 Maszyn learning, the reason why. This is the example, the right-hand character's recognition. This is the one of the basic example of the good approach for the neural networks. It would be very difficult to develop the algorithm for that. The neural network approach for this is a very good solution in terms of the time of the development and in terms of the accuracy. Czy to jest technologia na sklepie? Oczywiście nie. Pierwsza modela neural została inventowana w 1943 roku w USA. Pierwsza implementacja, jedna z najwyższych, najwyższych implementacji na neural networku, została implementowana w 1957 roku w USA. To była aplikacja na imię, używając dźwięków dźwięków, a rozwiązanie, tak jak na kamerę, było 20 do 20 pixel. Struktura na neural networku było bardzo proste. To była tylko jedna z nierunek nierunek. I teraz. Ta idea jest jeszcze w nierunek. To jest tzw. multilayer perceptron. I ta solucja jest bardzo dobre dla podróży time-serious. Na przykład na neural networku, który jest w porządku, 1, 2, 3, 4, etc. Możemy zniszczyć serię, jak 2, 3, 4, a potem na neural networku można kontynuować. Ale nie kontynuować, jak 5, 6, 7, ale 4.97, 5.90, coś, ponieważ to jest podróży time-serious. Dobra, wracamy do szkoły. Myślę, że wszyscy, pamiętajmy, ta struktura, to jest nierunek biologiczny. Mamy cell body, dendrites i axon terminal. Dendrites są equivalenty z nierunem. Cell body jest equivalent z procesorem. Axon to jest our output. I jedna z najważniejszych msg. Neural networks to nie digital world. Nie możemy powiedzieć, że podróż jest 100% na pewności lub 100% nie na pewności. Możemy rozmawiać o podróży podróży, między 0 i 1. To jest, myślę, najważniejszy msg. I nie jest w porządku podróży, jeśli podróży podróży jest dobrze czy nie. Wydaje się z rozwiązaniem i aplikacją. Nie mogę powiedzieć, że podróży podróży z nierunem, 0,95% to ok. Wydaje się z aplikacją. To jest biologia, to jest realna życie i to jest model matematyczny z nierunem. Mamy podróży i do każdego podróży możemy spotkać podróż. Mamy sety koeficzji W1 do Wm. Te koeficzje są tak zwane kawałki w nierunach. Więc do każdego podróży nierunem to jest podróż i to jest też podróż. To mamy kawałkę procesora. Więc musimy muntiplić co jest funkcjonacją kawałki procesora. Mamy kawałkę podróży podróży koeficzji kawałki i to muntiplić wszystkie skończenia podróży. Więc to jest kolejny podróż dla Was. Muntiplić i muntiplić. Więc jedyna operacja podróży nierunem. Ok, więc muntiplujemy podróży kawałki podróży podróży podróży kawałki podróży podróży podróży podróży podróży detonatur podróż Jeżeli ten wynik jest wyższy lub większy niż 0, ten wynik jest 1. Więc ta aktywacja funkcjonuje w tym, co nazywa unia. Więc to jest model matematyczny z networku neuro. I teraz zobaczmy, co jest możliwe brycie między biologią i bardzo zauważywane techniczne implementacje z strukturą networku neuro. To jest krawka limulosa. To jest nazywa latynna. Z drugiej nazwy jest krawka horczowa. I ta krawka ma 5 lub 7 ojczy. To są ojczy brycie. I ten ojczy jest tak zwany kompont ojczy. Kompont ojczy znaczy matrix ojczy. I to jest strukturą jednej wody. To jest kolejny przykład kompont ojczy. I to jest cylikon implementacja kompontu ojczy. Jak wiesz, po prostu po swoich ojczyach mamy ojczy brycie. I to jest krawka. I wiesz, co jest w tym, co jest w tym, co aktywują? W brycie, przed swoimi ojczymi. To jest 10 bryty w południu 9 bryty per sekund. Więc bardzo dużo. I dzięki temu, co przeprocesowało i nerwę, moja bryja musiała pracować tylko kilku bryty per sekund. I to jest ten przykład, czy ojczym zimplifowanym implementacją i nerwem tej limułowej krawki. Ten triangel jest symbolem nerwem. Więc mamy trzy wody i następujące bryty. Minus 0,5, potem 1, potem minus 0,5. I to jest bryja. W naszym razie, to może być considerowane jako brycie. Więc dwóch sensorów są wypozywane do bryty. I dwóch sensorów są wypozywane do bryty. I tutaj jest bryty bryty i bryty. Let's consider this, the first nerwem. This input is activated, so 1 multiplied by 0, minus 0,5. Plus 1 multiplied by 1. Plus 1 multiplied by minus 0,5. So the result is zero. The result of accumulation is zero. And it is less than 0,25. So the output is zero. And for this neuron, we can do the same evaluation of the output, we have one. And then zero and zero. What is the conclusion? Crap exposed to the full daylight can see nothing. And it makes sense from the biological point of view. Because the brain of crap is quite small. So there is no need for crap to see the landscape. But crap, thanks to this nerve eye, can very easily detect the border between shadow and light. Crap can easily count the objects. Crap can easily detect the edges, which makes sense from a biological point of view. And of course if exposed to the shadow, to the night, to the darkness, crap can see nothing as well. I this structure just reduced the amount of information and lets us to extract some features. Like moving objects like edges. And this structure, this neural network is so called sight breaking. Because the sight weights are negative. That's why the sight inputs are breaking somehow the activation of the neuron. I this structure is used for computer vision for image recognition. We just discussed the atomic neural model. But today we are talking about neural networks. And this is the basic structure of the neural network. The neural network consists of at least three layers. Input layer, hidden layer, or dense layer depends on the wording. And the output layer. And what is very interesting? The number of the neurons of the elements in the output layer is equal to the number of classes to predict. Today we will focus on the, and we will practice the audio acoustic scene classification application. We will use microphone to distinguish between different acoustic scenes. Like indoor, outdoor and in vehicle. So in our case we have three classes to predict. So the number of the neurons here will be equal to three. The activation functions. We discussed the most simple unit step activation function. But in practice of course there is more types of activation functions. Starting from unit step and the generalization of the unit step. So the unit step is vary from zero to one. The generalized one, so-called signum vary from minus one to one. Then linear, piecewise, logistic hyperbolic tangent. I think quite well known to us as electronic engineers characteristic rectified linear unit. So just a diode characteristic. And this activation function is very useful for the image recognition. Our reality is not linear. That's why we need not linear activation function for the image recognition neural network. Today we will focus on the supervised learning. It means that we need supervisor. We need teacher. We need teacher. And in our case this teacher it is of course the algorithm. What is the basic idea behind the learning process? It is so-called forward propagation. We are feeding the network with the batch of input data. Then we are analyzing. We are comparing the output on the neural network which is so-called result of the inference. Or just the inference. So the prediction means inference. We are comparing the result of the prediction to the expected value. And then if it is not in line, the teacher, in our case the algorithm, tunes the coefficients inside the neural network. And in our case the coefficients are the weights and the biases. I this is very complex process in terms of mathematics behind. This is the problem of the optimization of the multi variables functions. In fact we need to find the minimum of the loss function. I will explain later on and do it step by step. But do not worry, those algorithms are ready. The libraries are ready and we can just reuse the ready solutions. So the teacher algorithm job is just a feedback feeding the inputs with batch of data. Then comparing the output to the expected output and then adjusting the weights. And again the new wording. So the prediction in neural networks means inference. The expected value of the prediction of the inference. It is grant through data set. Is it clear? We are touching the surface only. I am fully aware. But we need to catch the basic idea. For the supervised learning process we need the teacher. We need the dedicated algorithm to tune the coefficients, the weights and biases. Our solution kube.ai tool supports Cortex M4 and Cortex M7. Because of floating point unit. Because of the hardware acceleration of the computation. Next important message to you. The neural network, the design of the neural network and the learning of the neural network. It is I would say this is 10%, 5% of the process. I think there are two very difficult parts. The data set collection. It could be very expensive process. And the preprocessing stage. Because it is better to avoid feeding the neural network with raw data. I think almost always we need the preprocessing stage. Which is not trivial as you will see today. And this is for the R&D managers message. You need really experienced DSP experts. In your team. This is important stage. Quite time consuming. And when it is time consuming it means that it is expensive. We discuss the most simple structure of the neural network. But of course in practice we have a lot of standard neural network types. Starting from Perceptron. So remember the 1957, the hardware implementation. Then we can simulate for example long short term memory. Just by implementing the feedback. We can implement some recurrent neural network structures. Again to simulate some memory process to be able to predict time series. Like stock exchange for example. And very practical structure in terms of industry. Convolutional network. So this is the most common structure. Which is used for the image recognition. And the structure of this network reflects the functionality. Because we have inputs. Then this triangle structure. It is filtering stage. Here we are reducing the amount of data. Do you remember our INERF. Which reduce the amount of data from 10 in power of 9 bits per second. Down to several bits per second. So this is the role of the convolution. Filtering of the data. And we can also describe this stage as a feature extraction. So the convolution layers can be used to extract the features. Like particular grayscale of particular areas of image. The edges counting of the points etc. Then we have brain of our neural network. So it is dense layer or dense layers. Dense means that the input of the following layer is connected to each output. Of the previous layer. And then we have output layer. And the number of the neurons is equal to the number of classes. Number of states of the system to predict. Deconvolutional network. So here we have convolution. Here we have deconvolution. Sometimes it seems to us that we have not so much data. But using deconvolution we can try to find hidden patterns.