Dialogue with AI trainer from Fudan University: Is MOSS derived from "Wandering Earth 2"? What are its future goals? Model | Training | Wandering Earth 2
Dialogue character: He Zhengfu, artificial intelligence trainer for the MOSS project at the Natural Language Processing Laboratory of Fudan University
Q: What is MOSS and what are its main functions?
Answer: MOSS is a conversational language model that can provide various direct or indirect assistance to people's lives. It can conduct Q&A on daily life knowledge, help check weather, plan travel, etc; It can assist in efficient office work, such as automatically processing tables, generating outlines, drafts, translations, etc., and also master professional knowledge in fields such as finance, healthcare, and education. Many industries are introducing conversational language models represented by MOSS, such as automotive voice assistants, customer service, etc., which will have the effect of cost reduction and efficiency improvement.
Q: What is your specific training process for MOSS?
Answer: The essence of parameters in a large model is a massive matrix, which performs simple, heavy, and repetitive numerical operations on the input text to ultimately obtain the content that needs to be generated. We can collect and "clean" the corpus on the internet, and allow large models to learn knowledge from these corpora. Specifically, the learning process involves constantly "reading" the text and adjusting the internal parameters of the large model to deepen its understanding of language, ultimately obtaining some kind of "intelligence". This process is called training.
In the process of building MOSS, we empower it with powerful capabilities through a three-stage "reading" process. The first stage is the acquisition of basic knowledge. MOSS extensively "reads" almost all texts on the network, and due to its large number of parameters, it is sufficient to cover a vast amount of knowledge. The second stage is the acquisition of dialogue ability. MOSS learns to answer human questions through dialogue by reading dialogue data and utilizing the knowledge acquired in the first stage. The third stage is alignment. Due to the possibility of misleading responses, MOSS will suppress the generation of content that does not comply with human laws and moral ethics based on human feedback, making the answers more objective and rational.
Q: What are the differences between MOSS and ChatGPT?
Answer: ChatGPT's training data covers a wide range and provides a good user experience. As an attempt in the academic community, MOSS aims to share more forward-looking theories and engineering experiences with the academic community by creating an open-source conversational language model.
Q: Is MOSS derived from the movie "Wandering Earth 2"? What are its future goals?
Answer: The name MOSS is related to the movie "Wandering Earth 2", in which the artificial intelligence robot MOSS exhibits extremely strong intelligence and rationality, becoming a powerful assistant to humans. We have seen the enormous potential of artificial intelligence in the development of conversational language models, so we named MOSS, which embodies our expectations for the future development of artificial intelligence technology.
The future MOSS will become increasingly intelligent. We will fully utilize the cloud computing power and resources provided by platforms such as volcanic engines, conduct model iterations and technical exchanges with more peers, continuously explore the technological frontiers of conversational and large-scale language models, and enable artificial intelligence technology to better benefit human society.
Artificial Intelligence Trainer: Making Machines Understand Humans More
Turn on the computer, input the collected sound data such as wind, rain, and stream sound, "clean" the mixed noise, and "train" the hearing aid data model to test its sensitivity in real scenes... Accompanied by the "clattering clattering clattering" sound of fingers tapping the keyboard, Tencent Teana Lab's artificial intelligence trainer Fu Cong's day of work begins.
In recent years, with the continuous development of artificial intelligence technology, this profession known as artificial intelligence trainers has gradually grown. As one of the "digital professions", the emergence of artificial intelligence trainers has accelerated the process of artificial intelligence from technological research and development to industry application, which will generate high economic and social value.
Continuously feeding data to the model
Every time he goes out, Fu Cong always wears a big "earring" on his ear.
This "earring" is actually a test version of a hearing aid. The sounds in earrings come in various forms, including whistling noise, sharp and piercing noise... These noises, amplified by hearing aids, have been a long-standing problem for many hearing-impaired individuals wearing hearing aids.
Fu Cong and his team are trying to use algorithm design and artificial intelligence technology to "train" data models, making hearing aids more "intelligent" in reducing noise, making hearing-impaired people hear clearly, understand, and feel comfortable.
Fu Cong explained that the data model of hearing aids is very small, so it needs to be optimized for different scenarios. Many scenarios are full of challenges, such as a hearing-impaired person eating in a restaurant, surrounded by many people talking, and wanting to chat with the opposite person. The sound around is particularly noisy, and as a normal person, one may not be able to hear clearly, let alone a person with hearing impairment. We hope to use the model to extract the necessary sound, reduce noise, and help more hearing-impaired people.
The ideal is full, but the actual process of developing model algorithms is like a repeated "battle".
The development process of the model can be roughly divided into the following steps: data collection, data cleaning, model training, scenario testing, and algorithm adjustment. After several iterations, testing and adjustment are carried out. If the test results are not ideal, this process needs to be repeated until the optimal effect is achieved. Fu Cong said.
Data collection should be targeted. In order to make the model smarter, it is necessary to collect various special data for different scenarios. Fu Cong and his team members not only need to go to the subway during rush hour, bustling restaurants, and crowded roads to collect hundreds of hours of sound data, but also need to wear hearing aids to experience the differences in these sounds. "For example, the wind may sound like a whirring sound to a normal person, but after wearing the hearing aids, it is very noisy, like going to a KTV to sing, and the sound hits the microphone hard.". To collect various wind noise data, Fu Cong recorded wind sounds in various scenes such as road cycling and seaside storms.
Data cleaning is the process of removing unwanted data. Fu Cong gave an example - the sound of the wind, which can be mixed with the sound of cars honking and people talking in real scenes. When organizing, these data should be excluded and kept as a relatively pure source of wind, so that the model can "recognize" the wind.
Model training is the process of feeding cleaned data to the model. In addition to the special data collected, Fu Cong and his colleagues will also include data such as languages from various countries around the world and some non speech sounds, which basically covers all the noise and speech encountered in people's lives.
Unlike humans, artificial intelligence models do not get tired, irritable, or lose their temper during the training process. Their intelligence depends on model parameters, training strategies, data volume, and so on. "They are like a 'child', becoming increasingly 'smart' and recognizing more and more sounds, which gives me a great sense of achievement," said Fu Cong.
Test patience, meticulousness, and endurance
After the model training is completed, it does not mean that it can be immediately applied to hearing aids for hearing-impaired people, and it also needs to go through a long process of iteration and adjustment.
For example, in order to provide suitable hearing aids for hearing-impaired individuals, the traditional approach is for patients to repeatedly go to offline fitting stores to try them on, which is a complicated process. Fu Cong explained that in general, hearing loss can be classified into three types based on the cause of the disease: sensorineural, conductive, and mixed hearing loss; According to the degree of hearing loss, it can be divided into mild, moderate, severe, and extremely severe hearing loss. The adaptation methods of hearing aids vary for different types.
Is it possible to move the adaptation process online and utilize artificial intelligence algorithms and deep learning capabilities to enable hearing-impaired individuals to perform accurate listening tests online? With this question in mind, Fu Cong began to develop adaptation algorithms. He likened this process to doing application problems, which require searching domestic and foreign literature, searching for existing solutions, using existing knowledge to carry out reasonable imagination, design experiments, and find answers based on specific usage environments.
This process tests the patience and meticulousness of artificial intelligence trainers. When testing the sound quality of hearing aids, different wearing methods correspond to different test results. Fu Cong and colleagues need to design different wearing methods in an "N x N" arrangement and repeatedly conduct experiments to study their impact on sound quality.
This process greatly tests the endurance of artificial intelligence trainers. "The basic literacy of an artificial intelligence trainer is to force oneself to listen to harsh sounds many times." Fu Cong said that this is because the trainer needs to quantitatively measure the sound limit points that hearing-impaired patients can hear normally, and the decibels of these sounds are unbearable for normal human ears. "Many times, I wish I could drop my headphones. After a day of testing, my whole head feels pain.".
After continuous iteration and adjustment, the built-in algorithm hearing aid has finally been completed. The most unforgettable thing for Fu Cong was their first time donating products to Shaoguan, Guangdong. They handed hearing aids to the deaf elderly one by one, turned them on, put on the equipment, adjusted the gain... "Although I had great confidence in the model, I still felt my heart in my throat because before that, the elderly could not communicate normally," Fu Cong said.
He cautiously asked an old man, "Can you hear what I'm saying?"
"Okay," the old man said three words slowly and firmly from his mouth.
"At that time, I felt that what we were doing was quite meaningful," said Fu Cong.
Using technology to solve human needs
Artificial intelligence trainers are a profession that needs to endure loneliness, as they spend a lot of time designing solutions, writing code, collecting data, and training models.
"My secret to overcoming loneliness is interest." Fu Cong's major is communication, many of which are related to signal processing. He usually likes music, so he combines his interests with his major and work, focusing on the audio field. After graduating from university, he participated in many works related to audio signal processing, experiencing various stages of audio algorithms from traditional algorithms to artificial intelligence algorithms, and then to large-scale deep learning.
In Fu Cong's view, artificial intelligence technology is a great tool aimed at liberating humans from a lot of mental labor and replacing currently costly individual labor in a large-scale manner. For the entire society, this is a progress in productivity with enormous social and commercial value.
What is a mature artificial intelligence technology like? Fu Cong believes that there are three stages to go through: first, perceptual intelligence, which focuses on simulating human visual, auditory, and tactile perceptual abilities, such as facial recognition, speech recognition, etc; The second is cognitive intelligence, which has characteristics such as human thinking comprehension, knowledge sharing, action coordination, or game theory. It can truly understand what people are saying and provide relatively complete answers based on some prompts; The third is behavioral intelligence, which, like autonomous driving, can truly play a role in the physical world.
To achieve this goal, it is necessary to continuously train the artificial intelligence model. Fu Cong stated that the first step is to prepare enough data for the problem, "as much as possible to cover all the situations encountered in solving this problem"; Secondly, it is necessary to design good algorithms and continuously optimize them based on user feedback.
"The field of artificial intelligence technology is advancing rapidly, requiring AI trainers to have broad perspectives, profound humanistic sentiments, and a sense of social responsibility. They should use the latest ideas, concepts, and correct ethics in the industry to help humans solve problems encountered in production and life." Fu Cong said.