As an AI enthusiast, I have always been fascinated by the developments taking place in the field, and recently I stumbled upon an interesting research paper Yann LeCun about how to truly create an artificial intelligence.
For those of you who don't know, he's the early pioneer of deep learning, who worked alongside with Geoffrey Hinton.
Specially, Yann talks about combining predictive world models, intrinsic motivation, and joint-embedding architectures, all trained through self-supervised learning.
This position paper, which has been cited by over 50 researchers, explores how intrinsic motivation may be introduced into AI systems to revolutionize how they learn.
The Key to Autonomous Learning in Artificial Intelligence
Intrinsic motivation is unique compared to other learning methods since it does not involve any hardwired programs, supervision, or external rewards.
Unlike reinforcement learning, it revolves around internal drives and natural curiosity. The paper points out the limitations of current world models and hardwiring behaviors, which are unable to teach machines to drive with just 20 hours of practice, like it does for teenagers.
Instead, the author argues that the secret lies in our ability to learn the world model, our internal understanding of how the world functions.
This idea reminds me of Jordan Peterson's description of phenomenology, which describes how humans perceive the world around them.
Yann's Proposal
The paper highlights three challenges that AI needs to overcome:
- Learning the world by observation and with minimal interaction
- Combining logic-based symbolic reasoning with gradient-based approaches
- Representing percepts and action plans hierarchically and at multiple abstraction levels and time scales
To address these challenges, Yann LeCun proposes a hybrid solution consisting of:
- A modular system where each component can be trained separately
- Joint Embedding predictive Architecture (JEPA) and Higher-JEPA, which learn hierarchical representations of the world
- A non-contrastive self-supervised learning paradigm that does not differentiate between similar and dissimilar data
- Employing Higher-JEPA for hierarchical planning under uncertainties
This approach suggests that animals can acquire common sense through observation and minimal interaction, enabling them to predict the outcome of their actions, reason, plan, explore, and create novel solutions to problems. For instance, humans can intuitively understand the physics of driving and adjust their speed accordingly, even if they have not studied it explicitly.
Common sense can effectively fill in missing information about the world around us, both in time and space. It allows us to make sense of ambiguous stimuli, a valuable trait that has helped humans survive and evolve.
The Paradox of Common Sense
As an enthusiast in debiasing and understanding the human condition, I believe that humans' common sense can sometimes put society at risk – whether through conflicts, racism, or misinterpreting the power of exponential growth and compound interest.
Nevertheless, common sense plays an essential role in our day-to-day activities, especially in physical movement and decision-making. In order to develop truly intelligent machines that can imitate human actions and thoughts, it's essential to help them grasp our common sense knowledge.
The hybrid approach outlined in this paper, combining self-supervised learning with joint embedding architectures and intrinsic motivation, holds potential to redefine the AI landscape.
The Future Is Moving
As we move into 2023, the possibilities for integrating intrinsic motivation and common-sense understanding into AI systems continue to excite not only researchers but also all AI practionars like myself. Instead of being swayed by OpenAI and Google's PR, we can predict where the true trend is going in the future.
In fact, 1 year afterwards, Yann LeCun and his Facebook team actually open sourced a new paradigm shift model called Image-JEPA. He's truly living his word.
Building machines that can learn, reason, and adapt like humans is a fascinating endeavor, and I eagerly anticipate the new developments and breakthroughs that await us in this rapidly advancing field.