Sound Research WIKINDX

"What happens when, after a brief chat with a colleague, I re-enter my office and visually perceive the hot, steaming, red cup of coffee that I left waiting on my desk? One possibility is that my brain receives a swathe of visual signals (imagine, for simplicity, an array of activated pixels) that specify a number of elementary features such as lines, edges, and color patches. Those elementary features are then progressively accumulated and (where appropriate) bound together, yielding shapes and specifying relations. At some point, these complex shapes and relations activate bodies of stored knowledge, turning the flow of sensation into world-revealing perception: the seeing of coffee, steam, and cup, with the steaming bound to the coffee, the color red to the cup, and so on.

[...]

As I re-enter my office my brain already commands a complex set of coffee-involving expectations. Glancing at my desk sets off a chain of visual processing in which current bottom-up signals are met by a stream of downwards predictions concerning the anticipated states of various neuronal groups along the appropriate visual pathway. In essence, a multi-layer downwards cascade is attempting to "guess" the present states of all the key neuronal populations responding to the present state of the visual world. There ensues a rapid exchange (a dance between multiple top-down and bottom-up signals) in which incorrect guesses yield error signals which propagate forward, and are used to extract better guesses. When top-down guessing adequately accounts for the incoming signal, the visual scene is perceived. As this process unfolds, top-down processing is trying to generate the incoming sensory signal for itself. When and only when this succeeds, and a match is established, do we get to experience (veridically or otherwise) a meaningful visual scene."

"Representations constructed at earlier stages of processing feed immediately higher levels in a feedforward manner...this process proceeds incrementally until access to a ''lexical conceptual'' representation has been achieved. In speech recognition...this involves a conversion from acoustic features onto phonetic representations, phonetic representations onto phonological representations, and finally access of the lexical item based on its phonological structure.
[...]
(1) the extraction of (necessarily brief and coarse) cues in the input signal to elicit hypotheses, that while coarse, are sufficient to generate plausible guesses about classes of sounds (for example, plosives, fricatives, nasals, and approximants), and that permit subsequent refinement; (2) the actual synthesis of potential sequences consistent with the cues; and (3) a comparison operation between synthesized targets and the input signal delivered from the auditory analysis of the speech."

David Poeppel and Philip J. Monahan “Feedforward and feedback in speech perception: Revisiting analysis by synthesis”, Language and Cognitive Processes 26:7, (2011): 935-95.