Reading time: 5min
Companies need to make use of their data with Artificial Intelligence (AI) to stay competitive: predicting sales to better allocate resources, detecting fraud to identify liabilities or simply to better understand customers by analyzing their behavior.
AI models are built on big piles of data to make better decisions and massively impact companies top- and bottom-line. The problem is, that developing those models is cost-intensive and requires specialized data science skills.This might prevent companies from tapping into their data’s potential. That’s why using already existing AI models—so called “pre-trained” AI models—is on the rise as it reduces development and training time, raises performance and allows access to models without the necessity of a large amount of data.
No one wants to reinvent the wheel. So why should we always create a new AI model from scratch when we want to solve a problem? Instead of developing your own AI model, you can use one that has already been trained on a similar problem with another data set. This will give you faster results because the existing AI model already “learnt” from another data set.
Take for example ResNet, a model used for image recognition: ResNet is trained on 1.28 million hand-annotated images from the ImageNet 2012 data set. This data set was created for the ILSRVC challenge. Countless hours went into creating this data set, developing the ideal algorithm setting, and the training of the final model. If you have a similar problem in the area of image recognition, it would help you tremendously to tap into the potential of this meticulously designed AI model. Using a pre-trained model like ResNet for a specific business problem may not yield sufficient results on its own but it is a head-start compared to a model from scratch.
We like to think of three different AI model types:
Type1 – single solution models: These models are created with one problem in mind, and one problem alone. The AI models is developed and deployed by a company to solve their business problem and it is likely to do so with a sufficient quality. Using such an approach creates high development cost and needs time to be developed.
Type 2 – pre-trained general-purpose models: These are models that are trained for a very general task. They come in very handy for tasks like image classification (e.g. ResNet) or natural language processing . They perform great on the general task they are trained on and are relevant for a very large user base. Those AI models can give you a head start when you develop a single solution model for a business problem that includes one of these general tasks. However, most business problems are very specific which prevents you from using those AI models.
Type 3 – specialized pre-trained models: These models are developed for a specific business problem many companies face. The model is trained to solve a specific business problem but is shared among all companies that face the same challenge (e.g. fraud detection or sales forecast). The quality of a specialized pre-trained model to solve a business problem is greater than a single solution model because it can potentially be developed on a greater database. Cost per user compared to a single solution models can be reduced because many companies can share the cost for development.
Existing companies that face a similar business problem could “simply” put all the required data in one central database and train an AI model together. But of course, for strategic, legal and ethical reasons we cannot and do not want to do that.
That is why we help companies with similar analytical problems to leverage transfer learning—a technology that allows the development of specialized pre-trained models. With the help of transfer learning we take one AI model, that has already been trained on data to solve a business problem and retrain it on a new data set for a similar problem. Additionally, we make sure that the AI model contains no information of the data it has been trained on. That means, we transfer only the abstract AI model, but not the data itself. The performance achieved by the transferred AI model potentially yields a better performance because the model is trained on a large and more diverse data set. Also, the development time of the AI model on the new data set is much shorter.
We believe by using transfer learning there is a way to combine the best of both worlds. With specialized pre-trained models, we can increase the user base compared to single solution models which reduces individual user cost, while at the same time increase the performance of the model.
Did we spark you interest? Get in touch: firstname.lastname@example.org