.png&w=256&q=75)
🤓 Latest Submissions
Wavelet-TSLM: Time Series Foundation Model
🏭 1. The Challenge (Track 1: Intelligent Industry) Industrial telemetry (turbines, power grids, factory sensors) generates continuous streams up to 131,072 steps. Existing time-series LLMs (AnomalyXL, OpenTSLM) discretize signals into text tokens, causing heavy quantization error and context explosion. On the standard benchmark, AnomalyXL collapses to a near-zero localization IoU of 0.008. ⚡ 2. The Solution: Wavelet-TSLM Wavelet-TSLM bridges continuous signal processing with LLM reasoning without quantization loss: Continuous Haar Wavelet Pyramid: 6 dyadic levels preserving macro-trends and micro-transients at native float resolution. 16-Frequency Fourier Tokenizer: Continuous float64 projection directly into Qwen's latent space. Dual-Stage Zoom & Pointer Readout: Global scanning followed by 32x physical zoom with a continuous pointer head, completely eliminating hallucinated text timestamps. 🚀 3. Native AMD ROCm & Instinct MI300X Acceleration Engineered for AMD Instinct MI300X (gfx942) with ROCm 6.4 and PyTorch 2.6: Triton FlexAttention (512-pad): Slashes attention memory from 65.4 GiB to 3.5 GiB at 16k context (18.7x reduction). Wavelet Checkpointing (CHECKPOINT_RAMI): Cuts activation memory by 4.2x (3.3 → 0.78 GiB), enabling 16-channel telemetry on a single GPU. 📊 4. Benchmark: Wavelet-TSLM vs AnomalyXL (ETH Benchmark) Localization IoU: 0.6034 vs 0.008 (+75x improvement) High Precision (IoU ≥ 0.5): 71.3% (112/157) vs <1% Ultra Precision (IoU ≥ 0.7): 49.0% (77/157) vs 0% Median Center Error: 0.0088 L (sub-1% error) vs >15% Specificity (Normal telemetry): 98.0% (zero false-alarm spam) Multichannel F1 Score: 0.8986 vs 0.038 (+23.6x) 💼 5. Business Value & Industrial Impact Prevents catastrophic downtime ($500k+ per failure) by delivering sub-second, actionable root-cause work orders for plant maintenance. Open source and AMD ROCm ready.
18 Oct 2026
