Google has released the TimesFM model, a foundational artificial intelligence model for predicting numerical data, such as when users input basic numbers, the model can predict the next value without the need for prior training data.
To train this model, Google utilizes real-world data sets of over a hundred billion, including Google Trends and Wikipedia page views, along with synthetic data that represents basic mathematical or physics models.
The performance evaluation is conducted using the Monash Forecasting Archive, which consists of tens of thousands of data records like traffic volume and weather conditions. TimesFM outperforms models that require upfront training by using zero-shot learning, even outperforming GPT-3.5 in prediction accuracy, despite its smaller model size of only 200 million parameters.
The model is available for download from HuggingFace.
Source: Google Research
TLDR: Google’s TimesFM model makes accurate numerical predictions without the need for prior training, outperforming larger models like GPT-3.5 in certain scenarios. It is available for download from HuggingFace.
Leave a Comment