One of the things that has bothered me the most since starting to learn about AI / ML systems is that they often treat the passage of time as a second-class citizen, if at all.
Time—in particular the continuous flow of time—plays an absolutely central role in the lives of humans and other animals. I won’t talk about consciousness—I’m not very good at wrestling such slippery topics into submission—except to point out that I could not even imagine it devoid of the flow of time.1
Even sticking to more tractable topics, the list of ways in which time factors into animal lives is long:
- Learning is continuous and ‘on line’
- Attractor states play important roles in the operation of human and other animal brains
- Feedback loops and recurrence are everywhere in the brain
- There are numerous short-timescale processes like temporal coding, spike timing, neural oscillations, phase locking mechanisms, temporal integration of stimuli and motor control, eye saccades, spontaneous activity patterns in the visual system, etc.
- We have the daily circadian rhythm and the sleep-wake cycle
- We have a continuous predictive coding process happening in which we have expectations / anticipations that are either met or not (and when the latter happens it kicks off other downstream processes)
- Our brains make heavy use of experience replay
- Our vision system has the dorsal visual stream, which deals specifically with motion
- We have several memory processes, from short-term to long-term; our memory systems suffer from temporal encoding clashes (learning two things at about the same time inhibits retention / retrieval)
- Our brains never stop developing, from the time we are in the womb until we die.2 There is some evidence that our genetics encode simple rules defining dynamic systems that then generate structure in the brain.3
But it feels terribly reductionist to even list those things; in reality they are all intertwined effects in a dance through time (at least for the short period that we’re alive).
This is mostly missing from current AI systems.
Yes, there are exceptions:
- Robotics are arguably the closest systems we have that—of necessity—deal with time
- Time plays a definite role in spiking neural net (SNN) systems
- Some video classification systems use the concepts of temporal attention or ‘flow’
- Some ML systems by their very nature deal with time-domain signals (e.g., audio)
- Work by Geoffrey Hinton and a few others has touched on sleep-wake cycles
- Some RL systems use a rather crude analog of experience replay
- There have been some notable efforts towards curriculum-based learning (primarily in vision systems and LLMs)
But these feel like they are either tangential invocations of time-based processes or modest attempts to deal with a small piece of the puzzle at a time. They are not time-based through and through, at their core.4
Something in my gut tells me that having time play a central role in our AI systems is going to be critical for getting them to understand causality, object permanence, interactions with 3D systems, and physics; be able to perform complex, hierarchical long-term planning; be robust to adversarial attacks, and so on. I would love to help figure out how we make that happen.
2024-08-21 update: I’m learning a bit about Global Workspace Theory (here is the Wikipedia entry, although I kind of prefer this summary from one of the originators), which dates all the way back to the 1980s. It is a leading theory about how the different components of the brain may work together, and time (not surprisingly) plays a central role.
2025-12-04 update: The recent Continuous Thought Machines paper by Sakana AI is an exciting step in the direction that I’m talking about here. Their blog post provides an accessible overview of the concept: https://pub.sakana.ai/ctm/
Footnotes
-
I quite like Karl Friston’s discussion on consciousness in this article. ↩
-
Check out The Development of Embodied Cognition: Six Lessons from Babies by Smith and Gasser (2005). ↩
-
See Why the Brain is So Noisy for an interesting and accessible read. ↩
-
I have to confess that I’m really failing to convey what I’m thinking about here, almost to the point that I need to regroup and start again. Apologies. ↩