1 What each paper does in a nutshell
Albarracin et al., “Mapping Husserlian phenomenology onto active inference.”
Goal. Turn Husserl’s philosophical analysis of consciousness—especially his account of temporal “thickness” (primal‑impression → retention → protention) and the cycles of anticipation/fulfilment—into the language of generative models used in the free‑energy / active‑inference framework.
Key moves.
- observations ↔ hyletic data,
- hidden states ↔ lived perceptual contents,
- A/B matrices ↔ sedimented background knowledge,
- forward/back‑messages in a POMDP ↔ retentions and protentions that tether the present to a just‑past and a just‑ahead flow of experience,
- variational‑free‑energy minimisation ↔ the phenomenological drive toward fulfilment and reduced surprise.
The paper’s broader pitch is a computational phenomenology: by formalising first‑person structures we can build testable models of consciousness and behaviour . Darlow et al., “Continuous Thought Machines.”
Goal. Re‑introduce real neural timing into deep learning through a new architecture (CTM) whose primitive representation is the evolving synchrony of neuron activations rather than static layer outputs.
Two core innovations. - Neuron‑level models: each neuron owns a tiny recurrent MLP that converts a sliding window of its pre‑activations into the next activation, generating rich internal dynamics .
- Neural‑synchronisation latent: at every “internal tick” the pairwise synchrony matrix of activations becomes the latent used to query attention and to project outputs; learnable exponential decays let the model privilege recent or distant activity as needed .
Because computation unfolds over flexible ticks, the model naturally “ponders” longer on hard inputs and stops early on easy ones, achieving adaptive compute and surprisingly good probability calibration across many tasks (ImageNet, mazes, sorting, RL) .
2 Where the two stories meet
| Phenomenology + Active Inference (Albarracin) | Continuous Thought Machine (Darlow) | Common thread |
|---|---|---|
| Retention / protention encode an implicit history and horizon of expectations that steer present perception | A running window of activations and their synchrony steers what the model attends to next; exponential decays tune horizon length | Temporal context as the substrate of cognition |
| Variational free‑energy minimisation makes an agent act to fulfil predictions and reduce surprise | Tick‑by‑tick loss uses certainty‑weighted cross‑entropy; the model “thinks” until prediction certainty peaks, implicitly minimising expected error/surprise | Self‑evidencing loop (predict → sample → update) |
| Generative model parameters (A, B, C, D) store sedimented knowledge that gets updated through experience | Neuron‑level weights and learnable decay rates are updated by SGD, gradually shaping the dynamical landscape the CTM inhabits | Learning as sedimentation of structure |
| Computational phenomenology aims to test first‑person theories with runnable models | CTM supplies a concrete, biologically‑inspired machine in which phenomenological ideas (continuous inner time, fulfilment dynamics) can be simulated at scale | A bridge from theory to implementation |
3 A synthetic perspective
- Temporal dynamics are not epiphenomenal; they are the representation.
Albarracin et al. argue that consciousness is constituted by a flowing structure of anticipations; Darlow et al. show that letting neural activity flow (rather than freezing it into static activations) yields practical gains and emergent reasoning. The CTM can thus be read as an engineering proof‑of‑concept for computational phenomenology. - Retention / protention ↔ state‑history indexing in CTM.
The CTM’s sliding windows of pre‑ and post‑activations functionally resemble Husserlian retentions (still‑living past) and protentions (immediate future expectations). Learnable decay rates modulate how quickly the “living present” fades—mirroring Husserl’s fading of retention . - Fulfilment and frustration manifest as certainty curves.
In CTM training the tick at which certainty peaks versus ground‑truth labels can be seen as experiential fulfilment; mis‑predictions that later self‑correct mimic phenomenological “frustration” events driving model update (cf. CTM visualisations in Fig. 5 of the paper) . - Free‑energy principle gives CTM a normative envelope.
Re‑casting CTM loss as variational‑free‑energy (e.g., by adding explicit epistemic‑value terms) would align it mathematically with active inference, potentially giving the architecture principled exploration and policy selection—an avenue invited by Albarracin’s mapping. - Experimental test‑bed for phenomenology.
By injecting phenomenological manipulations (e.g., altering decay to lengthen the “specious present”) and reading out synchrony patterns, one could generate quantitative predictions for neuro‑phenomenological experiments—delivering on the empirical ambitions of computational phenomenology.
4 Opportunities going forward
- Integrate objectives. Add an explicit expected‑free‑energy term to CTM training so that action‑selection (attention queries, output timing) becomes Bayes‑optimal in the active‑inference sense.
- Hierarchical time consciousness. Husserl emphasises nested temporalities (melody vs. symphony). A multi‑layer CTM where upper layers integrate lower‑layer synchronies over longer tick scales could capture this.
- Subjective‑state probing. The synchrony matrix is interpretable. Mapping clusters of synchrony to phenomenological categories (e.g., fulfilment vs. surprise) could supply a new kind of “virtual first‑person report.”
- Adaptive compute as phenomenological duration. Varying the tick budget lets one model how attention stretches subjective time under cognitive load, offering a quantitative handle on classic phenomenological observations.