Original source: Outshift by Cisco
This video from Outshift by Cisco covered a lot of ground. 8 segments stood out as worth your time. Everything below links directly to the timestamp in the original video.
This new AI model could dramatically simplify how businesses predict future trends and detect unusual activity in their operational data. Imagine an AI that can forecast critical system behaviors with unprecedented accuracy, making complex data analysis accessible without specialized effort.
Splunk to Release Advanced AI Model for Time Series Analysis
Splunk is preparing to release an advanced AI model designed to simplify time series prediction and anomaly detection, which will be available in its AI Toolkit by March. The model, which leverages a transformer architecture and extensive Splunk data, has demonstrated significantly lower mean absolute error on observability benchmarks compared to existing models. It also maintains balanced performance on general benchmarks, indicating its versatility beyond specific domains.
The upcoming release, with prediction capabilities launching in February and anomaly detection following in March, marks a notable advancement in handling complex time series data. This development is expected to streamline operations for data engineers and make sophisticated analysis more accessible, potentially transforming how industries approach data-driven insights and operational monitoring.
Cisco IQ Automates Level One Support with Data-Driven AI
Cisco is advancing its Cisco IQ platform to automate level one support tasks, including troubleshooting and compliance audits, by shifting from traditional expert systems to data-driven AI. This initiative, set for general availability later this year, integrates time series analysis with unstructured data using large language models (LLMs) and transformer architectures to develop comprehensive reasoning processes.
This evolution aims to provide IT operations with more intuitive and autonomous problem-solving capabilities, making network management more efficient. The integration of diverse data types and advanced AI reasoning marks a significant step towards networks that can dynamically understand and respond to complex issues, signaling a new era for network intelligence.
"Networks are becoming really sexy again thanks to LLMs and transformer architectures."
AI Enhances Time Series Forecasting with Synthetic Data and Transformers
Recent advancements in time series forecasting models are leveraging synthetic data to fill real-world data gaps and improve the robustness of predictions. Companies like Google, AWS, and Salesforce have been at the forefront, developing models such as TimesFM, Chronos, and Moira. These models incorporate enhanced transformer architectures, initially gaining prominence in language processing, to better understand and predict patterns in sequential data.
Cisco is also participating in this evolution, utilizing vast datasets from its acquisition of Splunk to refine these forecasting capabilities. The shift towards more sophisticated data handling and architectural improvements signifies a maturing field in AI, where foundational models are becoming more adept at navigating the complexities of real-world time series data for diverse applications.
AI Canvas Streamlines Network Troubleshooting with User-Guided Workflows
AI Canvas, currently in beta, is demonstrating how AI can streamline network troubleshooting by integrating user intuition into automated workflows. When an anomalous network activity alert is triggered in Splunk, AI Canvas ingests the case data and initiates a workflow, visually graphing the events. Users can guide the system with intuitions, such as correlating an event to login activities, which prompts the AI to generate Splunk Processing Language (SPL) queries and visualize new data.
This human-in-the-loop approach allows the system to identify issues like problematic user activity, as demonstrated by finding a user repeatedly running a specific curl command. AI Canvas can then automatically create alerts for such activities. The system enhances efficiency by automating query generation and data visualization, while still relying on human collaboration for mitigation and documentation, offering a blend of AI-driven automation and expert oversight.
Agent-Driven AI Pinpoints Root Causes in Kubernetes Applications
Researchers have developed a prototype for an agent-driven root cause analysis system designed for applications running on Kubernetes, capable of autonomously identifying critical issues. The system demonstrated its ability to iterate through multiple hypotheses and evaluate evidence over nine steps, ultimately pinpointing an "interface down" issue affecting a vote application. This iterative process, which includes forming dependencies and discarding incorrect paths, is tracked through an investigation graph and audit trail.
This AI-powered approach aims to mimic human troubleshooting by generating and refining hypotheses, even backtracking from unproductive lines of inquiry. The prototype illustrates a significant step toward automated problem resolution in complex IT environments, offering detailed audit trails and generating summary reports for management, thereby enhancing operational efficiency and transparency.
Deep Agents Aim to Automate Level One IT Support
The concept of "deep agents" is emerging as a solution to automate and offload entire level one IT support functions, mirroring human troubleshooting processes. These agents are designed to generate hypotheses, retrieve and evaluate evidence, correlate findings, and reformulate hypotheses to precisely identify and resolve problems. This initiative aims to embed the intuitive next-step decision-making that typically requires human insight directly into automated systems.
By systematically generating and evaluating potential causes for issues like a malfunctioning printer, deep agents can progressively narrow down the problem, moving from broad possibilities to specific components. This approach promises to significantly enhance the efficiency of IT support, reducing reliance on manual intervention for initial problem diagnosis and resolution.
Researchers Explore Single Large Model for All Time Series Use Cases
Researchers are investigating the feasibility of training a single, large model capable of addressing all time series use cases, similar to the approach used for large language models. The goal is to significantly reduce the cost and effort typically associated with developing specific models for each application. Initial attempts in 2023 involved training transformer architectures on massive datasets, yielding promising academic results, but these models struggled with the "dirty" and diverse frequencies of real-world data.
The challenge lies in moving beyond academic performance to robust real-world application, where data often lacks the clean structure assumed in benchmarks. This ongoing effort highlights the industry's aspiration to create a universal time series model, a "train once, infer many" paradigm that could revolutionize forecasting and anomaly detection across various sectors by making advanced analytical capabilities more accessible.
Cisco Patents Method for Enhanced Time Series Analysis in Transformers
Cisco has patented a novel method to enhance how transformer models process time series data, enabling them to remember longer sequences without exceeding context window limitations. This approach involves interleaving long-term historical data, presented as coarse roll-ups, with recent, fine-resolution data. This technique allows the AI to capture both overarching trends and immediate details, addressing a critical challenge in applying transformer architectures to time series analysis.
The innovation draws parallels with human memory, which retains broad historical context while focusing on recent events with high fidelity. This method, along with experimental approaches from companies like Datadog, which trained models on trillions of data points, signifies the rapid evolution of AI in handling complex, observational data. By optimizing how transformers interpret time series, Cisco aims to improve the accuracy of predictions and anomaly detection across various applications.
Also mentioned in this video
- The topic of telemetry in network operations, distinguishing between using… (0:08)
- The speaker details the classic approach to telemetry, involving data… (1:12)
- The speaker delves into the mathematical and architectural solutions for… (8:12)
- The speaker shifts focus to root cause analysis, explaining how detected alerts… (17:43)
Summarised from Outshift by Cisco · 31:52. All credit belongs to the original creators. Streamed.News summarises publicly available video content.
