What kind of deep learning is most suitable for your enterprise?

Deep learning （Deep Learning） It's machine learning (Machine Learning) A subset of , It uses artificial neural networks （Artificial Neural Network） To perform complex calculations on a large amount of data . This kind of algorithm imitates the structure and function of human brain , And have a strong learning ability 、 Scene adaptability and portability .

Benefit from more and more abundant data and higher machine computing power in the era of big data , Deep learning gradually shines in various application scenarios , For many traditional enterprises to reduce the labor cost of a large number of repetitive operations , It also provides innovative dimensions with great potential for start-ups .

In recent years, most of the popular technologies include deep learning , such as Autopilot 、 Computer vision （ Face recognition , Image recognition ）、 Audio recognition 、 natural language processing 、 supercomputer 、 Virtual Assistant 、 Automatic identification of agricultural crops 、 Manufacturing defect detection wait . Even in some more traditional scenes , Deep learning (DL) It also replaces some traditional machine learning because of its higher accuracy (ML) Model , such as Anti fraud technology （Fault Detection）, Customer relationship management , Automated customer service , E-commerce and investment modeling .

There are also some novel innovative applications , such as Google Colab released Disco Diffusion AI painting , And from MIT Affectiva “ mind-reading ”： Recognize facial features through computer vision algorithm to Recognize emotions .

Disco Diffusion AI Art

Such rich application scenarios , It must not be satisfied by a single algorithm . Deep learning (DL) Under the joint promotion of academia and industry , Get rapid iterative development , Dozens of different algorithms have evolved , Suitable for performing different tasks . and If you can choose the most appropriate algorithm in the scene , Can often bring the greatest benefits .

To help readers understand the different applications of different algorithms , Next, we will briefly introduce nine commonly used deep learning algorithms , And at the end, it discusses the use of deep learning under different conditions （DL） Or machine learning （ML）：

|  1. Multilayer perceptron Multilayer Perceptrons (MLPs)

MLP Is the simplest and ‘ original ’ The neural network of , It consists of a fully connected input layer 、 Hidden layer and output layer , Applicable to image classification , Speech recognition, etc , It is the most basic model of deep learning .

Loukas, George & Vuong, Tuan & Heartfield, Ryan & Sakellari, Georgia & Yoon, Yongpil & Gan, Diane. (2017). Cloud-Based Cyber-Physical Intrusion Detection for Vehicles Using Deep Learning. IEEE Access. 6. 3491-3508. 10.1109/ACCESS.2017.2782159.

MLP The classic example of is number recognition , That is, we randomly give a picture with handwritten digits on it , It can be used to judge the number written on the picture .

Credit to: 3Blue1Brown

|  2. Convolutional neural networks Convolutional Neural Networks (CNNs)

CNN It is usually suitable for image recognition , It consists of many layers of Neural Networks , Through the convolution layer （convolutional layer） Feature extraction for each small area of the input content , Get the characteristics of different dimensions or after stretching and compression ; Then the pool layer simulates the human visual system to reduce the dimension of the visual input object , This can not only reduce the data level of calculation , It can also avoid over fitting .

General image classification 、 Image recognition 、 Video recognition will be used CNN To deal with it .

|  3. Cyclic neural network Recurrent Neural Networks (RNNs)

Compared with ordinary Neural Networks ,RNN The biggest difference is that the results of the last output can be substituted into the next calculation for common training , Therefore, it is better at dealing with internal closely dependent ‘ Sequence data ’（ For example, text 、 Audio etc. ）.

It is often used in the application of natural language processing , Such as text generation 、 Machine translation 、 Audio and Video Tags , Or the prediction of stock trend with close connection between data .

|  4. Long and short term memory network Long Short Term Memory Networks (LSTMs)

LSTM It's a special kind RNN, It can remember the long-term feature by default in the setting , So it is very suitable for learning long-term dependence , It is often used in time series prediction and speech recognition .

|  5. Generative antagonistic network Generative Adversarial Networks (GANs)

GANs It is a generative deep learning algorithm , By generating false data ‘ generator ’ And distinguish the true and false information ‘ Discriminator ’ form . It can create new data instances similar to training data , Such as face photo generation and 3D Rendering , Is its common application .

Face generation , Credit to: Sarvasv Kulpati, 'A Brief Introduction To GANs'

GAN In the future, the game direction has great development prospects , But at the same time, it also needs very strong hardware support .

|  6. Restricted Boltzmann machine Restricted Boltzmann Machines( RBMs)

RBM It is a kind of randomly generated neural network , It consists of a visible layer and a hidden layer , The neurons in the apparent layer and the hidden layer were connected in both directions .

RBMs, Credit to Wikipedia
 

The probability distribution can be learned from the input data set , Commonly used in dimension reduction ( Less hidden layer )、 classification 、 Return to 、 Feature learning （ Hidden layer output is the feature ） And theme modeling .

|  7. Deep belief network Deep Belief Networks (DBNs)

DBN It is composed of multiple restricted Boltzmann machines , It is often used for video recognition and motion data capture .

|  8. Self organizing chart Self Organizing Maps (SOMs)

SOM It is an unsupervised artificial neural network , It can reduce the dimension of data in high-dimensional space , Form a low dimensional discrete map , It is often used for complex data visualization , Help people better understand data .

|  9. Automatic encoder Autoencoders

Automatic encoder is also an unsupervised neural network model , It can learn the hidden characteristics of data , And encode or decode the data , It is often used for feature extraction or data dimensionality reduction , Be regarded as PCA Superior substitution of .

Autoencoders, Credit to Arden Dertat, 'Applied Deep Learning - Part 3: Autoencoders'

Conclusion ： There are differences in the applicable scenarios and functions of the nine algorithms , Multilayer perceptron 、 Convolutional neural networks are mostly used in image processing ; Generative antagonistic network 、 Restricted Boltzmann machine 、 Self organizing chart 、 Automatic encoder is suitable for data processing ; Cyclic neural network 、 The long-term and short-term memory network is suitable for audio processing .

Deep learning (DL) The application in industry has effectively solved many problems that are difficult to deal with in traditional machine learning , For example, audio and video processing and more accurate prediction 、 Classification results ; meanwhile , Interesting ideas collide with deep learning , Often can burst out wonderful products . But we are the helpers of products and Technology , For the sake of customers , Occasionally I need “ persuade sb. to resign from an official position ” Some customers start up with a desire for deep learning ：

No algorithm is perfect , Deep learning (DL) The same is true . Compared with traditional machine learning （ML）, It needs to More training data 、 Also needed GPU Support for 、 as well as Long time training and adjustment , It also means higher investment （ get data 、 Model training And fine-tune、 Model performance maintenance 、 Time cost and labor cost ）. Some scenes can be selected directly pre-train Good model , such as NLP Of BERT and GPT3, these pre-train The model only needs a small amount of data to fine tune the model （fine-tune）, You can get good performance .

But even so , The amount of data required for fine-tuning is usually larger than that of traditional machine learning models . At the same time, the traditional machine learning model has better interpretability , And fast and convenient , Sometimes it is more suitable for prototypes that need fast iteration .

So when the data assets of start-ups were not rich enough , Can be used first pre-train Model , Or consider whether traditional machine learning can meet the early functions , And on this basis, establish a plan to smooth the transition to the deep learning model .

If you are also interested in deep learning and machine learning , Or want to consult whether the enterprise is applicable 、 What kind of deep learning or machine learning model is applicable , Welcome to scan the QR code below to add wechat , Contact us directly ：