News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
Why do you have deep learning?
Why do you have deep learning?
Why do you have deep learning?
Deep learning can use more data or better algorithms to improve the results of learning algorithms. For some applications, deep learning performs better on big data sets than other machine learning (ML) approaches.
In terms of performance, deep learning explores the probability space of neural networks. Compared with other tools, deep learning algorithms are more suitable for non-monitoring and semi-monitoring learning, more suitable for strong feature extraction, and more suitable for image recognition categories, text recognition categories, Voice recognition categories, etc.
Why deep learning is so keen, because it is not characterized by any loss function, nor is it limited by specific formulas, which makes the algorithm more open to scientists, it can be better than other traditional machine learning tools. The way to stop using and expanding.
Further, the use of a concept in the book "Deep Learning Book MIT" (free e-book: http://www.deeplearningbook.org/) may further give reasons why it is necessary to learn from machine learning to deep learning.
After 1960, the limitations of linear classifiers were recognized. It only splits the input space into very simple regions, two halves separated by a hyperplane. With regard to issues such as image and speech recognition, the demand input-output function is insensitive to input uncorrelated changes (position changes, direction changes, illumination changes, treble and bass changes in speech), and is sensitive to categories (eg white wolves) And Samoyed).
At the pixel level, two different poses, the photos of the Samoyed dogs in different environments will be very different, and the same background, the same position of the Samoyed and white wolf photos may be very similar. A linear classifier that directly manipulates image pixels or other "shallow" classifiers may not be able to distinguish the last two photos while placing the first two in the same category. This is why shallow classifiers require a good feature extractor—selectively generating representations of important category information in the image, while at the same time having invariance on irrelevant information such as gestures—to deal with the choice-independent dilemma.
In order to make the classifier more powerful, it is possible to apply generalized nonlinear features and kernel functions. However, generalized features (such as Gaussian kernel function) can be poorly distributed. The conventional method is to manually design a feature extractor, which requires a large number of engineering experience and domain experts to complete. If good features can be learned automatically through learning, the above problems can be prevented, which is the central advantage of deep learning.
It can be seen that deep learning has its center more than traditional machine learning algorithms, and sometimes it can only use deep learning to complete certain identification tasks by using exercise neural networks.
Deep learning can use more data or better algorithms to improve the results of learning algorithms. For some applications, deep learning performs better on big data sets than other machine learning (ML) approaches.
In terms of performance, deep learning explores the probability space of neural networks. Compared with other tools, deep learning algorithms are more suitable for non-monitoring and semi-monitoring learning, more suitable for strong feature extraction, and more suitable for image recognition categories, text recognition categories, Voice recognition categories, etc.
Why deep learning is so keen, because it is not characterized by any loss function, nor is it limited by specific formulas, which makes the algorithm more open to scientists, it can be better than other traditional machine learning tools. The way to stop using and expanding.
Further, the use of a concept in the book "Deep Learning Book MIT" (free e-book: http://www.deeplearningbook.org/) may further give reasons why it is necessary to learn from machine learning to deep learning.
After 1960, the limitations of linear classifiers were recognized. It only splits the input space into very simple regions, two halves separated by a hyperplane. With regard to issues such as image and speech recognition, the demand input-output function is insensitive to input uncorrelated changes (position changes, direction changes, illumination changes, treble and bass changes in speech), and is sensitive to categories (eg white wolves) And Samoyed).
At the pixel level, two different poses, the photos of the Samoyed dogs in different environments will be very different, and the same background, the same position of the Samoyed and white wolf photos may be very similar. A linear classifier that directly manipulates image pixels or other "shallow" classifiers may not be able to distinguish the last two photos while placing the first two in the same category. This is why shallow classifiers require a good feature extractor—selectively generating representations of important category information in the image, while at the same time having invariance on irrelevant information such as gestures—to deal with the choice-independent dilemma.
In order to make the classifier more powerful, it is possible to apply generalized nonlinear features and kernel functions. However, generalized features (such as Gaussian kernel function) can be poorly distributed. The conventional method is to manually design a feature extractor, which requires a large number of engineering experience and domain experts to complete. If good features can be learned automatically through learning, the above problems can be prevented, which is the central advantage of deep learning.
It can be seen that deep learning has its center more than traditional machine learning algorithms, and sometimes it can only use deep learning to complete certain identification tasks by using exercise neural networks.
PREVIOUS:Passport OCR
NEXT:Deep learning based OCR recognition tech