The Natural Input Memory Model

The Natural Input Memory Model Joyca P.W. Lacroix (j.lacroix@cs.unimaas.nl) Department of Computer Science, IKAT, Universiteit Maastricht, St. Jacobsstraat 6, 6211 LB Maastricht, The Netherlands Jaap M.J. Murre (jaap@murre.com) Department of Psychology, Universiteit van Amsterdam, Roeterstraat 15, 1018 WB Amsterdam, The Netherlands Eric O. Postma (postma@cs.unimaas.nl) H. Jaap van den Herik (herik@cs.unimaas.nl) Department of Computer Science, IKAT, Universiteit Maastricht, St. Jacobsstraat 6, 6211 LB Maastricht, The Netherlands recognition-memory studies. Finally our main conclusion will be given. Abstract A new recognition memory model is proposed which differs from the existing memory models in that it operates on natural input. Therefore it is called the natural input memory (N IM ) model. A biologically-informed perceptual pre-processing method takes local samples from a natural image and translates these into a feature-vector representation. The feature-vector representations reside in a similarity space in which perceptual similarity corresponds to proximity. By using the similarity structure of natural input, the model by-passes assumptions about distributional statistics of real-world input. Our sim- ulations on the list-strength effect, the list-length effect, and the false memory effect support the validity of the proposed model. In particular, we conducted a face recognition simula- tion with the N IM model and found that it is able to replicate well-established recognition memory effects that relate to the similarity of the input. The N IM Model The N IM model encompasses the following two stages. 1. A perceptual pre-processing stage that translates a natural image into a number of feature vectors. 2. A memory stage comprising two processes: (a) a storage process that simply stores feature vectors; (b) a recognition process that compares feature vectors of the image to be recognized with previously stored fea- ture vectors. Memory Representation Many computational memory models represent an item by a vector of abstract features (e.g., the S AM model, Raaijmakers & Shiffrin, 1981; the R EM model, Shiffrin & Steyvers, 1997, the model of differentiation, McClelland & Chappell, 1998). The feature values are usually drawn from a mathematical distribution (e.g., a geometric distribution). Since the com- putational models artificially generate vector representations, they do not address the contribution of the similarity struc- ture intrinsic to natural data. However, we believe that the similarity structure contains important information. There- fore, we propose a memory model that operates on natural data and represents the similarity structure of these data. The similarity structure of natural data can be represented in any type of space that fulfills the compactness criterion (Arkadev & Braverman, 1966). This criterion is fulfilled when similar objects in the real world are close in their rep- resentations. Several researchers developed so called ‘simi- larity spaces’, in which representations of similar items are in close proximity of each other (e.g., Nosofsky, 1986; Steyvers, Shiffrin, & Nelson, in press). An analysis of human similarity judgments or of free association data often forms the basis of a similarity space. However, we propose to derive the similar- ity space from the natural data by employing a biologically- informed transformation. In the next section, a new recognition memory model that operates on natural images is introduced and described. We call this model the natural input memory (N IM ) model. We will conduct a face recognition simulation with the N IM model and will evaluate its ability to replicate findings from Figure 1: The natural input memory (N IM ) model. Figure 1 presents a schematic diagram of the N IM model. The face image is an example of a natural image. The two boxes correspond to the perceptual pre-processing stage and the memory stage. The Perceptual Pre-Processing Stage In this section, we first provide some background on the sources of biological inspiration and on the computational considerations. Then, we discuss some relevant implemen- tation details. Biological Inspiration and Computational Considerations The human visual system is our main source of biological inspiration. The eye sequentially fixates on those parts of a visual scene that are most informative for recognition (e.g., Yarbus, 1967). Early visual processing in the brain leads to the activation of millions of optic nerve cells (Palmer, 1999). The nerve-cell activations may be conceived as a high di- mensional vector. The high dimensionality enables the rep- resentation of a large amount of information without suffer- ing from interference (Rao & Ballard, 1995), but it also ham- pers the memory performance, as the number of examples