI breakthrough.

“There has been a long debate about whether neural networks learn the same way humans do,” said Vsevolod Kapatsinsky, a linguist at the University of Oregon.

A study published last month suggests that natural and artificial networks learn the same way, at least when it comes to language. Researchers led by Gasper Begush, a computational linguist at the University of California, Berkeley, compared the brainwaves of people listening to a simple sound with the signal produced by a neural network analyzing the same sound. The results were strikingly similar. “To our knowledge,” Begush and colleagues wrote, “observed responses to the same stimulus are the most similar brain and ANN signals reported so far.”

Most importantly, the researchers tested networks made up of general-purpose neurons that are suitable for a variety of tasks. “They show that even very, very general networks that don’t have any sort of evolved bias towards speech or any other sound show a match for human neural coding,” said Gary Lupyan, a psychologist at the University of Wisconsin-Madison. did not participate in the work. The results not only help to demystify how ANNs learn, but also suggest that the human brain may not yet be equipped with specially designed language hardware and software.

To establish a baseline for the human side of the comparison, the researchers played the same syllable “ba” multiple times over two eight-minute blocks in front of 14 English and 15 Spanish speakers. During playback, the researchers recorded fluctuations in the average electrical activity of neurons in each listener’s brainstem, the part of the brain where sounds are first processed.

In addition, the researchers loaded the same “ba” sounds into two different sets of neural networks, one trained on English sounds and the other trained on Spanish. The researchers then recorded neural network processing activity, focusing on artificial neurons in the network layer where sounds are first analyzed (to reflect readings from the brainstem). It was these signals that very accurately corresponded to the waves of the human brain.

The researchers chose a type of neural network architecture known as a generative adversarial network (GAN), originally invented in 2014 for imaging.

In this study, the discriminator was initially trained on the sound set of English or Spanish. Then the generator, which had never heard these sounds, had to find a way to reproduce them. At first, it made random sounds, but after about 40,000 cycles of interaction with the discriminator, the generator got better, finally producing the correct sounds. As a result of such training, the discriminator also improved in distinguishing between the real and the generated.

It was at this point, after the discriminator was fully trained, that the researchers reproduced the sounds “ba” on it. The team measured fluctuations in the average activity of artificial discriminator neurons that generated a signal so similar to human brain waves.

The experiment also revealed another interesting parallel between humans and machines. Brainwaves showed that English and Hispanic participants heard the sound “ba” differently (Spanish speakers heard more than one “pa”), and GAN signals also showed that the English-trained network processed sounds slightly differently than this. trained in Spanish.

“Now we are trying to understand how far we can go, how close to human language we can get with the help of general purpose neurons,” Begush said. “Can we achieve human-level performance with the computing architectures we have just by making our systems bigger and more powerful, or will that never be possible?” While more work is needed before we can know for sure, he said, “Even at this relatively early stage, we are surprised at how similar the inner workings of these systems, humans and ANNs seem to be.”