1. Tencent Cloud creates a one-stop industry large model selection store; 2. VideoComposer, a text-to-video diffusion model that is easier to use than Gen-2; 3. Another financial AI tool—BondGPT, a corporate bond investment assistant; 4. Tsinghua University The university proposed LiVT, which uses visual Transformer to learn long-tail data and improve model generalization capabilities; 5. Microsoft’s blessing to developers, integrating the “Generate Insights” tool in VS to improve compilation efficiency.
Today’s Highlights
1. Tencent Cloud creates a one-stop industry model selection store;
2. VideoComposer, a text-to-video diffusion model that is easier to use than Gen-2;
3. Another financial AI tool – BondGPT, a corporate bond investment assistant;
4. Tsinghua University proposed LiVT, which uses visual Transformer to learn long-tail data and improve the generalization ability of the model;
5. Microsoft’s gospel to developers, integrate the “Generate Insights” tool in VS to improve compilation efficiency;
Daily Wisdom AI
1. Go straight to the Tencent Cloud Large Model Technology Summit—Solve the Difficulty in Large Model Application
Tencent Cloud announced the launch of MaaS (Model-as-a-Service) large-scale model service, relying on Tencent Cloud TI platform to build industry large-scale model selection stores, based on Tencent HCC high-performance computing clusters and large-scale model capabilities, to provide customers with one-stop large-scale Model service.
At present, the difficulties faced by enterprises in the application of large models mainly include:
Fewer computing resources; large-scale model training and reasoning have high requirements for computing and storage resources, and the threshold is relatively high for many customers; data quality is poor; data is the basis for training large models, and low quality will lead to model failures. It is difficult to guarantee the training effect and efficiency; the input cost is high; the model also needs continuous optimization and debugging to adapt to the special functions of the enterprise; security compliance; data security is the most worrying issue for enterprises to use the model; there is a shortage of professional talents;
See Wisdom Comments:
From the perspective of protecting enterprise data property rights and privacy, Tencent Cloud provides customers with low-cost, convenient and fast large-scale model services. Combining its own advantages in computing power, Tencent Cloud breaks through the difficulties of enterprise application large-scale models from the dimensions of models, data, and applications. For different application scenarios, provide more suitable computing power network and intelligent application assistants, such as AI code assistants, conference assistants, etc. More importantly, we have observed that MaaS services can meet the diverse needs of customer model pre-training, model fine-tuning, intelligent development, etc., and support customers to join private domain data for training, which greatly solves the problem of large-scale model data security for enterprises. concerns.
Regarding how to solve the problem of large-scale enterprise application, which is the focus of attention in the industry, Wall Street Insights and Wisdom Research is very honored to invite the founder and CEO of Lanma Technology[Zhou Jian]to bring you the most core interpretation: the next important wave of AI On the track, how to break through enterprise applications and find the core profit secrets? Everyone is welcome to make an appointment for the live broadcast: Tuesday, June 20 at 19:00.
2. This text-to-video diffusion model VideoComposer is easier to use than Gen-2
VideoComposer, a diffusion model for text-generated video, enables simultaneous control of spatial and temporal patterns in various forms of synthetic video, such as text descriptions, sketch sequences, reference videos, and even simple hand-crafted actions. It seems to be stronger than the Gen-2 effect.
See Wisdom Comments:
VideoComposer enables users to compose videos with textual conditions, spatial conditions, and temporal conditions in a flexible way by introducing motion vectors in compressed videos as explicit control signals, combined with a spatio-temporal conditional encoder (STC-encoder). The method is able to effectively control spatial and temporal patterns, including textual descriptions, sketch sequences, reference videos, and handcrafted actions, etc. Experimental results show that VideoComposer has good performance and interaction-frame consistency.
This work makes significant progress in achieving controllable video synthesis by addressing the challenges of temporal dynamics and temporal consistency across frames, further advancing the development of customizable visual content creation.
3. Another financial AI tool – BondGPT, a corporate bond investment assistant
LTX, a subsidiary of Broadridge, an American financial technology company, recently announced the launch of BondGPT, a chat robot APP based on the GPT-4 model, for corporate bond investment. BondGPT is mainly for corporate bond investors, including hedge funds, traders, etc. It can answer various bond-related questions and help users solve related problems.
See Wisdom Comments:
The development of financial AI tools is currently very popular. In the AI News Weekly, we also analyzed that many financial institutions have begun to develop AI projects. The application scenarios of AI technology in the financial industry are also very rich, such as AI traders, financial public opinion analysis, assistant Code writing, research report knowledge base retrieval, etc. are all worthy of attention.
4. Tsinghua University proposed LiVT, which uses visual Transformer to learn long-tail data and improve model generalization ability
Tsinghua University’s CVPR 2023 paper, Learning Imbalanced Data with Vision Transformers, discusses in detail how to effectively use long-tail data to improve the performance of visual Transformers, and explores new methods to solve the problem of data imbalance in the real world.
Through a series of experiments, the article found that under the supervised paradigm, the visual Transformer will experience serious performance degradation when dealing with unbalanced data, while the visual Transformer trained with balanced distribution of labeled data shows obvious performance advantages. Compared with the convolutional network, this feature is more obvious on the visual Transformer. On the other hand, unsupervised pre-training methods do not require label distribution, so under the same amount of training data, Visual Transformer can exhibit similar feature extraction and reconstruction capabilities. Based on the above observations and findings, the study proposes a new paradigm for learning imbalanced data, aiming to make the visual Transformer model better adapt to long-tail data. Through the introduction of this paradigm, the research team hopes to make full use of the information of long-tail data and improve the performance and generalization ability of the visual Transformer model when dealing with unbalanced labeled data.
See Wisdom Comments:
The method of learning long-tail data with visual Transformer not only achieves significant performance improvement in experiments, but also does not require additional data, which is feasible for practical applications. For example, it can be applied in scenarios such as medical image analysis and security monitoring.
Data are often imbalanced, i.e. some classes have far fewer samples than others. As a result, traditional model training may be biased towards predicting a large number of categories, while the performance of a small number of categories degrades severely. By learning long-tail data with a visual Transformer, it can better adapt to real-world data imbalance problems. At the same time, it can effectively improve the recognition performance of minority categories, so that the model can better discover and understand sample features and improve the generalization ability.
5. Microsoft’s gospel to developers, integrate the “Generate Insights” tool in VS to improve compilation efficiency
Microsoft announced that in the latest version 17.7 of Visual Studio 2022, the “Generate Insights” tool will be integrated to improve developer efficiency. Microsoft announced that “Generation Insights” is now available in VS 2022, a tool that provides developers with in-depth insight analysis data, enabling developers to better understand and improve the compilation process.
The “Generation Insights” tool will issue a report after analyzing the compilation process, which will display the impact of “Generation Insights” analysis of each code variable on the total compilation time, allowing developers to intuitively see which specific codes have problems, It takes a lot of compilation time and can provide developers with some solutions to improve compilation efficiency.
It allows developers to have a deeper understanding of C++ development. By visually displaying the status of each part of the code at compile time, the tool allows developers to better understand the in-depth running process of C++, and by understanding the impact of each part of the code on the total compilation time, find out The method of optimizing the compilation process can improve the development ability to a certain extent while maintaining the code quality.
Risk Warning and Disclaimer
Market risk, the investment need to be cautious. This article does not constitute personal investment advice, nor does it take into account the particular investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, opinions or conclusions expressed herein are applicable to their particular situation. Invest accordingly at your own risk.