Foundation Models are designed to serve as versatile embedding machines, with strong zero shot capabilities and superior generalization performance when fine-tuned on diverse downstream tasks. While this is largely true for language and vision foundation models, we argue that the inherent diversity of time series data makes them less suited for building effective foundation models. We demonstrate this using forecasting as our downstream task. We show that the zero-shot capabilities of a time series foundation model are significantly influenced and tied to the specific domains it has been pretrained on. Furthermore, when applied to unseen real-world time series data, fine-tuned foundation models do not consistently yield substantially better results, relative to their increased parameter count and memory footprint, than smaller, dedicated models tailored to the specific forecasting task at hand.

Transformer-based architectures achieved breakthrough performance in natural language processing and computer vision, yet they remain inferior to simpler linear baselines in multivariate long-term forecasting. To better understand this phenomenon, we start by studying a toy linear forecasting problem for which we show that transformers are incapable of converging to their true solution despite their high expressive power. We further identify the attention of transformers as being responsible for this low generalization capacity. Building upon this insight, we propose a shallow lightweight transformer model that successfully escapes bad local minima when optimized with sharpness-aware optimization. We empirically demonstrate that this result extends to all commonly used real-world multivariate time series datasets. In particular, SAMformer surpasses the current state-of-the-art model TSMixer by 14.33% on average, while having ~4 times fewer parameters. The code is available at https://github.com/romilbert/samformer.

A comprehensive understanding of interested human-to-human interactions in video streams, such as queuing, handshaking, fighting and chasing, is of immense importance to the surveillance of public security in regions like campuses, squares and parks. Different from conventional human interaction recognition, which uses choreographed videos as inputs, neglects concurrent interactive groups, and performs detection and recognition in separate stages, we introduce a new task named human-to-human interaction detection (HID). HID devotes to detecting subjects, recognizing person-wise actions, and grouping people according to their interactive relations, in one model. First, based on the popular AVA dataset created for action detection, we establish a new HID benchmark, termed AVA-Interaction (AVA-I), by adding annotations on interactive relations in a frame-by-frame manner. AVA-I consists of 85,254 frames and 86,338 interactive groups, and each image includes up to 4 concurrent interactive groups. Second, we present a novel baseline approach SaMFormer for HID, containing a visual feature extractor, a split stage which leverages a Transformer-based model to decode action instances and interactive groups, and a merging stage which reconstructs the relationship between instances and groups. All SaMFormer components are jointly trained in an end-to-end manner. Extensive experiments on AVA-I validate the superiority of SaMFormer over representative methods. The dataset and code will be made public to encourage more follow-up studies.