BioelectricaAlterspan Labs
Goals and control  ·  Article

Temporal depth of generative models and conscious versus unconscious inference

Temporal depth describes how far into the future a generative model reaches when an organism infers and acts. Friston, Wiese and Hobson tie the split between deep and shallow models to the split between conscious and unconscious inference.4 Kirchhoff and colleagues argue that living homeostatic systems need such depth to act on probable futures.1 Friston, Wiese and Hobson call the distinction vague and quantitative.4

Earliest held
2018, Kirchhoff M, Parr T, Palacios…
Most discussed in
Sentience and the Origins of Consciousness…, 2020
In the library
9 passages in 5 works
Rewritten
2026-10-26
18501875190019251950197520002025Kirchhoff M, Parr T, Palacios E…, 20181Kirchhoff M, Parr T, Palacios E…, 20182Levin M, 20193Friston KJ, Wiese W, Hobson JA, 20204Friston KJ, Wiese W, Hobson JA, 20205Friston KJ, Wiese W, Hobson JA, 20206McMillen P, Levin M, 20247Levin M, Resnik DB, 20268
01

Depth and autonomy

Kirchhoff, Parr, Palacios, Friston and Kiverstein (2018) linked temporal depth to autonomy. They argued that a biological system is homeostatic and has dependencies over several time scales. It must therefore have a generative model with temporal depth. Such a model lets the system sample among options and pick the one with least expected free energy.1 They also argued that action depends on inference about the outcomes of the organism's own actions. Because those outcomes cannot be observed before an action is taken, the organism needs a generative model of future states.2

02

Perceptual control and reflexes

Levin (2019) placed cells within active inference and perceptual control theory. On this view behaviour is the control of perception, and systems tune internal states so as to predict future stimuli.3 Friston, Wiese and Hobson (2020) noted that predictive coding, homoeostasis and thermostats can be described as perceptual control. They called this a reflexive form of active inference, with models built on differential equations.4 They held that any generative model of dynamics has some temporal depth, because differential equations entail some inference over time.

03

Single or multiple trajectories

Friston and colleagues said that the line between deep and shallow is a matter of degree. They suggested a better line. A shallow model entertains one trajectory into the future. A deep model entertains several counterfactual, action-dependent trajectories, and so faces a selection problem, which is choosing an action or planning.5 They related the time a creature takes to return to its attracting manifold to the depth of its policies, and to the idea of planning as inference.6

04

Use in collectives

McMillen and Levin (2024) wrote about decision-making and proto-cognitive functions in multicellular collectives. They stated that they make no claims about first-person experience in such collectives. They also did not claim the same degree of capacity as human cognition.7 Levin and Resnik (2026) likewise held that intelligence comes in degrees. The highest forms involve consciousness and metacognition. Lower forms need only persistence and plasticity in promoting a goal.8

SourcesEach quotation was checked word for word against the passage it opens.
  1. a biological system must possess a generative model with temporal depth, which, in turn, implies that it can sample among different options and select the option that has the greatest (expected) evidence or least (expected) free energy.Kirchhoff M, Parr T, Palacios E, Friston K, Kiverstein J, 2018 · The Markov blankets of life: autonomy, active inference and the free energy… · open at passage 33
  2. The second observation is that for any organism to make inferences of this kind it must have a generative model of future states.Kirchhoff M, Parr T, Palacios E, Friston K, Kiverstein J, 2018 · The Markov blankets of life: autonomy, active inference and the free energy… · open at passage 44
  3. behavior is the control of perception (Powers, 1973). Systems minimize surprise and optimize variational free energy by tuning their internal states in a way that optimally predicts future stimuliLevin M, 2019 · The Computational Boundary of a "Self": Developmental Bioelectricity Drives… · open at passage 42
  4. predictive coding, homoeostasis and thermostats can all be articulated in terms of perceptual control [46,110,111] and a reflexive form of active inference using generative models based upon differential equations.Friston KJ, Wiese W, Hobson JA, 2020 · Sentience and the Origins of Consciousness: From Cartesian Duality to Markovian… · open at passage 66
  5. Perhaps a better distinction would be between generative models that entertain a single trajectory into the future, versus multiple (counterfactual action dependent) trajectories that incur a selection problemFriston KJ, Wiese W, Hobson JA, 2020 · Sentience and the Origins of Consciousness: From Cartesian Duality to Markovian… · open at passage 66
  6. This is functionally equivalent to the temporal depth or extent of policies; namely, courses of action, and internally consistent with the notion of planning as inferenceFriston KJ, Wiese W, Hobson JA, 2020 · Sentience and the Origins of Consciousness: From Cartesian Duality to Markovian… · open at passage 88
  7. we here do not make any claims about first-person experience of unconventional collectives73–76, nor are we saying that the phenomena we describe are of the same degree as familiar human-level cognitive capacities.McMillen P, Levin M, 2024 · Collective intelligence: A unifying concept for integrating biology across… · open at passage 5
  8. The highest forms of intelligence involve consciousness, perception, problem-solving, and metacognition, but lower forms need only involve persistence and plasticity with respect to goal-promotion.Levin M, Resnik DB, 2026 · Mind Everywhere: A Framework for Conceptualizing Goal-Directedness in Biology… · open at passage 14
Linked ideas
Temporal depth is a property of the generative models that the free energy principle and active inference describe.
Choosing among multiple future trajectories is the policy selection problem that marks deep models in the Friston account.
Kirchhoff and colleagues tie temporal depth to homeostatic systems that must handle perturbations over several time scales.
Anticipation in simple organisms bears on how little temporal depth a system can have and still predict.
Levin and Resnik define lower forms of intelligence by persistence and plasticity toward goals, without requiring consciousness.
Basal cognitionrelated to
McMillen and Levin study minimal decision-making in cells and tissues without claims about first-person experience.
B elaborates A: within active inference, temporal depth of generative models lets systems minimise expected free energy over futures, moving from reflexive homoeostasis to allostasis and intentional action.
Where it is discussedPassages matching temporally deep generative model, temporal depth, perceptual control, conscious inference
2020Sentience and the Origins of Consciousness: From Cartesian Duality to Markovian… · Friston KJ, Wiese W, Hobson JA4
2018The Markov blankets of life: autonomy, active inference and the free energy… · Kirchhoff M, Parr T, Palacios E…2
2024Collective intelligence: A unifying concept for integrating biology across… · McMillen P, Levin M1
2026Mind Everywhere: A Framework for Conceptualizing Goal-Directedness in Biology… · Levin M, Resnik DB1
2019The Computational Boundary of a "Self": Developmental Bioelectricity Drives… · Levin M1