Beyond the 'black box': Toward AI that is both generative and explainable

The latest breakthroughs in artificial intelligence (AI) build on deep neural networks, a specific type of AI system. Today, their applications are known to the general public in several areas, in particular, the so-called large language models capable of producing human-resembling text and conversations (ChatGPT, Bing AI, Bard) and the text-to-image generative models which can produce striking images from a text captions (DALL-E 2, Imagen, Stable Diffusion, Midjourney), as well as others connected to the recognition, production and translation of speech and sound. What is behind the success of these techniques? The domains in which those successes happened are not haphazard or random, but rather perfectly suited for the strengths of neural networks as they exist today. Indeed, neural networks are proficient at emulating complex actions, even when they are difficult to define in precise terms, as long as we can feed them enormous amounts of examples to learn from.