Data science neural network project life cycle
Download as PPTX, PDF0 likes171 views
This document outlines the machine learning model development process. It involves getting and cleaning data, defining a neural network strategy, splitting the data into training, validation, and test sets, training and evaluating the model through multiple iterations to reduce underfitting and overfitting, improving the model through techniques like adding more data or adjusting hyperparameters, and ultimately deploying the optimized model into a production system. The goal is to develop a model that generalizes well to new real-world data through an iterative process of training, evaluating performance, and making improvements.
Data science neural network project life cycle
- 1. Explore data Code one or multiple models Experiment and train the model 1.Evaluation & error analysis Balance between underfitting & overfitting Improve the model performance 1.Clean incorrectly labeled data 1.Get more data or synthesize artificial data •Define a Neural Network strategy •NN, CNN, RNN, … Business problem •Clean •Format •Label •Pre-process Get Data •Training •Cross- validation •Test Split Datasets with as many cycles (epochs) as needed Underfitting (high bias) Overfitting (high variance Multiple if the infrastructure allows it(if needed) (if needed) Expose model to real world data Real World data Insert model into production system Production Tune model’s hyper- parameters Keep tuning Start Customer Acceptance Many sources Streaming vs Static Structured vs Unstructured Model accuracy is high? Start iterate here as many time as needed Regularization, Dropouts, Hyper-parameters, Network architecture, Optimizers & Loss function, ... Transfer learning from existing model? Training loss Validation loss Accuracy Vincent Pommier, Sept 2018