[Paper Review] Learning in cognitive systems with autonomous dynamics

Claudius Gros, Gregor Kaczor|ArXiv.org|Apr 8, 2008

Neural dynamics and brain function16 references6 citations

TL;DR

This paper proposes a biologically inspired learning framework for autonomous cognitive systems where unsupervised feature extraction occurs during transient transitions between self-sustained internal activity states. By leveraging a dense homogeneous associative network (dHAN) with competitive dynamics, the model autonomously generates learning signals during sensitive periods—transition phases between transient attractors—enabling fast, unsupervised independent component analysis on bar-stripe inputs without external supervision.

ABSTRACT

The activity patterns of highly developed cognitive systems like the human brain are dominated by autonomous dynamical processes, that is by a self-sustained activity which would be present even in the absence of external sensory stimuli. During normal operation the continuous influx of external stimuli could therefore be completely unrelated to the patterns generated internally by the autonomous dynamical process. Learning of spurious correlations between external stimuli and autonomously generated internal activity states needs therefore to be avoided. We study this problem within the paradigm of transient state dynamics for the internal activity, that is for an autonomous activity characterized by a infinite time-series of transiently stable attractor states. We propose that external stimuli will be relevant during the sensitive periods, the transition period between one transient state and the subsequent semi-stable attractor. A diffusive learning signal is generated unsupervised whenever the stimulus influences the internal dynamics qualitatively. For testing we have presented to the model system stimuli corresponding to the bar-stripes problem and found it capable to perform the required independent-component analysis on its own, all the time being continuously and autonomously active.

Motivation & Objective

• To develop a paradigm for autonomous cognitive systems driven by intrinsic, self-sustained neural dynamics rather than external input.
• To address the challenge of avoiding spurious correlations between external stimuli and internally generated activity patterns in highly autonomous systems.
• To identify when and how sensory input can meaningfully influence internal dynamics without disrupting the system’s autonomous operation.
• To implement an unsupervised learning mechanism that activates only during specific transition phases—sensitive periods—between transient attractor states.
• To test the feasibility of this learning mechanism in performing independent component analysis on sparse temporal inputs, such as bar-stripe patterns.

Proposed method

• Models internal dynamics using a dense homogeneous associative network (dHAN) that generates a time series of transient, semi-stable attractor states.
• Defines neural activity via continuous-time rate dynamics governed by growth rates $ r_i(t) $, where $ \dot{x}_i = (1 - x_i)r_i $ for $ r_i > 0 $ and $ \dot{x}_i = x_i r_i $ for $ r_i < 0 $.
• Introduces 'sensitive periods'—transition intervals between transient states—during which external stimuli can qualitatively influence internal dynamics and trigger learning.
• Generates a diffusive learning signal unsupervised whenever sensory input alters the trajectory of the internal dynamics during these sensitive periods.
• Applies an orthogonalization procedure (Eq. 6) to synaptic weights $ v_{ij}^{pq} $, enabling overlapping winning coalitions and preventing saturation, consistent with 'learning by mistakes'.
• Uses a 15-site linear chain dHAN with 14 potential winning coalitions (e.g., (0,1), (1,2), ..., (14,15)) to simulate receptive fields and feature responses to 10 bar-stripe input patterns.

Experimental results

Research questions

• RQ1 How can a cognitive system maintain autonomous internal dynamics while still learning from external stimuli without forming spurious correlations?
• RQ2 What are the temporal conditions—specifically, transition phases—during which external stimuli can meaningfully influence internal dynamics and trigger learning?
• RQ3 Can a self-sustained dynamical system perform unsupervised independent component analysis using only internal transient-state transitions as learning triggers?
• RQ4 How do overlapping winning coalitions and non-saturating synaptic updates affect feature extraction performance in a competitive, autonomous network?
• RQ5 To what extent is the proposed learning mechanism robust under varying input conditions, including sparse and quasi-continuous stimuli?

Key findings

• The model successfully performed unsupervised independent component analysis on a 10-bar-stripe input problem using only transient-state transitions as learning triggers.
• Learning occurred rapidly and unsupervised, with distinct receptive fields emerging for each of the 14 potential winning coalitions, as shown in Fig. 7.
• Receptive fields exhibited both positive and negative weights, reflecting complex, orthogonalized synaptic configurations due to the orthogonalization procedure (Eq. 6).
• The system avoided synaptic saturation by balancing active, orthogonalized, and inactive contributions (Eq. 6), consistent with the 'learning by mistakes' principle.
• The model demonstrated robustness in feature extraction despite overcompleteness (more winning coalitions than independent input patterns), though some bars (e.g., second-last vertical bar) were not fully resolved.
• Preliminary results indicate the framework remains functional under diverse input conditions, supporting future extension to natural scene analysis and continuous input.