Google DeepMind has unveiled a new AI training method called ‘JEST.’ This innovative approach can shorten AI model training time by 13 times and reduce the required computing power by 10 times. The introduction of this technology is expected to transform AI training into a faster and more efficient process.
Innovative Approach of JEST
JEST maximizes AI training speed and efficiency through ‘Multimodal Contrastive Learning.’ Traditionally, data was selected randomly or based on the relevance of individual data points. However, JEST optimizes training data by selecting ‘batches’ that include high-quality data.
JEST revolutionizes the way training data is selected. Instead of randomly selecting data, it evaluates and ranks data quality using a pre-trained model. This allows the selection of batches that include high-quality data to maximize the performance of AI models.
Importance of High-Quality Data
JEST’s success hinges on the curation of high-quality data. Google DeepMind researchers emphasize that the performance of JEST, pre-trained with a small set of high-quality data, determines the overall performance. This model then filters larger datasets to identify high-quality data.
High-quality data is the key factor in determining AI model performance. JEST carefully selects data through a smaller model, which is then used to train multimodal AI models, creating more efficient and robust models.
Energy Efficiency and Environmental Protection
JEST optimizes the energy-intensive AI training process to reduce energy consumption. This contributes significantly to environmental protection. Just as Microsoft’s water consumption surged by 34% due to increased AI computing demand, AI training requires enormous energy. JEST demonstrates the potential to significantly reduce this energy consumption.
Future Potential of JEST
JEST has opened new horizons for AI training methods. This approach will enhance the efficiency of AI model training, enabling more companies and researchers to effectively utilize AI technology. The future impact of JEST on the AI field is highly anticipated.
Source: decrypt.co