Automated Deep Neural Network Architecture Selection for Time Series Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for building time-series prediction models, especially with complex data, face challenges in accuracy due to multi-level seasonality, high correlation, and noise, and lack automation in selecting optimal deep neural network architectures.
Innovation Solution
A cloud-based system and method for automatically synthesizing and training optimized deep neural network architectures for time-series predictions by selecting existing architectures, replicating them, modifying parameters, and determining the fittest model through parallel training, utilizing a hardware processor and meta-learning techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If traditional manual methods are used to build time-series prediction models, then model customization and control are possible, but automation and efficiency are reduced
Solution Approach 1:
The system enables self-service automated model building where the platform automatically selects architectures, tunes hyperparameters, and trains models without manual intervention. The automated DNN architecture selection process allows the system to serve itself by autonomously navigating the complex model selection space and delivering optimized predictions.
Solution Approach 2:
The platform provides a universal automated machine learning system that handles multiple time-series prediction tasks across different domains (IoT, retail, transportation, energy). It consolidates various functions including architecture selection, hyperparameter tuning, model training, and deployment into a single multi-functional system that works across diverse use cases.
2Measurement precision
If simple neural network architectures are used, then training speed and resource consumption are reduced, but accuracy on complex time-series data is insufficient
Solution Approach 1:
The system automatically changes DNN architecture parameters (number of layers, neurons per layer, activation functions, learning rates, batch sizes) based on the specific time-series data characteristics. This parameter optimization allows the model to achieve high accuracy on complex multi-level seasonal data while the automation manages the complexity of exploring different parameter configurations.
Solution Approach 2:
The DNN architecture is dynamically selected and adapted rather than fixed. The system evaluates multiple candidate architectures and dynamically chooses the optimal configuration for each specific time-series task, allowing the model structure to adapt to the underlying patterns in the data while the automated process manages the complexity of this dynamic selection.
3Reliability
If multiple DNN architectures are trained and evaluated, then model accuracy and suitability are improved, but training time and computational resources increase
Solution Approach 1:
The system performs preliminary actions by pre-selecting a curated set of candidate DNN architectures that are most likely to perform well on time-series data before the actual model training begins. This preliminary architecture selection based on meta-learning from previous tasks reduces the number of architectures that need full training, thereby improving deployment speed while maintaining model suitability.
Solution Approach 2:
The system uses copying by replicating successful architecture patterns from previously solved tasks. Through meta-learning, it copies effective architectural configurations from the repertoire of learned patterns and applies them to new time-series problems, reducing the need to train entirely new architectures from scratch and thus improving productivity while ensuring reliability through proven patterns.
4Adaptability or versatility
If existing DNN architectures are used without modification, then implementation is faster and easier, but adaptability to specific time-series patterns is reduced
Solution Approach 1:
The system makes the DNN architecture dynamic and adaptable by automatically modifying architecture parameters based on the specific time-series patterns detected in the data. The automated process dynamically adjusts layer configurations, neuron counts, and hyperparameters to adapt to seasonal patterns, trends, and noise characteristics while maintaining ease of use through full automation of this adaptation process.
Solution Approach 2:
The system applies local quality by customizing specific parts of the DNN architecture based on the local characteristics of the time-series data. Different segments of the architecture (input layers, hidden layers, output layers) are selectively modified to match local patterns in the data such as seasonal periodicities or trend components, while the automated system manages the complexity of these localized adaptations.
Data Source
AI summary
A system and method for automatically generating deep neural network architectures for time series prediction. The system includes a processor for: receiving a prediction context associated with a current use case; based on the associated prediction context, selecting a prediction model network configured for a current use case time series prediction task; replicating the selected prediction model network to create a plurality of candidate prediction model networks; inputting a time series data to each of the plurality of the candidate prediction model network; train, in parallel, each respective candidate prediction model network of the plurality with the input time series data; modifying each of the plurality of the candidate prediction model network by applying a respective different set of one or more model parameters while being trained in parallel; and determine a fittest modified prediction model network for solving the current use case time series prediction task.