In the past few years, we have a complete team dedicated to artificial intelligence research and experimentation. The team is focused on developing new evolutionary computational methods (ECs), including designing artificial neural network architectures, building business applications, and using challenging methods inspired by natural evolution to solve challenging computing problems. The development momentum in this field is very strong. We believe that evolutionary computing is likely to be the next major issue in artificial intelligence technology.
EC, like Deep Learning (DL), was introduced decades ago, and EC can also be improved from the available computing and big data. However, it addresses a very different need: we all know that DL focuses on modeling what we know, while EC is focused on creating new knowledge. In this sense, it is the next step in DL: DL can identify objects and speech in familiar categories, and EC enables us to discover entirely new objects and behaviors - objects and behaviors that maximize specific goals. As a result, EC has enabled many new applications: designing more effective behaviors for robots and virtual agents, creating more effective and cheaper health interventions, and promoting mechanization and biological processes in agriculture.
Not long ago, we published five papers to report significant progress in this area. The report focuses on three areas: (1) The DL architecture has reached the latest state of the art in three standard machine learning benchmarks. . (2) Development techniques are used to improve the performance and reliability of practical application development. (3) Prove the solution to the evolutionary problem on very difficult computational problems.
This article will focus on the first area in which to optimize the DL architecture with EC.
Sentient reveals a breakthrough study of neural evolution
Much of the deep learning depends on the size and complexity of the network. With neural evolution, DL architectures (ie, network topologies, modules, and hyperparameters) can be optimized outside of human capabilities. We'll cover three examples in this article: Omni Draw, Celeb Match, and Music Maker. In these three examples, Sentient successfully surpassed the most advanced DL benchmarks using neural evolution.
Music production (language modeling)
In the field of language modeling, the system is trained to predict the next word in the "language library", such as the large collection of texts in the Wall Street Journal for several years. After the network makes predictions, this input can also be cyclically entered. Thus the network can generate a complete sequence of words. Interestingly, the same technique applies equally to music sequences. The following is a demonstration. The user enters some initial notes and the system creates a complete melody based on the starting point. Through neuron evolution, Sentient optimizes the design of gated periodic (long-term short-term memory or LSTM) nodes (ie, the "memory" structure of the network), making the model more accurate in predicting the next note.
In the field of language modeling (predicting the next word in a language corpus called Penn Tree Bank), the benchmark is defined by the puzzle point and is used to measure how the probability model predicts the real sample. Of course, the lower the number, the better, because we want the model to "confuse" as little as possible when predicting the next word. In this case, the perceptron defeated the standard LSTM structure with a puzzle point of 10.8. It is worth noting that in the past 25 years, although humans have designed some LSTM variants, the performance of LSTM has not improved. In fact, our neuronal evolution experiments show that LSTM can significantly improve performance by adding complexity, memory cells and more nonlinear, parallel pathways.
Why is this breakthrough important? Language is a powerful and complex intelligent structure of human beings. Language modeling, the next word in predictive text, is the benchmark for measuring how machine learning learns language structures. Therefore, it is the agent for building natural language processing systems, including speech and language interfaces, machine translation, and even medical data including DNA sequences and heart rate diagnostics. And we can do better in language modeling benchmarks, and we can use the same technology to build better language processing systems.
Omni Draw
Omniglot is a handwritten character recognition benchmark that recognizes 50 different alphabetic characters, including real languages ​​like Cyrillic (written Russian), Japanese and Hebrew, and such as Tengwar (written language in Lord of the Rings). Artificial voice.
The example above shows multitasking learning, which allows you to learn all languages ​​at the same time and take advantage of the relationships between characters in different languages. For example, the user enters an image, and the system outputs the meaning of the different languages ​​according to the match, "This will be X in Latin, Y in Japanese, Z in Tengwar, etc." - using the relationship between Japan, Tengwar, and Latin. Find out which roles are the best match. This is different from a single-task learning environment where the model trains only one language and does not establish the same connection on the language dataset.
Although Omniglot is an example of a data set, there is relatively little data in each language. For example, it may have only a few Greek letters, but many are in Japanese. It can use the knowledge of the relationship between languages ​​to find a solution. Why is this important? For many real-world applications, the acquisition of tagged data is very expensive or dangerous (eg medical applications, agriculture and robotic rescue), so models can be automatically designed using relationships with similar or related datasets, to some extent Can replace lost data sets and improve research capabilities. This is also a good proof of the ability of neurodevelopment: there can be many connections between languages, and evolution has found the best way to combine their learning.
Celeb Match
Celeb Match's demo is also suitable for multitasking, but it uses large data sets. The demo is based on the CelebA dataset, which consists of approximately 200,000 celebrity images, each of which has 40 binary tag attributes such as "male and female", "with or without beard" and so on. Each attribute produces a "classification task" that directs the system to detect and identify each attribute. As a fun add-on, we created a demo to do this: the user can set the required level for each property, and the system will determine the closest celebrity based on the evolved multitasking learning network. For example, if the current image is Brad Pitt's image, the user can add a "grey hair" attribute that has been found to be similar to him but with different hair.
In the field of CelebA multitasking face classification, Sentient used evolutionary computation to optimize the network of these detection attributes, successfully reducing the error of the overall three models from 8% to 7.94%.
This technology has made artificial intelligence a big step forward in predicting the various attributes of humans, places and the physical world. Unlike training networks that find similarities based on abstraction and learning functions, it makes similar semantics and interpretability possible.
Our company`s current transformers have high precision,wide range,small volume and good linearity that can be used to the field of electronic watt-hour meter, electric energy metering, electronic detection.
Performance
1.Power frequency insulation strength:The insulation between the primary winding and the secondary winding and the ground part of the CT can bear 4kV power frequency voltage for 1minute
2.Interturn insulation strength:The secondary open circuit, the primary winding through the rated current 1min, no inter-turn damage in the transformer
3.The deviation is better than the industry standards and national standards
Voltage Transforvmer,Durable Voltage Transformer,New Voltage Transformer,Good Voltage Transformer,High-class Voltage Transformer
Anyang Kayo Amorphous Technology Co.,Ltd. , https://www.kayoamotech.com