Research shows progress in reproducing neural activity inside artificial objects. But if it is true that research is advancing rapidly, there is still little talk about the consequences of the applications
Many, and rightly so, have admired what increasingly complex artificial intelligences are achieving.
But what if live neural networks and artificial neural networks were directly integrated with each other?
Let’s start thinking: Neurons are specialized cells, able to respond not only to electrical stimuli, but also to light, pressure, chemicals, and magnetic fields; For this reason, as they do in living organisms, they can be used to collect stimuli from the external environment and translate them into electrical impulses.
Moreover, researchers have already been shown to be able to keep complex biological networks of neurons alive for times long enough for them to be able to exert certain functions within artificial bodies: Living brain cells have already been used to make some robots process information And They navigate their environment based on external stimuli that neurons process.
Going further, a research team from the University of Illinois Urbana-Champaign puts on an amazing demonstration Work at the American Physical Society meeting held this month in Las Vegas.
The researchers transplanted about 80,000 neurons derived from reprogrammed mouse stem cells onto plates. For comparison, consider that the brain of an adult fruit fly contains about 200,000 neurons, and the human brain has over 86 billion.
Cultivation of the platelet neurons made it possible to obtain a two-dimensional biomechanical network, which was placed under an optical fiber and on an electrode grid; In this way, it was possible to stimulate the network itself with mixed sequences of light pulses and electrical signalsand then record the electrical signals produced by the neural network in response to the electrode network.
Each device was placed in a palm-sized box, as can be seen in the figure accompanying this text, which was in turn placed in an incubator to keep the cells alive. The electrical signals produced by the neural network were sent to an ordinary computer chip and fed into a neural network, which was used to recognize the specific electrical patterns produced by the biological network.
At this point, the entire apparatus was ready to answer the first essential question: Once a specific stimulus using light and electricity is sent to a biological neuron network, the electrical response will be specificthat is, will the same response always be elicited by presenting the same stimulus?
The researchers then created 10 distinct sequences of electrical impulses and flashes of light, each played multiple times for one hour.
After this “training” for an hour, the first significant result was found: The neurons produced the same signals each time the same pattern was presented.
The chip, which runs the artificial neural network, just had to learn to distinguish between those signals, classifying them as 10 different types – that is, grouping them according to their similarity. Artificial neural networks often take a great deal of time and many iterations to train, but the division of labor between biological neurons, which are able to generate a precise electrical stimulus in response to environmental conditions they “saw,” and the artificial neurons allowed the researchers to train in a big way. Reducing the time and energy required for training.
Not only that: At the end of the training hour, the researchers let the neurons rest; They were then re-exposed to each of the 10 series of light and electricity.
Good, The biological network retained the memory, reproducing the same electrical patterns produced during training and thus feeding the artificial neural network used for its original recognition. In particular, to assess how well the device works, they calculated a score commonly used to evaluate AI recall and classification capabilities, ranging from 0 (the worst possible case) to 1 (the most efficient), to obtain a hybrid device designed score of 0.98.
At the moment, the device cannot compete with traditional neural networks, but here I would like to make a number of points.
First of all, the device will make it possible to assess whether the sensory memory of biological brains is really consistent in an architectural change of synaptic connections between different neurons, and to study this parameter in response to stimulation and in relation to the increase in the discriminatory abilities of external stimuli and the continuity of this ability over time.
Second, the system is scalable: Many plates can also be biologically attached to each other, allowing neurons in one culture to communicate with those in another primary culture., resulting in complex three-dimensional architectures, comparable to those of cerebral organoids and the brains themselves. This will most likely allow to achieve performance that is still difficult to assess today.
Still, Here we have a further demonstration of the pathway that leads to the creation of biomechanical brains Which, remotely and in an unlimited way, can control machines of all kinds, receive stimuli from the environment in which these machines will operate and control the response to them.
In addition, the significant savings in terms of data, time and energy needed to train networks of neurons compared to current artificial intelligence, which was also demonstrated in these first and simple experiments, opens the door to making a leap forward in capabilities that we can call “cognitive” artificial systems. , if not necessarily with the same computational capabilities, depend on different inference methods than those used to date.
Finally, and this seems to me to be the most important element of all, we may be closer to tackling the problem of the organic/physical source of mind empirically, with all the moral, philosophical, and scientific consequences of the case.
The search is going pretty quickly, however Social awareness of what is being done and its possible applications, for better or worse, I think is still very thinly spread; It seems to me that this is an example of the gap between science and sharing that I was discussing yesterday on this page.
“Infuriatingly humble alcohol fanatic. Unapologetic beer practitioner. Analyst.”