Model Retraining

Retrain AI models on your own data to achieve peak performance

Models need tailoring. That's why Modzy platform gives you tools to retrain models on your custom data.

While humans have a natural ability to transfer knowledge and experience across different tasks and domains, machine learning models still struggle with generalizability. This means that countless hours and resources are spent training models from scratch to perform against slightly different datasets, rather than leveraging existing models and retraining them to be more generalizable. Transfer learning refers to the concept of improving a machine learning model’s performance by transferring the knowledge acquired during training to a different, but related, dataset. In simpler terms, transfer learning allows you to quickly retrain a model, that was already trained, on your own dataset, which makes the model more performant on your dataset.

Retraining at Modzy

At Modzy, we’ve pioneered a novel approach to retraining models based on transfer learning that allows data scientists to quickly and easily retrain models on their own data for improved performance, enabling major time and cost savings.
  • Our retraining solution, which is different than simply training a model from scratch, is engineered and designed to consume as little computational power in as short a time as possible given the size of the target dataset and task.
  • Our solution allows for retraining of object classification models, a solution to a much trickier problem. Retraining image classification models is easy, and solutions have existed for this problem for years. 
During training, we use large datasets from a wide range of distributions to ensure that our models can make correct predictions across diverse scenarios. Our retraining solution improves performance and minimizes the requirement for a large labeled dataset, which reduces the amount of computational resources and time required for training.

Modzy Retraining Highlights

Today, our solution allows for retraining a model on a dataset of 1600 images on one GPU in 16 minutes

Retrained models detect people at a 45% mean average precision (as compared to a 5% benchmark), and vehicles at a 76% mean average precision (as compared to a 20% benchmark)

Problems with Existing Approaches to Retraining

There are several challenges with existing approaches to retrain models :  

  • Models needs to be trained on large dataset of labeled data points to be generalizable
  • Collecting a sufficient sample of labeled training data is expensive, time-consuming, or may be impossible for some scenarios
  • Current approach primarily involves training new models from scratch to do slightly different tasks, with high infrastructure and data curation costs 

Modzy’s approach to retraining offers a solution to many of the challenges with existing workaround solutions to retraining, enabling resource efficiencies and better performant solutions.

Get Details on Quick Start Program