Исходя исключительно из входных данных невозможно сделать качественный inference сцены - нужны вводные (bias, aprioi) от вышестоящих зон. Но вышестоящие зоны не могут дать прогнозов без данных от нижестоящих зон.

Эта мысль много где встречается:

1992 [GemanBiasVariance]

To mimic substantial human behavior such as generic object recognition in real scenes - with confounding variations in orientation, lighting, texturing, figure-to-ground separation, and so on - will require complex machinery. Inferring this complexity from examples, is, for all practical matters, not feasible: too many examples would be needed. Important properties must be built-in or “hard-wired,” perhaps to be tuned later by experience, but not learned in any statistically meaningful way

1994 [MumfordNeurArc]

The first idea is that to sucessfully reconstruct the world variable, one must learn to synthesize the coded signals that one observes, so that tentative reconstructions of world variables can be checked by comparing the actual observer signal with synthesized signals. This means that architecture is not purely feedforward, bottom-up, but fundamentally recursive combining feedforward actions with feedback, top-down processing with bottom-up.

1997 [HintonGenModelsForSDR]

Many neural network models of visual perception assume that the sensory input arrives at the bottom, visible layer of the network and is then converted by feedforward connections into successively more abstract representations in successive hidden layers. Such models are biologically unrealistic because they do not allow for top-down e¡ects when perceiving noisy or ambiguous data (Mumford 1994; Gregory 1970) and they do not explain the prevalence of top-down connections in the cortex.)

2006 [Olshausen85Perc]

Arriving at the correct interpretation of an image, then, constitutes something of a chicken-egg problem between lower and higher levels of image analysis. The low-level shape features that are useful for identifying an objectedges, contours, surface curvature, and the like-are typically ambiguous in natural scenes, so they cannot be computed directly based on a local analysis of the image. Rather, they must be inferred based on global context and higherlevel knowledge. However, the global context itself will not be clear until there is some degree of certainty about the presence of low-level shape features. A number of theorists have argued from this that recognition depends on information circulating through corticocortical feedback loops in order to disambiguate representations at both lower and higher levels in parallel

But years of research in computer vision have now conclusively demonstrated that it is impossible to compute the presence even of simple edges in an image in a purely bottom-up fashion (i.e., a filter such as a simple or complex cell model). … Our perception of crisp contours, corners, and junctions in images is largely a post hoc phenomenon that is the result of massive inferential computations performed by the cortex. In this sense our initial introspections about scene structure may actually be a poor guide as to the actual problems faced by the cortex.

a number of studies have now shown that human observers are quite poor judges of what constitutes a boundary or junction when these features are shown in isolation (Elder, Beniaminov, and Pintilie 1999; McDermott 2003).

2010 [ErhanUnsupervisedHelp]

If this disentangling hypothesis is correct, it would help to explain how unsupervised pre-training can address the chicken-and-egg issue explained in Section 2: the lower layers of a supervised deep architecture need the upper layers to define what they should extract, and vice-versa. Instead, the lower layers can extract robust and disentangled representations of the factors of variation and the upper layers select and combine the appropriate factors (sometimes not all at the top hidden layer).

2021 [HintonGlomFinal]
Из модели следует, что есть иерархия модулей, и есть bottom-up и top-down тракты. И понимание есть многотактный процесс. При стабилизации описаний на каждом из уровней и рождается то самое понимание того, что мы видим/слышим и т.д.

2022 [LeCunPathToAutonomMachIntel]

I submit that the ability to represent sequences of world states at several levels of abstraction is essential to intelligent behavior.

Prediction takes place at all levels.

Многотактность приводит к мысли, что архитектура должны быть: 1) либо явно выражающая многотактность. Т.е. сетка многократно потребляем информацию и потихоньку стабилизуерсят (похоже на RNN) 2) либо unwrapped, когда мы разворачиваем много модулей в длинную цепочку и стабилизация происходит по мере прохождения информации по цепочке. Трансформеры? Например, в GPI-3 96 слоев.