Artificial intelligence: small, but it's a success
From complex to complex: a new model of artificial brain has been created
As a result, their brainchild, called Spaun (Semantic Pointer Architecture Unified Network), performs a number of different types of cognitive functions. This artificial brain contains certain properties of neural transmission, reflects the functional organization of some real parts of the brain, and as a result, the artificial brain demonstrates a complex and flexible behavior of the whole object. And this is the first working artificial intelligence with similar abilities.
Any living being with a brain demonstrates flexible and complex behavior consisting of a whole array of elementary acts. Each of these acts requires the inclusion of different brain functions – from pattern recognition to information analysis, from performing a specific action to assessing its consequences. Imitation of individual neurons performing these elementary functions, combining them into a single working system is the ultimate task of building artificial intelligence. The complexity of this super task is generated by an incredibly large number of neurons in the real brain and a fundamental misunderstanding of how the interaction of different blocks of information processing is arranged. The creators of artificial intelligence and their colleagues in neurophysiology face a painful question: is it enough to simply increase the number of neurons to reconstruct complex behavior? And if it's enough, then for how long? And if not enough, what determines behavioral flexibility and complexity? So far, as everyone knows, there is no answer to these questions, otherwise we would have been discussing the advantages and disadvantages of a reasonable robot for a long time. And we are still at the stage of anecdotal conversations with computer Alice (the best of the "virtual interlocutor" programs, which is still very far from passing the Turing – VM test).
But it's not so hopeless: artificial intelligence specialists are creating new models with might and main. They already have a brain model with 1 million neurons in their arsenal, in which an adequate spatial structure of the laying of neurons, the properties of their interaction, the parameters of pulse transmission and some others are recreated: Henry Markram, 2006. The Blue Brain Project. (See also the article "A billion for brain powder?" – VM.)
There is also a more cumbersome model with 1.6 billion neurons, comparable in number of neurons and synapses to the cat brain (Rajagopal Ananthanarayanan et al., 2009. The cat is out of the bag: cortical simulations with 109 neurons, 1013 synapses, PDF, 2.07 MB). An even more impressive model has also been built, in which the interaction of 100 billion neurons is programmed (E.M. Izhikevich, G.M. Edelman. Large-scale model of mammalian thalamocortical systems). In all these models, emphasis is placed on the number of conducting paths, on the number and parameters of connections between them. An indicator of realism is, as a rule, similarity with the electrophysiological characteristics of the whole brain and its parts or the performance of any one elementary function.
Fig. 1.This is what a supercomputer looks like, servicing one of the models of an artificial brain, in which 100 million neurons are involved; this is approximately the same as the mouse brain. Now a new model has been created, it is 50 times smaller in volume, and the connections between the elements are fundamentally different. Image from the website plus.google.com
None of the models proposed to date can reproduce the diversity of responses and functions. But a new model could do it, not so impressive in terms of the number of neurons, but it was arranged in a new way. It was named Spaun; this name is an abbreviation of Semantic Pointer Architecture Unified Network, which roughly means "unified network of organized semantic pointer". That is, we mean a system capable of converting a symbol into an object and vice versa, and using this ability for different actions. Spaun demonstrates how a complexly organized brain generates complex behavior. The merit of its creation belongs to specialists from the Center for Theoretical Neuroscience at the University of Waterloo (Canada).
Spaun consists of two parts. The first is actually an artificial brain, which has an eye (camera). The second is an artificial arm connected to the brain. Thus, the whole system is able to perceive, analyze and act: the eye sees, the brain thinks and commands, the hand executes. Together they are able to perform eight different tasks. At the same time, the designer himself should not interfere with the "thoughts" of his artificial brainchild. He cannot tell the machine what is required of it, in other words, he cannot solder contacts or enter additional codes. The brain is told what task it should perform at the moment, and it must choose the way to solve it.
The set of tasks – it is not trivial at all for a machine – is proposed as follows (see video):
- Drawing: recognize an object and, observing the style, draw a drawing with your hand. There are only 10 objects, these are numbers from 0 to 9.
- Pattern recognition: recognize a handwritten digit and represent it in a predefined format (Fig. 2).
- Choosing the best: out of three possibilities, choose the one for which the reward is due. The reward varies randomly from experience to experience.
- Memorization: Reproduce the displayed series of numbers.
- Addition: Calculate the sum of two numbers and write it.
- Answers to questions about spatial arrangement. Here Spaun has to answer one of two questions about the order (place) of the digit in the depicted row: either which digit is in a certain place, or which place is occupied by a certain digit.
- Semantic recognition, for example, from 0024 to make 24.
- Creative analysis: you need to solve one of the simple tasks from the IQ test: continue the series by analogy, for example, add a series 123, 567, 23?.
Spaun performs these tasks in any order, and, I repeat, the creators of the model do not interfere, do not prompt, but only examine their creation.
Fig. 2. Task #2: what the "eye" of the artificial brain named Spaun sees and what his hand draws. Image from the website nengo.ca
When performing such tasks, a number of cognitive functions are simultaneously involved. It was them that the neuroconstructors tried to model. For example, to complete task 8 – let it be such a test (1, 11, 111, 2, 22, ?) – you need to start by recognizing the written numbers. Then it is necessary to encode the received information, reduce its volume by 10-15 times, as it happens in real visual centers, and send it to the working memory block. In the working memory block, the information is compared with the existing one, and as a result it is possible to distinguish the order of the images, for example, to divide the options 12 and 21. Then the relationships between the semantic units are established, that is, the similarity and difference in the ordering of all images is evaluated, in our example it is 1 and 11, 11 and 111, 2 and 22. Next, the area of averaged relationships is displayed. Taking into account the categories of the greatest similarity, the solution is determined – 222. Then the information enters the decoding and expansion block of the information. It is called a motor unit, since signals go from it to the hand, and it draws, in accordance with its weight and inertia, the desired figure.
The model includes 2.5 million neurons. The properties of neural transmissions and their conductivity were copied by the creators of Spaun from real nerve cells of a real mammalian brain. The model has analogues of dopamine and GABA receptors with their specific parameters of pulse transmission.
Spaun neurons are grouped into separate blocks (Fig. 3).
Fig. 3. The functional blocks of the artificial brain Spaun are based on real brain areas and the connections between them.
The blocks mimic the work of specialized departments of the cortex and each is responsible for a specific function: memorization, encoding and compression of information, etc. So, the first block deals with the perception of visual stimuli, this is an analogue of the visual cortex. His task is to recognize the image, isolate meaningful information from it, separating it from all the accompanying information, and then encode neural impulses understandable to the brain. Then the encoded information goes to the working memory department, where it is compressed even more and stored in this form. The reward block selects the final action from several possible options. This is followed by the obviously necessary block of decoding neural information into motor commands of the hand. Thus, the blocks are not at all designed to solve any specific tasks, for example, adding numbers or recognizing pictures. On the contrary, in the model everything is arranged in such a way that any (!) task based on visual perception is solved in principle.
A is a diagram of the human brain, which shows the areas used to create Spaun. The names of the regions are circled and filled with the colors that represent the corresponding blocks in diagram B. V1 – striar (primary visual) cortex, V2 and V4 – extrastriar cortex, IT – inferior temporal cortex, AIT – anterior inferior temporal cortex, VLPFC and DLPFC – ventrolateral and dorsolateral prefrontal cortex, OFC – orbitofrontal cortex, PPC – posterior parietal cortex; Str – striatum (D1 and D2 denote different dopamine inputs), STN – subthalamic nucleus, VTA – ventral region of the tire, GPe and GPi – outer and inner parts of the pale globe, SNc and SNr – compact and reticular parts of the substantia nigra (see Substantia nigra). M1 is the primary motor cortex, SMA is the additional motor region, PM is the premotor cortex. Within the orange area, lines with circles show GABAergic (inhibitory) connections, and with squares – dopaminergic (modulating).
B – functional architecture of Spaun. Thick lines show connections between cortical regions; thin lines show connections between the action selection unit (basal ganglia) and the cortex. Squares with rounded corners show that when an action is selected, the input of information to certain blocks changes. A small square on the line between the reward evaluation block and the action selection block shows that this relationship modulates the significance of the action.
Image from the discussed article in Science
For each specific task, you can conduct a series of experiments and evaluate how Spaun copes with the tasks. For example, let's analyze the memorization task: recall and write down a number of numbers. In humans (what artificial intelligence specialists call a biological brain), the accuracy of the reproduction of a series depends on its length, the first and last elements of the series are best remembered. In the experiment, Spaun reproduced a series of four, five, six and seven digits over and over again. Surprisingly, the artificial subject also demonstrated the best reproduction of the first and last digits (Fig. 4).
Fig. 4. The accuracy of reproduction of a series of four, five, six and seven digits by people (A) and the Spaun machine (B). It is clearly seen that in both cases the average values of the series are reproduced worse than the final ones. Image from the discussed article in Science
The creators of Spaun emphasize that their task was not to analyze the execution of individual tests. They intended to create such a machine that could act according to the tasks presented, and at the same time the tasks can be of different types. And they created such a machine. As a result, the machine simulates both the properties of individual cells – and they are originally embedded in the system as parameters for transmitting impulses, and the properties of the behavior of an integral object – we saw this on the example of a test with the reproduction of a series. Thus, with the help of the coordination device of individual neural elements, it was possible to simulate the complex and diverse behavior of the whole. And this is an important step towards the creation of artificial intelligence.
Source: Eliasmith et al., A Large-Scale Model of the Functioning Brain // Science. 30 November 2012. V. 338. P. 1202-1205.
Portal "Eternal youth" http://vechnayamolodost.ru06.12.2012