Learning and disrupting invariance in visual recognition with a temporal association rule

Abstract

† These authors contributed equally to this work.Learning by temporal association rules such as Foldiak's trace rule is an attractive hypothesis that explains the development of invariance in visual recognition. Consistent with these rules, several recent experiments have shown that invariance can be broken at both the psychophysical and single cell levels. We show (1) that temporal association learning provides appropriate invariance in models of object recognition inspired by the visual cortex, (2) that we can replicate the "invariance disruption" experiments using these models with a temporal association learning rule to develop and maintain invariance, and (3) that despite dramatic single cell effects, a population of cells is very robust to these disruptions. We argue that these models account for the stability of perceptual invariance despite the underlying plasticity of the system, the variability of the visual world and expected noise in the biological mechanisms.

Toward a unified model of face and object recognition in the human visual system

“…Network simulations have demonstrated how a minor modification to standard Hebbian association (called the trace rule) can produce view change tolerant representations in self-organizing systems (Földiák, 1991; Becker, 1993; Wallis et al, 1993; Wallis, 1998)—see Rolls (2012) and Bart and Hegdé (2012) for recent reviews. Subsequent electrophysiological studies have leant further support to this theory (Cox et al, 2005; Cox and DiCarlo, 2008; Li and DiCarlo, 2008, 2010) which has prompted developers of other hierarchical models of object recognition to experiment successfully with trace-rule learning (Isik et al, 2012). …”

Temporal contiguity training does not affect size-tolerant representations in object-selective cortex

“…This could be the case if the effects of the exposure are confined to the small subset of cells with a high preference for one of the objects involved in the training, as was the case in the model of Isik et al (2012). It is not sure that this can explain the discrepancy with the monkey results, because much of the evidence came from multi-unit recordings that also already pool across multiple neurons.…”

“…While the electrophysiological methods used by Li & DiCarlo allowed them to measure responses of single neurons or small groups of neurons and adjust stimulus selection to their selectivity, the smallest measurement unit of fMRI, a voxel, consists of hundreds of thousands of neurons. A computational modeling study by Isik, Leibo, and Poggio (2012) suggested that, depending on the size of the measured cell population and the amount of exposure, large changes at the level of single cells might be negligible at the population level.…”

Exploiting temporal continuity of views to learn visual object invariance

“…The authors augmented HMAX to incorporate a simplified, but effective, implementation of learning by temporal association, called the “modified trace rule.” This augmented model was able to reproduce a diagnostic feature of invariance learning: when trained with smooth temporal variations of a given object, such as a face [see Figure 1, top left , of Isik et al (2012)], the neurons in the topmost layer of the network, individually and collectively, did learn an invariant representation of that object.…”

“…In each such sequence, images at all positions showed the same object (e.g., a face), except for one position (called the “swap position”) that showed a different object [e.g., a car; see Figure 1, top right , of Isik et al (2012)]. As expected, invariance tuning of each cell trained with such a sequence was disrupted, with the cell responding to the main object at most locations, but responding more strongly to the swap object at the swap location.…”