Data Machina #261

2019-03-12 · Source: Data Machina · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Advanced, short

Summary

Recent advancements in Generative AI and time-series (TS) forecasting are highlighted through six new research papers from major organizations. JPMorgan's LETS-C method uses language embedding models for lightweight TS classification, achieving state-of-the-art accuracy with 14.5% fewer parameters. Mercedes-Benz introduces TeVAE, a temporal variational autoencoder for discrete online anomaly detection in multivariate TS data, identifying 65% of anomalies with only 6% false positives. Datadog's Toto is the first general-purpose TS forecasting foundation model for observability data, trained on one trillion data points, outperforming existing models. Tesco applies a transformer architecture with multiple-resolution tokenization for price optimization. ViTime, a novel Vision-Language Model, transforms numerical TS into binary images for zero-shot forecasting, surpassing supervised models. Finally, TimeS from another study utilizes the T5 transformer for short-term stock price "excitement prediction" by integrating stock, event, and sentiment data.

Key takeaway

For AI Engineers and Research Scientists developing forecasting or anomaly detection systems, these advancements indicate a shift towards integrating generative AI and large language models. You should explore methods like language embedding for TS classification or visual intelligence models for zero-shot forecasting to improve accuracy and efficiency. Consider adapting transformer architectures and specialized VAEs for domain-specific challenges like price optimization or industrial anomaly detection.

Key insights

Generative AI and LLMs are being adapted to achieve state-of-the-art performance in diverse time-series tasks.

Principles

• Embeddings bridge language and time series data.
• Foundation models benefit from massive pre-training.
• Visual transformation enhances time series analysis.

Method

Methods include tokenization, base model selection, and prompt engineering to re-program LLMs for time series. Others involve embedding time series with language models, using VAEs for anomaly detection, or transforming TS into images for VLM processing.

In practice

• Use LETS-C for efficient TS classification.
• Apply TeVAE for industrial anomaly detection.
• Consider Toto for observability TS forecasting.

Topics

• Time Series Forecasting
• Time Series Classification
• Anomaly Detection
• LLMs for Time Series
• Vision-Language Models

Code references

• ashishpatel26/500-AI-Machine-learning-Deep-learning-Computer-vision-NLP-Projects-with-code
• virattt/financial-agent-ui
• fferegrino/cycle-station-predictions

Best for: AI Engineer, AI Scientist, Research Scientist, AI Researcher, Machine Learning Engineer, Data Scientist

Related on AIssential

• Speaking Numbers to LLMs: Multi-Wavelet Number Embeddings for Time Series Forecasting — TempoWave improves LLM time series forecasting by aligning continuous numerical data with discrete language models using multi-wavelet embeddings.
• The brewing GenAI data science revolution — Foundation models are extending beyond text/image to transform tabular and time-series data analysis.
• EVIL: Evolving Interpretable Algorithms for Zero-Shot Inference on Event Sequences and Time Series with LLMs — LLM-guided evolution can discover compact, interpretable algorithms for zero-shot dynamical systems inference, outperforming complex neural networks.
• TimeVista: Exploring and Exploiting Vision-Language Models as Judges for Time Series Forecasting — Vision-Language Models reliably judge time series forecasts, aligning significantly better with human preferences than traditional metrics.
• Representing Time Series as Structured Programs for LLM Reasoning — Representing time series as structured symbolic programs enables LLMs to reason effectively without fine-tuning.
• Moving To Substack — The Transformer architecture is fundamental to modern generative AI, enabling advanced language model capabilities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Data Machina.