Human category learning is consistent with Bayesian generative but not discriminative classification strategies
-
1
Imperial College London, Department of Computing, United Kingdom
-
2
Imperial College london, Department of Bioengineering, United Kingdom
Since an early age, humans have adopted the ability to group visual entities into categories. Mechanisms that correspond to categorization processes, are typically investigated at both neuronal (Freedman, 2011) and behavioural contexts. Here we step back and design tests with the aim of realising which computational process is used by our brain for forming categories (i.e. classification). In machine learning and pattern recognition two types of classification algorithms are known: 1. Generative and 2. Discriminative approaches. Generative approaches solve the categorization problem by building a probabilistic model of how each category was formed and infer category labels. In contrast, the discriminative approach learns a direct mapping between input and category labels. Recent work (Hsu and Griffiths, 2010) shows human classification is consistent with discriminative and generative classification depending on task conditions. We hypothesize that humans employ generative mechanisms for classification, when not encouraged otherwise. To test this we exploit a counterintuitive prediction for generative classification, namely how the discrimination boundary between two classes shifts if one category's distribution is revealed to be broader during learning. We trained N=17 subjects to distinguish two classes, A and B of visual objects in two different tasks (two Persian-characters and armadillo-horse stick-drawings). The classes in each task were parameterized by two scalars; objects for each class are drawn from Gaussian parameter distributions, with equal variance and different means (class "prototypes"). Next, subjects classify unlabelled examples drawn between the two classes, so we can infer their discrimination boundary. This process is then repeated but includes training data for class A, which lie far away from B. Counter-intuitively, generative classification predicts a shift of the discrimination boundary closer to B. Conversely, discriminative classifiers will show either no shift of the boundary or a shift of the boundary away from class B. Similar to the generative classifier, subjects were strongly influenced by knowledge of the distribution associated with alternative categories as classification boundaries shifted towards B for both tasks across all subjects. Our results indicate that categorization in both tasks is consistent only with generative and not discriminative classification. Our experimental design is directly amenable for neurophysiological studies to investigate the neuronal substrates of generative classification in the brain.
References
A Proposed Common Neural Mechanisms for Categorization and Perceptual Decisions, Freedman D.J. and Assad J.A., Nature Neuroscience, 14: 143-146, 2011.
Hsu, Anne Showen and Griffiths, Thomas E., Effects of Generative and Discriminative Learning on Use of Category Variability (Feb 1, 2010). 32nd Annual Conference of the Cognitive Science Society, 2010.
Keywords:
categorisation,
Human Learning,
machine learning,
object recognition,
visual learning
Conference:
Bernstein Conference 2012, Munich, Germany, 12 Sep - 14 Sep, 2012.
Presentation Type:
Poster
Topic:
Sensory processing and perception
Citation:
Mehraban Pour Behbahani
F and
Faisal
AA
(2012). Human category learning is consistent with Bayesian generative but not discriminative classification strategies.
Front. Comput. Neurosci.
Conference Abstract:
Bernstein Conference 2012.
doi: 10.3389/conf.fncom.2012.55.00095
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
11 May 2012;
Published Online:
12 Sep 2012.
*
Correspondence:
Miss. Feryal Mehraban Pour Behbahani, Imperial College London, Department of Computing, London, SW7 2AZ, United Kingdom, feryal.mehrabanpour10@imperial.ac.uk