SpikingVTG: Saliency Feedback Gating Enabled Spiking Video Temporal Grounding
Keywords: Video language models, Video temporal grounding, Efficient LMs
TL;DR: A low-powered spiking video language model framework for video temporal grounding tasks.
Abstract: Video Temporal Grounding (VTG) seeks to retrieve consecutive intervals or specific clips from a video based on specified natural language queries. VTG requires accurately aligning video segments with corresponding natural language instructions, highlighting the need for effective methodologies to capture semantic correspondence and maintain temporal coherence. Spiking neural networks (SNNs), previously underexplored in this domain, present a unique opportunity to tackle VTG challenges from both the architectural and energy-efficiency perspectives. In this paper, we leverage sparse spike-based communication of SNNs to propose a multimodal architecture tailored for VTG tasks, namely SpikingVTG, providing a biologically inspired and efficient solution. Leveraging temporal saliency feedback, our proposed spiking video-language model (VLM) achieves competitive performance with non-spiking VLMs across diverse moment retrieval and highlight detection tasks. We introduce a Saliency Feedback Gating (SFG) mechanism that improves performance while reducing overall neural activity. To efficiently train our spiking VLM, we analyze the convergence dynamics of each neuronal layer and utilize equilibrium states to enable training using implicit differentiation at equilibrium. This approach eliminates the need for computationally expensive backpropagation through time while also enabling the use of knowledge distillation for efficient model training. To further improve operational efficiency and facilitate the on-chip deployability of our model, we leverage a multi-stage training pipeline that focuses on eliminating non-local computations, such as softmax and layer normalization, leading to the development of the Normalization Free (NF)-SpikingVTG model. Additionally, we create an extremely quantized variant, a 1-bit NF-SpikingVTG model, which vastly improves computational efficiency during inference while maintaining minimal performance degradation from our base model. Our work introduces the first spiking model to demonstrate competitive performance on VTG benchmarks, including QVHighlights and Charades-STA.
Submission Number: 95
Loading