Vivchaєmo neuron frizzy: why to start. Pombs in the book by Simon Haykin "Neural framing: a new course"


We described the simplest power of formal neurons. We talked about the fact that the boundary summer is more precise about the nature of a single spike, and the line summer allows the modeling of the neuron to be formed from a series of pulses. It was shown that the value of the line summer can be set at the frequency of the adhesions of a real neuron. Now we are waking up on the basis of power, such as the volodymy of such formal neurons.

Filter Hebb

Far away we will often turn to neuromechanical models. In principle, practically all the basic concepts from the theory of neural framing can be directly related to the real brain. Lyudin, stinging with singing staff, having come up with helpless tsikavi neuromerezhevy constructions. Evolution, going through all the possible neuron mechanisms, saw everything that appeared for her to be corny. It’s not surprising that for even more sophisticated models invented by people, you can know readable biological prototypes. Oskilki our decision not to put in place of the meta I want to be a detailed wiklade of the theory of neural framing, but we will only have to deal with the most remote moments that are necessary to describe the basic ideas. For a great deal of intelligence, I highly recommend turning to a special literature. My best friend on neural fences is Simon Haykin “Neural fences. Secondary course ”(Haykin, 2006).

At the heart of the neuromechanical models are the good rule of Hebb's rule. It was propounded by the physiologist Donald Hebb in 1949 (Hebb, 1949). In three vilny interpretations, there is even a simpler sensation: neuron connections, which are activated spirally, to blame for the activation, neurons' connection, which is just right, blame for weakness.
The output of the line summer can be written:

As soon as we use small numbers of cobs for input, we never ask us to try a neuron based on the Hebb rule:

de n- discrete croc per hour, - parameter of efficiency.

With such a procedure, we will quietly enter, on which a signal is given, it is more sensitive than the active reaction of the neuron itself. If there is no reaction, then it will not be seen and started.

True, such a vagi will not be intertwined with growth, so for stabilization it is possible to fix the norm. For example, podіliti on dozhinu vector, cut off from the "new" synaptic vag.

With this, it is possible to overwhelm the vag with synapses. The intelligence is the essence of a lighter change, like quilting behind the snake vag in two priyomi. With the spatula, if the neuron is active, at the synapse, at which a signal comes up, take the supplement. Vagi synapses without a signal zalishayutsya without changes. Then the normalization is changed to change the length of all synapses. If there is a whole lot of synapses without signals, they spend at the same time, but synapses with signals overwhelm them.

Hebb's Rule is not really inshe, as the implementation of the method of gradient descent over the surface of the grave. Along the way, the neuron grows out of the way before the signals are sent, destroying the skin once it is in the bik, the anti-gratification, so to speak of the antigradient. The gradual descent has grafted us to the local extremum, but not having slipped through it, the speed of the descent is to blame, but it is too small. Scho in Hebbovskiy navchannі vrahovuєtsya as the child of the parameter.

Crumbs of the speed parameter allow you to rewrite the front formula in the viewer in the following order:

If you see a different order of things, then you see the rule of Oja (Oja, 1982):

The positive addition is the positive for Hebbovskoe navchannya, and the negative for the out-of-home stability. Recording in such a view allows you to see it, as it can also be realized in the analogue middle without the calculation, operating only with positive and negative links.

So the axis, so in the edge of it is simple to create power. If we will step by step change the speed of the beginning, then the synapses of the neuron will go down to such values ​​that the first step starts to show the first head component, as the component went out, as if the components were caught before they were given the procedure. This design is called Hebb's filter.

For example, a pixelated picture is given to the input of the neuron, so that one point of the image can be adjusted to the skin synapse of the neuron. We will feed all two images to the input of the neuron - images of vertical and horizontal lines, which pass through the center. One croc navchannya is one image, one line, either horizontal or vertical. If the image is averaged, then you will see the chrest. Ale, the result of navchannya will not be similar to averaged. There will be one z line. The one that will often be seen in the images presented. The neuron sees not averaging or repetitiveness, but those points that are most often developed spirally. If the image is foldable, the result may not be the same. Ale tse will be the same component.

Starting a neuron leads to the fact that a singing image is seen (filtered) on its teresa. If a new signal is given, then the more accurate the signal and the adjustment of the vagina, the more accurate the neuron. A neuron can be called a detector neuron. At the same time, an image that can be described by a yogo wagami is accepted as a characteristic stimulus.

Head components

The very idea of ​​the head component method is simple and genetic. It is permissible that we have the last of the stories. The skin of them is described through the flow of water onto the receptor, which is just as light. Admittedly, we have sensors to describe a sign. All podії for us is described by vectors of dimension. The skin component of such a vector is to be applied to the meaning of the specific sign. All at once the stench is set to a vipadkov value X ... We can draw points at the viewer in the virtual space, and the axes appear to help us to perceive the signs.

The averaged value is even mathematically related to the drop value X, To mean, yak E ( X). We will center the data so that E ( X) = 0, then the dots will be concentrated around the cob of coordinates.

Qia ghmara can wake up in a tight spot. Having tried every possible straightforward, we can also know that the variance of the given will be maximal.

So the axis, such a straight line, is the first head component. The head component itself is defined as a single vector, which goes from the cob of coordinates and goes straight from the cob.

Far away we can know the largest perpendicular to the first component, as well as the bridging dispersion, as well as the maximum average of all perpendicular directions. Knowing yogo, we can accept another component. Then we can continue the shudder, wondering how shukati the demand is in the middle of the straight lines that are perpendicular to the already known components. As long as the coordinates are linearly independent, then it is possible to enter once, until the end of the vastness of space. In such a rank, we recognize the mutually orthogonal components, ordered according to the amount of variance given by the stink to explain.

Naturally, it is natural that the main components visualize the internal laws of our gifts. Ale є more simple characteristics, also describe the essence of the obvious laws.

Let's admit that for everything we have n go. Skin pod_ya is described by a vector. Vector component:

For dermal signs, you can write down how you see it in your dermal pod:

For be-like two signs, on which there will be a description, it is possible to have a size, as to show the steps of their spiritual manifestation. The quantity is called covariates:

Vaughn show, some of the views from the middle value of one of the signs are displayed on the basis of similar views of the signs. If the mean value signifies zero, then the covariance is typed in a glance:

As soon as the covariance on the root-mean-square outputs is scored, the power of signs, then we recognize the linear coefficient of correlation, the titles of the same coefficient of the correlation of Pirson:

Coeffіcіnt korelatsії mak miracle power. The value is taken from -1 to 1. Moreover, 1 means the direct proportion of the two values, and -1 to speak about the vortex lineage.

With more pairwise covariances, the sign can be combined with a covariance matrix, yak, yak awkwardly perverted, є mathematically supported by the following:

So the axis appears, but for the head components it is true:

That is, the main components, or, as they call them, factors are the power vectors of the correlation matrix. You can see the power numbers. If the number is larger, the larger the variance will explain the factor.

Know all the head components, for the skin podії, so є implementation X You can record this projection on the head component:

In this way, it is possible to show all the input pods in the new coordinates, the coordinates of the head components:

Consider the procedure for making head components and the procedure for determining the basis from factors and then further wrapping up, but also for interpreting the officials, as the procedure is ideologically close and give a similar result, let us analyze the factor.

During the downtime of the procedure of factorial analysis, one screeches even a glibky sensation. On the right, in the fact that the vastness of the vigilante signs is that the space is spared, then the factor is the sign that I want to describe the power of the modern world, but in a fantastical look (as if it is not too much of a warning) to be remembered. That is a formal procedure of factorial analysis, allowing for the appearance of appearances, you can move to the appearance of appearances, if you want to be invisible, but not less, to be seen in the new world.

You can let it go, but our brain is actively seeing the factors as one of the procedures for knowing the new light. Seeing bureaucrats, we will recognize the possibility of being new to describe how to visit us. The basis of the new descriptions is the turn in the fact that there are quiet appearances, that are for seeing the factors.

Trochi I will explain the essence of the factors on the byte rivn. Suppose we are HR manager. Before you come the helpless people, and even for the skin you will memorize the singing form, then write down the details of the message about the driver. Having looked over at the bottom of your notes, you can see how the deyaky graphy may have a singing connection. For example, a haircut for men will be in the middle short, not for women. You fox people, shvidshe for everything, you will set up only the middle of the people, and the farbuvati will only be the women. Even before the questionnaire data, the factor analysis is taken, then the very same factor will emerge as one of the factors, which will explain at once a number of regularities. Ale factorial analysis allows you to know all the factors that explain the correlation of fallowness in the set of data. It means that, except for the factor of stati, which can be promoted, seen and insinuated, including implicit, unwarranted officials. If the pidlog is clearly going to be shown in the questionnaire, then the only important factor is to get lost in the row. Assessing the health of people is tied to your thoughts, assessing the bugs of success, analyzing their assessments in the diploma and to that other signs, if you come to the office, you don't write down the assessment of people її point. Assessment of intellectu - tse і bonding factor, the head component with a high explanatory effect. Obviously, the component is not sposterіgmo, ale mi fіksuєmo signs, which are correlated with it. Mayuchi lives dosvid, we can pidsvidomo for some signs of the formulation about the intelligence of the spy. That procedure, like our brain, is, by the day, factorial analysis. Spostering for the time, as ti chi іnshi appearances are manifested spіlno, the brain, vikoristovuchi formal procedure, seeing the factor, as the representation of strong statistical regularities, the power of navkolishnogo light.

Seeing a set of factors

We showed yak filter Hebba vidilayak pershu head component... To appear, behind the help of neural framing, it is possible to easily remove not only persh, but all other components with ease. The price can be zrobiti, for example, in such a way. Admittedly, we have a sign of input. Vizmemo line neurons, de.

Hebb's ghosting algorithm(Haykin, 2006)

We will start the first neuron, yak filter Hebb, and see the first component of the brain. And the axis of the cutaneous advancing neuron will be directed to the signals, which are completely infused with all the anterior components.
Activity of neurons on a crotch n start yak

A correction to the synoptic wag yak

from 1 to, and from 1 to.

For all neurons, the view is similar to the Hebb filter. It’s less important that the dermal advancing neuron does not batter the entire signal, but only those who “didn’t bump” the neurons in front of them. The whole principle is called repeated assessments. In fact, according to the interconnected set of components, the viral update of the outbound signal is noticeable; This algorithm is called the Hebb algorithm.

In the zazagalnuyu Algorithm of Hebb it is not so kind as to have a "numerical" character. The neurons are guilty of being in order, and for the processing of their performance, they are guilty here strictly last. It’s not even possible to learn the principles of robotics from the brain, the skin neuron wants to interconnect with the ones, and it doesn’t work autonomously, because there’s no way of a swirling "central processor", which will bias the last of the podiums. Of these, the world is matched with algorithms, which are called decorrelation algorithms.

Obviously, we have two balls of neurons Z 1 and Z 2. The activity of neurons in the first ball will set up a picture, like projecting along axons to the attack ball.


Projection of one ball onto the

Now it is clear that the skin neuron of another ball of mechanical synaptic connections from the usim axons, which came from the first ball, could smell the stench in the middle of the neuron (the little one is lower). Axoni, scho to treat in such an area, to establish the receptive field of the neuron. The receptive field of a neuron is a whole fragment of the brain's activity, which is available to you for caution. The whole thing for the whole neuron is simply not sensible.

At the edge of the receptive field of the neuron, the region of the trochlea of ​​the lesser size is introduced, which is called the zone of strangulation. There is one skin neuron with its own susides, which can be used in the zone. Such calls are called bichny, or, as accepted in biology, terminology, lateral. Zrobimo lateral sound to gallow, to reduce the activity of neurons. The logic of their robots is an active neuron, the activity of all quiet neurons, which can be used to enter the galvanized zone.

The connection can be generated strictly from the axons or neurons in the interconnection between the two regions, and it can be set in a general way, for example, to the strong points of some central The structure is simpler for modeling, the vague rosette is larger anatomically from the point of view of organizing the connections in the real cortex.

The function of neuron activity can be recorded:

de - pidsumkov activity, - helpless axons, which are able to tap into the receptive area of ​​the opposite neuron, - beleless neurons, into the zone of suppression of which they will consume the inverted neuron, - the force of the apparent lateral galmuvannaya, taking in negative meanings.

Such a function of activity is recursive, as the activity of neurons appears as a fallow one from one. It should be done before the practical development is carried out iteratively.

The start of synaptic vaginas is similar to Hebb's filter:

Lateral vagi are triggered by the anti-Hebbian rule, which increases the galvanization of “similar” neurons:

The essence of the construction is in the fact that Hebb's navchannya is guilty of bringing to the vision on the neuron wagons the meaning that is the first head factor characteristic of giving gifts. If a neuron is built up, it will move towards any factor, only if it is active. If a neuron repairs a factor і, apparently, reacts to it, it’s fixing to block the activity of neurons, so that it’s smothering into its zone. As for the activation of the claim to the number of neurons, then we must create competition before the reversal of the strongest neuron, attracting all of them. Some neurons don’t miss anything, they’ll be surrounded by them at that moment, if they were instructed to do so with high activity. In such a rank, decorrelation is seen, so that the skin neuron in the boundaries of the region, the size of which is determined by the size of the zone and the setting, starts to see its own factor, orthogonal to all of them. The tsey algorithm is called the adaptive head component learning algorithm (APEX) (Kung S., Diamantaras K.I., 1990).

The idea of ​​lateral galvanization is close in spirit to the goodness of the mind behind the new models by the principle of “the best take away everything”, which also allows the decorrelation of this area, in such a way to make fun of it. The whole principle of vikorisovutsya, for example, in the neocognitron of Fukushima, how to self-organize the cards of Kohan, the same principle is stuck in the newly created widely visible temporal memory of Jeff Hawkins.

Visually, it is possible to forgive the relative activity of neurons. There is a bit of a bunch of easy implementations on computers, but no analogies to the real bark are seen. If you put your own for the meta of death, everything on the basis of the interconnection of neurons without learning of the called algorithms, then the same result can be achieved, as well as the lateral galvanization of the neuron, which will produce a positive ringing sound. Such a trick for a joke can be victorious, for example, in the framing of the adaptive Grossberg resonance.

As the ideology of neural fancy is tolerated, then the rule of "winners take everything" is more manual, so as the maximum of activity is simpler, the lower the iteration of the activity for the recovery of a lot of galmuvanny.

It's time to end the part. It was time to reach the end of the day, even though they didn’t want to crush them, they were tied for a snake notification. Do not be amazed by the KDPV, the picture proasocyuvati for me overnight piece Intellectі with a bureaucrat.

A freshly translated fundamental pedagogue of S. Khaikin (translated from another American version 1999r) is a full pretender to the title of the 2006 rock festival in the Russian literature on neuroinformatics. However, it is necessary that the translation would like to be displayed without obvious blunders, but there were no additions and comments that were added in order to clarify the terminology oblasts, for giving lists of synonyms - not all readers will be broad-minded). Comments could be used to represent and progress in the area of ​​piece neuronal framing, which was placed at the time of publication of the English original. I am encouraged that the book will be drunk and will be introduced when the circulation is completed. Tim more, well, the number of pardons in mathematical formulas is significant. Correction of gratuities by the head rank and the corresponding part is given. But if it is necessary, it means that I am not guaranteeing that the list of inaccuracies brought up here will be revised - having read the book "diagonally", in words and at a reasonable degree of respect, then there must have been a chance to let it go (or to have mercy myself).

Chapter 1

  • P.32 another paragraph. Only here the word "productivity" can be raised as the speed of the robot, the effort of calculating. Distant in the bottom of the "productivity" (performance) will be the sense of accuracy, the quality of the robotic neuromeasure (for example, on p. 73 in another paragraph below).
  • Page 35 clause 7. "VLSI Implementability" is more beautiful than the reshaping not as "scale", but as "effective implementation on NVIS - over-great integrated circuits".
  • Page 39 part 7. The word "spike" - "wikid, impulse" in the Russian Committee of Neuroscience is often translated and simply transliterated as "spike".
  • P.49 I will name the paragraph. Singingly, we will paint the buv bi term "orinting the graph" instead of "rectifying the graph".
  • C.76 third paragraph. Replacing the posilannya is humbly guilty of being posilannya on the book of Yeshbi.
  • P.99 visnovok 1. It is necessary to give a one-hour satisfaction with the same minds with a sign "
  • С.105 paragraph 2. It is necessary to insert the word "visible" before (visible).

chapter 2

  • P.94 wine 2. Posilannya on shvidsha for everything is wrong, it’s not a book and especially not to go for a name.
  • P.122 last paragraph. Having laughed at the phrase "deformation of the structure of neurons": leave the call of the brain to the brain, it’s not lost. It’s better for all that has been done, so that the memory is realized only from the connections of synaptic inputs (end) from the tentacles of dendrites, or to the mixing of one tentacle on the bottom (term from Fig. 1.2 on page 40, about So that our music is alive and breaking with you.
  • C.129 formula (2.39). substitute NS guilty bootie NS.
  • C.129 formulas (2.40), (2.41), (2.44). The top index is guilty of buty q substitute m.
  • C.137 first paragraph i formula (2.61). E maє buti in italics. I in formulas (2.64), (2.65), (2.67), (2.68) on page 138, etc.
  • C.142 formula (142). Pislya pershoї shooter dodati 0.
  • C.142 last paragraph. front the last word insert "minus".
  • P. 147 first paragraph. | L|=l... Tobto zminna l in the right part of the viraz, the guilt is given in italics (examples of the variant in the bottom of the rogue from the odinitsy).
  • P.151 formula (2.90). Insert the upper row of the figurine bow F.
  • C.151 formula (2.91). Insert "at" before N.
  • C.160 The last paragraph of the wine. "With a small amount" substitute for "a great amount".

chapter 3

  • P.173 Figure 3.1. Dates in italics are indicated to those accepted in the lower definitions, to that which are in scalar.
  • P. 176 formulas (3.5), (3.7). Guilty bootie w* replace w* .
  • C.176 stop row. Shvidshe for everything you need to ask for, if you want the value of food, you can get it.
  • P.179 wine. Guilty of being "kidnapped by f (w) by w"
  • P.180 stop row in front of wine. Replacement can be more beautiful than taking, but power can be neural.
  • P.184 between the viraz at the upper row of the formula (3.30). substitute x(n) guilty bootie x(i)
  • C.200 paragraph written for the formula (3.59). Having chuckled at the "nervousness of Guchi-Schwartz". It is to blame for all kinds of high school courses that the nerves of Koshy-Schwartz (Cauchy) are.
  • P.204 The first paragraph to section 3.10 is about the conversion of the Bayesian classifier into a linear distribution box in the minds of the middle man. Rely on respect for the mind, however, the covariance matrices of both classes (to be introduced in the distribution on p. 207), but at the phrase "Gaussian middle" , and give a quadratic surface.
  • C.206 formula (3.77). Dal will replace the meaning of the formula λ, the number of developments in the text and in Fig. 3.10 will be handled by Λ.
  • P.216 zavdannya 3.11. Those that are given in the upper border of the sumi, maє buti are transferred to the sign of sumi (and the minus can be taken in front of the sum). Also in the paragraph of the text of the formula w T x guilty bootie w T x

chapter 4

My comment to the chapter: a nightmare, a newcomer in neural framing and optimization methods. At least, when looking at only students of provincial technical VNZ, they are ready to compete for high rates. Viklad was mixed in a series of and necessary, and little-needed speech, as without placing emphasis and re-assignments of the viclades (if we go "all or nothing" to replace the additional procedures). Plus a lot of empirics. Why not just the method of calculating the gradation of the folding function (the neural network plus the central function over the input and, if necessary, over the powers of the neurorecord) methods of quadratic programming), і in the range of historical applications of correct and incorrect approaches to victoriousness, to be calculated as a gradual gradation from the point of view of the theory of gradient optimization and maximization of efficiency.

I would like to bachiti in the head (or lower) to the same. According to the first, the method is based on the method of the least squares of the central function, especially for the new classifier (for example, cross-entropy function). In another way, there is a greater view of the power of the mother to be stored with decile storage functions: on the application of regularization according to Tikhonov through an explicit reduction of the size of the scale of the scale of the robot in the scale of the scale either using the Flat minina search method, Hochreiter and Schmidhuber, or using the CLearning method, clearing the input signals in the grid by Andreas Weigand with the co-authors. Thirdly, there is more detailed description of the possibility of counting other old ones in a fancy (meaning LeCun's and Dracker's robots, methods that were re-insured in an eye-catcher). Fourthly, a more detailed description of the methods of calculating the informational-corrosiveness of the new elements and signals in the measure ). As a matter of fact, it’s explicitly a request (you don’t know yourself) to the ability to change gradients based on input signals from a fence (for interconnecting tasks on neural networks that are necessary for solving direct tasks, for the CLearning method). Plus for the whole series of chapters, de-posting of the teaching with the teacher, writing in detail the idea of ​​the curves for the neuron framing.

chapter 5

  • C.357 for the formula (5.23). Dalі on decіlkoh sides E may be given in italics or in bold, and the form will be written down without a system. Correctly - in italics, for E (F), E s (F), E c (F), E (F, h).
  • C.361 formula (5.31). Replace the lower index H guilty bootie H .
  • P.363 last paragraph. "... with a linear combination ..." replace "... with a linear superposition ...".
  • C.364 formula (5.43). Come 1 / λ.
  • C.367 formula (5.59). σ replace δ.
  • Page 369 for the formula (5.65). I am guilty of the "line combination" replacement of the "line superposition".
  • P.373 third row of the formula (5.74). Insert the crooked shackle in front of the other t i .
  • P.382 formula (5.112). At the bottom of the sumi, add "not ready k".
  • P.390 name of the product 5.12. In Russian science, win over the term "nonparametric regression" (the Russian statistics method is called so) or "nuclear regression" (like translating "head-on").
  • C.393 formula (5.135). Insert "... for all ...", like in (5.139) on the next page.
  • P.399 "middle" paragraph. "... the clustering algorithm by k-middle... ", give the word" average "will not be missed anymore.
  • P.403 non-numbered list. It is also necessary to globally and unambiguously manipulate authorization from one experiment, if you want a lot of it.
  • P.404 is the first item on the list. I do not understand, especially in the part "pouring into the input parameters." Shvidshe, the greater the value of λ, the less the amount of money added to the model's power.
  • P. 408 first paragraph. Summaries of power on, maybe pidijde.
  • P.408 row 6 paragraph 2. "basic function" replaces "fundamental function".

chapter 6

  • P.431 stop the list before the distribution 6.4. I did not understand the "shortness" of the proponated vibration through the middle vibration b 0 will not be possible).
  • C.434 formula (6.35). index i at the rest x butti is not guilty.
  • P.435 unnumbered formulas in Mercer's theorem. Replace ψ is guilty but φ.
  • P.444 wine. The nickname of Huber was previously translated into Russian as Yak Huber, not Haber (for example, the translation of his book in an hour of the CPCP: Huber, "Robustness in Statistics").

Chapter 7 (in general)

  • P.459 third row from the top. The decoding of the term "weak algorithm of navchannya" is given on page 467 in another paragraph above.
  • P.459 unnumbered subparagraphs in clause 2. The term "sluice gating" is a shift of the term "gating network" over the same clumsy, albeit (and at all good) version of the Russian leave. Melodiously, more beautifully bulo b vikoristovuvati term "zvazhuє the fringe", a universal yak for vypadku zhorystogo mixing (coefficients multipliers 0 or 1 for a kerovan signal), so also for the soft control of multipliers.
  • P.463 clause 2. The part of "not" is tidied up in the sense of the proposition - the dispersion of the ensemble is less than the dispersion of the outer functions.
  • P.471 first rows. "Productivity" (nagaduєmo, "productivity-performance" here is not in the sense of speed, but in the sense of the accuracy of the decision and publicity - see our commentary to p. 32), the outward method of making it possible to lay down in the process of developing for other and advancing experts.
  • P.472 table 7.2 last row. Guilty bootie F fin ( x)=…

Bibliography

  • A bunch of times the words application, approximation, approach, applied, support, mapping, applicability, upper are written in one p.
  • ... The correct spelling of the nickname of one of the authors can be written in.
  • ... Mueller's nickname is correct - like his namesake.
  • ... First author - B u ntine.
  • ... It was in the same NIPS, well i.
  • ... Remaining from the authors of the correct titles of Art.
  • ... Require weak substitution week.
  • ... The last author of the correct names of Art.
  • ... First - Landa u.
  • ... Tse chapter in the book.
  • ... Sch ö lkopf.
  • ... The name is "... bia s term ". The duplicate is spelled correctly.
  • ... The name is "... gamm on".
  • ... Repeat.

In the given statty, selected materials - in the main Russian committee - for the basic implantation of piece neuronal fences.

Piece of neural framing, or INS is a mathematical model, as well as software or hardware integration, motivated by the principle of organizing and function of biological neural fences - fingering nerve cells living organism. The science of neural fences has been growing for a long time, however, in conjunction with the remaining advances of scientific and technical progress, the sphere of initiatives has begun to gain popularity.

books

It’s a good thing to do more with the classic method of reading - for the help of books. We picked Russian books from a great number of applications:

  • F. Wasserman, neurocomputer technology: theory and practice. 1992 r
    At the bottom, in the outwardly accessible form, the foundations of the neurocomputers are prompted. The structure of neural framing and development by algorithms and adjustment is described. Outside the chapters of assigning nutrition to the implementation of neuronal fences.
  • S. Haykin, Neuron lines: A new course. 2006 r
    Here we look at the main paradigms of piece neural fences. Submissions of material to revenge on Suvore are mathematically based on all neuromegalic paradigms, illustrated with butts, description computer experiments, To take revenge on impotent practical workers, as well as great bibliography.
  • D. Forsyth, Computer zir. Happy pidhid. 2004 r
    Computer zir is one of the most demanded areas for this stage of development of global digital computer technologies. It is required for virobniths, when controlling robots, when automating processes, in medical and vital supplements, when being guarded from supporters and when robots with personal computers, zooms, digital images.

Video

There is not much more accessible and light-hearted, less visual available for additional video:

  • To be smart, to see the machine, to wonder at the axis two lectures from Shaday Yandex.
  • Entry In the main, the principle of designing neural fences is to visibly go for the advancement of knowledge with neural fences.
  • Course of lectures on the topic "Comp'yuterne zir" from VMK MSU. Computer zir is the theory and technology of the creation of piece systems, which are used to display and classify objects in images and video recordings. The goals of the lectures can be brought before the introduction of the curriculum vitae and folding science.

Osvіtnі resources і korisnі power

  • Portal for piece intelligence.
  • Laboratory "I am an Intellect".
  • Neural lines in Matlab.
  • Neural patterns in Python (eng.):
    • Classification of the text for additional help;
    • Simple.
  • Neuron of the framing on.

A series of our publications on topics

Earlier we have published the same course #[Email protected] along neural hem. There is a whole list of publications for your happiness.