![]() IWSSL-22: July 28th & 29th, 2022The Third Annual International Workshop on Self-Supervised Learning |
||||||||||||
Home | Program | Venue | ||||||||||||
Future artificial intelligence (AI) systems must be able to adapt to new environments, be more robust in the face of complexity and uncertainty, and improve their skills incrementally, in synchrony with their experience in the real world. Their control and learning strategies, whatever form they may take, may rely on a variety of knowledge representation schemes to address incremental, open-ended learning and cognitive development, and must be able to handle mistakes - and learn from them. They must be able to learn not only their operating domain but also the limits of their own knowledge. Knowledge acquired through active exploration or mimicking must be more robust than hand-crafted knowledge and control systems. Self-supervised learning is squarely on our path towards more general machine intelligence; time will tell whether self-motivated learning will fuel the next wave of intelligent systems or the one after it, but its arrival is inevitable. In the rush to produce high-performing 'intelligent' systems (e.g. problem solvers and classifiers) with immediate appliation, AI system creation has significantly relied on hand-crafted frameworks, coded using human-readable programming languages (e.g PDDL, propositional logic, etc.). These systems are notoriously brittle as the number of rules tends to be relatively small and overly rigid, underspecifying even moderately complex tasks and environments. Modern machine learning methods are no exception - they rely extensively on hand-crafting and close supervision, leaving unaddressed many cognitive processes required for AI systems with broader, more general skills. In short, current AI paradigms seem inadequate for realizing the advanced intelligent robotic systems of the future that the founding fathers of AI hoped to see. Animals learn about themselves and their environment, and how the two interact effectively and efficiently, through self-motivated exploration and experimentation. High-level learning, e.g. the use of specialized tools, can be learned both by example and by building upon existing learned competencies. Approaches exploring mixed learning mechanisms in open worlds in AI has for the longest time seemed out or reach. An underemphasis on important topics such as reasoning, cumulative (life-long, incremental) learning, and cognitive development, has lead fundamental mechanisms of thought to receive less attention than they deserve, including attention, analogy making, and creativity. A self-supervised learner in a complex, open world must possess a variety of cognitive processes - including these, and more - to allow it to learn - and even invent - a variety of new skills, develop new cognitive capabilities, compare situations and objects for various purposes, and how to use those objects as tools, through exploration. To make sure such systems materialize within the next decades we must revive the original goals of the field by raising the bar and broadening the range of target cognitive features existant in AI systems. Participants in IWSSL-22 will discuss the development and evaluation of self-supervised learners, setting a research agenda for a wide-horizon long-view developmental approach to future intelligent systems and robots. Venue
View tentative programRegistration
Papers & PublicationAll participants are invited (but not required) to submit a paper before the workshop to stimulate discussion. This workshop will focus on the future of AI - we encourage you to consider submitting a position statement on the field's future, thoughts on fulfilling the original goals of the founding fathers, commentary on the state of AI, and/or relevant methodological considerations for the coming decades.
Important Dates
Organzing Committee
Steering Committee
Local Organizers
If you have any questions, please send an email to organizers.iwssl22@gmail.com as soon as possible.
|