Extend pre-trained models beyond their original purpose by fine-tuning them to your use case.

With fine-tuning, you take a pre-trained computer vision or natural language model and replace the output layer of the neural architecture with a custom, task-specific layer that is trained on your own dataset. This way you leverage the potential and extensive training of pre-trained AI models, while still allowing for customization.

Less training time.

A pre-trained AI model has already seen a lot of text or images already and has adjusted the model to reflect the information it has seen. As a result of that, training the outer layer of the model takes significantly less time, because they only need to be tuned to take in the “base” of the pre-trained model, and fit it to a specific task.

Less data.

Because the base information is already present in the AI model, it needs a lot less task-specific data to start performing well as compared to custom training a model from scratch. If you have a specific task you need to tackle and don’t have an abundance of data, fine-tuning might be what you’re looking for.

Less configuration.

Fine-tuned models have been shown to perform as good, or even better than custom architectures that are often very complex and made to work only on a specific task. Simply fine-tuning of a pre-trained model with minimal task-specific adjustments will allow you to create faster models, built on an architecture that you can easily extend and adjust to your needs.

When applying AI, we take the following approach: we evaluate your exact use case and then see what type of model you can use to reach your goals. For models, we have a hierarchy. In the best-case scenario, your use case is fulfilled by just pre-trained models without any adjustment or configuration. These models are as stable and efficient as possible, need the least amount of work, and save you the most amount of effort in the end. At the end of the line, we make use of custom training. You can completely customise the model to your needs, but need a lot of data to achieve your goals and aspirations. Next to that, there are custom-trained models need a lot more maintenance. Fine-tuning a pre-trained model is our in-between solution. You can get started (relatively) quickly, and have more options to customise. It solves cons of pure pre-trained models and custom-trained models, and is therefore a good option when you’re in the middle of the spectrum.

To be able to fine-tune pre-trained models, you need data. It might sound like fine-tuning will still get you out of the data collection step, but like always in AI and data science, the more the merrier. Immediately start collecting data, preferably in a structured way, so that you can adjust your pre-trained models with task-specific data. Then, like how we described it in our top tips on using pre-trained models, pick a pre-trained model that is pre-trained on data that is closest to your use case.