Researchers at MIT have discovered a mechanism to identify photos using synthetic data that, according to them, can compete with models trained on actual data.
Researchers developed a specific sort of machine learning model to generate exceptionally realistic synthetic data, which was then used to train another model to perform vision-related tasks as part of the research.
According to the researchers, huge volumes of data are now necessary in order to train a computer to execute picture classification tasks, such as spotting damage in satellite photographs taken during a natural catastrophe, among other things. The datasets necessary to train the model, on the other hand, can be generated at a cost of millions of dollars.
According to the researchers, their specific machine learning model, known as a generative model, uses far less memory to keep and distribute than a dataset, and so requires significantly less computing power.
A comparison was made between the results of their learning model, which was trained solely on this synthetic data, and the results of many other picture classification algorithms, which were trained on actual data. The findings imply that their strategy can occasionally outperform the other models when it comes to learning visual representations.
“We were certain that this strategy would ultimately succeed; all we had to do was wait for these generative models to become better and better,” said Ali Jahanian, the study’s main author. The researchers were particularly impressed when they demonstrated that their approach may occasionally do even better than the genuine thing.
AI education and training AI
This approach entails giving the generative model millions of pictures that contain items in a certain class, like as automobiles or cats, and then teaching it how to recognise these things so that it can produce objects that are similar to them. This process is known as training.
According to the researchers, these generative models may also learn how to change the data that they are given. For example, if it has been trained on photos of vehicles, it will be able to ‘imagine’ how a car might seem in a variety of settings that it did not witness during training. It could then generate graphics of an automobile in a variety of positions, colours, and sizes, among other things.
A approach known as contrastive learning, in which a machine learning model is given several unlabeled photos in order to learn which pairs of images are similar and which pairs are different, relies on having many views of the same image.
A contrastive learning model was developed when the researchers coupled a pretrained generative model to an item identification model. The contrastive learner could tell the generative model to generate distinct pictures of an object and then learn to recognise that object from different perspectives.
“It was like putting two construction pieces together,” Jahanian explained. As a result of the fact that the generative model may provide us with several perspectives on the same issue, it can assist the contrastive technique in learning better representations.
According to Jahanian, there are certain restrictions to employing generative models since they might expose source data, which can pose a risk to people’s personal information. If they are not adequately checked, they may potentially increase biases in the datasets on which they are trained. It is the intention of the study team to resolve these limitations in future investigations.