A.I. at the Crossroads of

NLP and Neuroscience

Workshop Description

AIxRoads is a one-day workshop to be held in Macao in August (10/11/12, 2019), in conjunction with the 28th International Joint Conference of Artificial Intelligence (https://ijcai19.org).

After many years of experimentation in research labs, Artificial Intelligence (AI) has moved into the arena of the real world. AI systems are currently used in many domains (e.g., medicine, finances and communication), outperforming humans in a broad range of acoustic-, visual- and natural language tasks.

This success may have several reasons: (a) the growth of computational resources and storage devices, (b) the availability of huge amounts of data (Internet), and (c) the development of smart learning algorithms. Progress is also due to the fact that researchers have managed to leverage and integrate discoveries made in disciplines that seemingly had nothing in common (Linguistics, Psychology, Mathematics, and Neuroscience). While being obviously different, these disciplines can nevertheless be complementary, an asset, allowing us not only to build artifacts, but also to gain a better understanding of the human information processor.

Finally, nature played an important role, as it inspired researchers by providing a model that, in order to be turned into sophisticated working solutions had to be understood, formalized and recast in engineering terms. Neural networks are a good example of this process as they represent a loose, yet very effective, imitation of the neural system. State-of-the-art neural architectures, such as Convolutional Neural Networks and Transformers, are directly inspired by biological (e.g., visual cortex) and cognitive (e.g., attention processes) models.

The goal of this workshop is to stimulate cross-fertilization between the different communities of the AI universe (e.g., Mathematicians, Linguists, Cognitive Scientists, Neuroscientists) in order to identify the knowledge needed to bridge the gap between Natural and Artificial Intelligence. More precisely, we would like to discuss whether and how the usage of knowledge concerning the human brain may enable engineers to produce better software.

Here are some of the questions for which we would like to find answers:

  • Can we get machines to learn as ordinary people do? Children induce rules on the basis of very few examples, containing even noisy data (few-shot learning ; learning of abstractions). Can we replicate this by a machine?
  • How can the learned knowledge be reused for new tasks?
  • How to improve the interaction between humans and machines?
  • Can knowledge of the brain mechanisms involving intelligence (sound, vision and language) help us to develop better architectures?
  • In what ways can the techniques developed in AI inspire cognitive scientists to get new ideas/theories, or, to help them to refine existing ones?
  • Is there a way for AI to exploit embodied representations ?
  • How can AI help us to solve problems in other disciplines, for example, NLP?
  • Can we make Natural and Artificial Intelligence cooperate in problem-solving, or, should the two be applied separately ?
  • If there is an interaction between the two, what should this look like? What are the interfaces and workflows?
  • What are the benefits for AI to mimic humans or the human mind while processing language?
  • Where in the development cycle and how shall AI engineers consider specific human aspects, such as the human brain/mind?
  • Specificities of humans and machines: how relevant is deep learning in modeling human thought?
  • Do we still need theories in the age of deep learning? Are there ways to interpret their results?
  • Is it possible to build a glass box and open the neural network black box?
  • What can NLP practitioners learn from network science (complex graphs)?
  • Can machines liberate us from the boring and mechanical aspects of problem-solving (logical proofs), allowing us to focus more on the creative aspects of the task?
  • How to build AI augmenting human intelligence, or, how to use human intelligence to augment AI?
  • Can we impose order and logic on an unordered set of ideas, by detecting the nature of the links between them automatically, to help authors in producing coherent texts?