Analysis by Synthesis: A (Re-)Emerging Program of Research for Language and Vision

By Thomas G. Bever, David Poeppel. 2010

127 citations recorded by the source corpus.

This contribution reviews (some of) the history of analysis by synthesis, an approach to perception and comprehension articulated in the 1950s. Whereas much research has focused on bottom-up, feed-forward, inductive mechanisms, analysis by synthesis as a heuristic model emphasizes a balance of bottom-up and knowledge-driven, top-down, predictive steps in speech perception and language comprehension. This idea aligns well with contemporary Bayesian approaches to perception (in language and other domains), which are illustrated with examples from different aspects of perception and comprehension. Results from psycholinguistics, the cognitive neuroscience of language, and visual object recognition suggest that analysis by synthesis can provide a productive way of structuring biolinguistic research. Current evidence suggests that such a model is theoretically well motivated, biologically sensible, and becomes computationally tractable borrowing from Bayesian formalizations.

Subfields: Experimental and Cognitive Psychology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Experimental and Cognitive Psychology

Citations in the corpus

Towards causal generative scene models via competition of experts (2020, methodology): “Vision as inverse graphics & probabilistic programs Another way to programmatically introduce information about scene composition is through analysis-by-synthesis, see Bever & Poeppel (2010) for an overview.”
Single Image 3D Interpreter Network (2016, methodology): “obust to imperfect detection results and can be jointly learned with keypoint estimators. Our work also connects to the traditional eld of vision as inverse graphics [14,21] and analysis by synthesis [58,20,5,52], as we use neural nets to decode latent 3D structure from images, and use a projection ”
Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations (2019, methodology): “Recently, following the spirit of “vision as inverse graphics” [59, 60], deep neural networks have been applied to the task of inverting graphics engines [61–65].”
Discovering linguistic structures in speech: models and applications (2014, background): “Can this type of information be included as part of a generative model? While we choose to build our model on a feature representation that only implicitly reflects this type of information [9], we see the potential in taking the analysisby-synthesis approach to building models that more closely res”
Visual Object Networks: Image Generation with Disentangled 3D Representations (2018, background): “Motivated by the philosophy of “vision as inverse graphics” [Yuille and Kersten, 2006, Bever and Poeppel, 2010], researchers have made much progress in recent years on learning to invert graphics engines, many with deep neural networks [Kulkarni et al., 2015b, Yang et al., 2015, Kulkarni et al.,…”