From Simple Competences to Intelligent Human-Compatible AIs
To accommodate participants, the paper submission deadline has been extended to December 7.
A two-day Developmental AI workshop will be held starting on February 26-27, 2024, at AAAI 2024 in Vancouver, B.C. to discuss new findings and open challenges for creating developmental AIs. Developmental AIs start from innate competences. They learn by interacting with objects in the environment, including people and other AI agents.
The mainstream approaches for creating AIs are the deep learning AI approaches (e.g., generative LLMs) and the traditional symbolic AI approach. These approaches have led to valuable AI systems and impressive feats. However, manually constructed AIs are generally brittle even in circumscribed domains. Generative AIs make strange mistakes and do not notice them. In both approaches the AIs cannot be instructed easily, fail to use common sense, and lack curiosity. They have abstract knowledge but lack social alignment.
The promise of developmental AIs that they will acquire self-developed and socially developed competences like people do. They would address the shortcomings of current mainstream AI approaches, and ultimately lead to sophisticated forms of learning involving critical reading, provenance evaluation, and hypothesis testing. Developmental AI approaches have demonstrated capabilities including visual and multimodal perception, object recognition and manipulation. Powerful models for hierarchical planning, abstraction discovery, curiosity, and early language acquisition exist but need to be adapted to a developmental learning based approach.
However, developmental AI projects have not yet passed the capabilities of young children. The AIs have not bridged the communications gaps to interact with people or to skillfully and skeptically draw on online information resources like those that power today’s LLMs. The research needs to bridge competence gaps involving nonverbal communication, speech, reading, and writing. This workshop is about next research. It is about the gaps, challenges, and prospects for creating developmental AIs that are robust, resilient, and human-compatible.
Topics
Topics are organized in two groups corresponding roughly to science and engineering (including bio-inspired reverse engineering).
The Science of Developmental AI (“science”)
- How kids learn and think: findings and experiments from neuroscience, psychology, machine learning, AI, and research on how educational techniques promote learning.
- Testing competence boundaries and evaluating learning by developmental AIs.
- Efficient and ethical participation by humans in training and testing AIs.
- Learning environments for ask-focused teaching, multi-agent free play, training paradigms.
- Perceptual grounding, cognitive grounding, and common grounding.
Bootstrapping Developmental AI (“engineering”)
- Machine learning for embodied AIs.
- Simulators and robot platforms for training embodied AIs.
- Tradeoffs in cognitive bootstrapping trajectories and curriculum design.
- Communication and linguistic competences for speech and reading.
- Techniques for ensuring that developmental AIs learn human-compatible values and drives.
Other topics: Should developing AIs grow physically as they mature cognitively? How should research on bootstrapping developmental account for economic and social issues and opportunities for ubiquitous embodied AIs.
Format
This workshop will include keynote presentations, short research presentations, and interactive working sessions. The workshop will run from 9:00AM to 5:00PM on the workshop days (February 26 -27, 2024).
Attendance
Attendance will be limited to about 40 participants.
This will be a working workshop – with cross-cutting multidisciplinary working groups. Participation is open to established researchers and to students interested in advancing this area and the virtuous research cycle with adjacent research areas.
Submissions
Submissions are due December 7, 2023 (extended). They should be electrically uploaded at the Open Review web site. Submissions can be uploaded via the OpenReview BDAI website https://openreview.net/group?id=AAAI.org/2024/Workshop/BDAI.
Submissions can be in a long form or a short form.
- Long Form Submissions — a technical paper or position paper of 4-6 pages. These papers will be considered for a subsequent publication after the workshop.
- Short Form Submissions — a technical or position statement of 1-2 pages plus a list of previous publications on a topic suitable for the working group discussions.
Note: Start early. If you do not already have an account, OpenReview advises that it can take up to a day for a new account to become active.
The conference organizers will select 4-8 workshop participants to give short research talks to seed the working group discussions. Acceptance letters from the workshop organizers will be emailed by December 11, 2023.
Participants will be invited to participate in the publication of workshop proceedings (TBD). The organizers of this workshop plan to announce a publication venue ahead of the workshop.
Workshop Organizers
Mark Stefik (Workshop point of contact); SRI International – PARC; mark.stefik@sri.com
Angelo Cangelosi University of Manchester; angelo.cangelosi@manchester.ac.uk
Benjamin Kuipers; University of Michigan; kuipers@umich.edu
Celeste Kidd; UC Berkeley; celestekidd@berkeley.edu
Bob Price; SRI International – PARC; bprice@sri.com
Charles Ortiz; SRI International – PARC; Charles.ortiz@sri.com
Sample Readings
Cangelosi, A., Schlesinger, M. (2015) Developmental Robotics: From Babies to Robots. The MIT Press, Cambridge, MA.
Cangelosi, A., Asada, M. (eds.) (2022) Cognitive Robotics. The MIT Press. Cambridge, MA. https://direct.mit.edu/books/oa-edited-volume/5331/Cognitive-Robotics
Juett, J., Kuipers, B. (2019) Learning and Acting in Peripersonal Space: Moving, Reaching, and Grasping. Frontiers in Neurorobotics 13(4) pp. 1-20.
Kidd, C., Hayden, B.Y. (2015) The Psychology and Neuroscience of Curiosity. Neuron 88(3) pp. 449-460. http://dx.doi.org/10.1016/j.neuron.2015.09.01
LeCun, Y. (2022) A Path Towards Autonomous Machine Intelligence 0.9.2, OpenReview. https://openreview.net/pdf?id=BZ5a1r-kVsf
Levine, S. (2021) Understanding the World through Action. https://arXiv.org/abs/2110.12543
Oudeyer, P.-Y. (2017) Autonomous Development and Learning in Artificial Intelligence and Robotics: Scaling Up Deep Learning to Human-Like Learning. Behavioural and Brain Sciences. 40, pp. 45-46. http://www.pyoudeyer.com/oudeyerBBS17.pdf
Spelke, E. (2022) What Babies Know. Oxford University Press.
Stefik, M., Price, R. (2023) Bootstrapping Developmental AIs. arXiv http://arxiv.org/abs/2308.04586