How do generative adversarial networks (GANs) differ from conventional neural networks?
Conventional generative machine learning algorithms learn the most likely prediction given the input signal. Maximizing likelihood of the prediction encourages smooth or blurry predictions that represent a compromise between competing labels or missing information. GANs maximize realism of the predictions by training two or more networks that work in opposition. A generator network produces realistic predictions and a discriminator network differentiates real and fake output signals. The generator and discriminator networks compete adversarially until the fake predictions are so real that the discriminator can no longer differentiate between real or fake.
GANs were first introduced for data augmentation but are now broadly adopted for generative text and image modeling. Retrace is proud to have an expansive catalog of generative dental AI solutions that function under the hood of our core products.
How do adversarial learning and the actor-critic concept differ?
Conceptually, GANs are closely related but very different to the actor-critics concept in reinforcement learning. Games such as chess and go are usually paired with reinforcement learning solutions. Previous gameplay solutions train neural networks to play a style of chess that is stylistically indistinguishable from the gameplay of a specific real game player. Controlling style also helps mitigate modal collapse which occurs when the network learns to predict a single style that maximizes realism. This might be problematic if the problem demands highly diversified predictions which is often the case during the imitation phase of many reinforcement learning problems such as playing chess, go, or Starcraft. This problem is also present in smart dentistry, when imitating a dentists decision making tendencies in the clinic.