slow ritornellos

VI. Predictive Coding 1. A large body of experimental evidence has shown that, following training with temporal sequences of sensory stimuli, BOLD response is greater when a stimulus violates the temporal pattern established by prior training, than when a stimulus is consistent with that learned pattern. 2. A cortical region learns temporal models of its input and uses these models to continuously generate predictions based on its current input in the context of its prior inputs. Prediction errors are then used to update the learned models. 3. The predictive coding hypothesis accounts for the decreased activity due to predicted stimuli, as well as for simple repetition suppression and priming effects. With a hierarchical model of predictive coding, whereby each level of cortex generates top-down predictions which are compared with novel input at a lower level. 4. Only the difference between the two, the prediction error, is transmitted back up to the higher level.

V. Temporal Slowness 1. One reason visual invariance problems are difficult is because they are not easily solved by either of these methods. Two input patterns representing the same object with some variation in position (or scale, rotation, etc.) may be separated by a very large Euclidean distance, while two patterns representing different objects may be much closer. 2. Error driven learning has been proven more capable of making the correct discriminations, but is not very good at generalizing these discriminations beyond its training set. 3. Temporal slowness is based on the idea that sensory data changes on a much more rapid timescale than do the relevant properties of the world. While the identities of nearby objects and their configuration in the environment tend to change only gradually, the visual input reflecting that environment can change markedly from one moment to the next, due for example to the movement of objects or a change in view angle. 4. Temporal slowness takes advantage of the fact that the many possible visual impressions that can be projected by an individual object make up a single, contiguous manifold in the high dimensional space of visual input. Because of this, different visual impressions of the same object will tend to transform continuously from one to another, tracing a path on that manifold. 5. This information can be used to learn invariant representations in an unsupervised manner, or rather by using time as a supervisor; those input representations that tend to occur adjacently in time are transformed into a common, invariant output representation. 6. Slow feature analysis restates the temporal slowness principle as a nonlinear optimization problem, in which, given an input signal, it will find the functions that will produce those output signals that vary most slowly over time, while still carrying significant information. 7. Unlike the earlier models, however, slow feature analysis remains a mathematical algorithm; there is not yet a biological theory for how it may be implemented by the cortex. However, it presents a credible solution to the question of how the disjunctive cell properties in a MHWA model such as HMAX may be learned.

IV. Multi-Stage Hubel Wiesel Architectures 1. According to MHWA, simple cells respond mostly to lines of a particular orientation, spatial frequency, and position, while complex cells introduce some degree of invariance to position. These MHWA models are composed of alternating layers of conjunctive and disjunctive units. 2. The conjunctive units perform template matching, responding to a particular combination of inputs from the previous layer, such as a contrasting edge at a particular orientation and position (often described using a Gabor filter function). 3. The disjunctive units are wired to respond when any of several related conjunctive units from a local area in the previous layer are active, for example when any unit is active that represents a given line orientation and spatial frequency, but at any position within a local range. Each disjunctive unit pools over a particular set of conjunctive input units. 4. The disjunctive unit layer then feeds into a second conjunctive unit layer, which learns to respond to specific combinations of those partially invariant representations, and so on. The result is that, over the course of several levels, representations are learned that are selective to individual whole objects but are also invariant to changes in position and scale. Shortcomings :- a. However it does not address how all of those response properties may develop in the first place. Specifically, HMAX's disjunctive units are hard-wired to the particular set of conjunctive units that they pool over, in order to respond invariantly to the activity of any of those conjunctive units and so introduce spatial and scale invariance; these connections are set parametrically, rather than through a learning process. The HMAX model remains agnostic to the nature of this learning process. b. It is unclear what role the disjunctive layers may play in non-visual cortical functions. In addition, there are other cortical functions such as the representation of temporal sequences, crucial for auditory and motor processing, that are not modeled at all within the HMAX framework. c. Even within the domain of visual object recognition, HMAX relies on a separate, supervised mechanism to learn other types of invariance, such as pose, rotation and lighting invariance.

Propositional Modularity and Autoepistemic Closure from Being No One by Metzinger

III. Deep Learning 1. Deep learning architectures exploit the assumption that the generating causes underlying observations about the world are organized compositionally or categorically into multiple hierarchical levels, and that higher level representations can build upon combinations and transformations of lower level representations. 2. These architectures generally use a combination of supervised and unsupervised learning to extract statistical regularities at each level, passing the transformed input representation up to the next level where further regularities may be extracted. 3. Each layer of the network acts as a filter, capturing a subset of regularities from the input signal, and so reducing the dimensionality of the data. For example, when applied to visual object recognition problems, a deep learning architecture may learn representations at the lowest level that correspond to lines and edges; at the next higher level it may learn representations corresponding to corners and intersections of lines; at the next level it may represent parts of objects, and so on.

Notes on "Towards Universal Cortical Algorithm" by Michael Ferrier II. Bayesian Inference 1. A number of theories have proposed that top-down feedback between cortical layers may provide contextual priors to influence inference at lower levels and attention, seen as biased competition within a cortical region, may be just one aspect of a process of biased inference that is mediated by top-down feedback. 2. Building on pattern theory proposed that a cortical hierarchy performs Bayesian belief propagation. In this view, a cortical area treats its bottom-up input (from sensory thalamus or from hierarchically lower cortical areas) as evidence, which it combines with top-down inputs from higher cortical areas that are treated as Bayesian contextual priors, in order to determine its most probable hypotheses and to activate their representations. 3. Over time, the region moves toward an equilibrium in which the optimum set of representations is activated, in order to maximize the probability of the active representations given both the bottom up and top down data. 4. In cases where direct input differs from perception, as with binocular rivalry or illusory contours, or where input is ambiguous and may be perceived in more than one way, short latency (100ms-200ms) responses in V1 correspond well with bottom up thalamic input, while longer latency responses instead correspond partially with what is perceived.

Constraints on a Universal Cortical Algorithm I. Sparse Distributed Representation 1. In this model, the activity of a small fraction of the cell population within an area acts as a representation. a. Because such distributed representation is combinatorial, storage capacity increases exponentially as the number of units grows. b. Distributed representation facilitates generalization among representations based on the degree of overlap between their respective patterns of activity. c. The activity of the individual units that make up a distributed representation may have meaning on their own, which allows for semantically relevant overlap between various representations. d. A representation's pattern of activity may be made up of subpatterns that represent classes in a type hierarchy, or sub-parts in a compositional relationship. With distributed representation, new representations can differentiate themselves progressively by gradual weight modifications, whereas localist representations are formed abruptly and discretely. e. By virtue of their redundancy, distributed representations are more fault tolerant than localist representations. 2. While the representational capacity of a sparse representation is very high, the representational capacity of a dense representation is unnecessarily high, resulting in a great degree of redundancy between representations. The high information content of dense distributed representations make associative mappings between representations much more complex. a. The mappings would generally not be linearly separable, and so would require multilayer networks and learning algorithms that strain biological plausibility. Linear separability is much more readily achieved with sparse distributed representations, which allows many more associative pairs to be learned by a single layer of connections modified by a biologically plausible local Hebbian learning function. b. Both types of distributed representation support generalization between overlapping patterns, but because the activity of a single unit in a sparse representation has lower probability and can therefore be more selective, there is greater opportunity (depending on the learning algorithm) for a sparse representation to mix several semantically meaningful subpatterns in a combinatorial manner.

Notes on "Towards Universal Cortical Algorithm" by Michael Ferrier Basic concepts 1. All areas of cortex have a laminar structure that is typically divided into six layers, each with stereotypical patterns of connectivity and cell type concentrations. 2. Crossing perpendicular to the layer structure are vertically aligned patterns of connectivity by which groups of ~100 cells are densely interconnected. These columns (also known as minicolumns) are often theorized to make up cortical “micro-circuits”, individual functional units. 3. Cells within a single minicolumn typically have similar receptive fields and response properties, leading to the hypothesis that each minicolumn acts as an individual pattern recognizer or feature detector. 4. Connectivity and receptive field characteristics also indicate a larger columnar organization, often called the hypercolumn, which is made up of ~100 minicolumns. 5. Inhibitory connections between the minicolumns within a hypercolumn supress activity in all but those minicolumns that best match their current pattern of input, which facilitates the self-organization of the minicolumns' receptive fields and produces a sparse re-coding of the input pattern. 6. The similarities in architecture between cortical areas and between species is so great as to suggest a common underlying structural and functional template that is repeated many times, with some modification, throughout the cortex. 7. Input is broken up into statistically coherent chunks at the “fault lines” of lower predictive probability, and these learned chunks are then made available as components for higher level associative learning.

We must destroy the rumination machine. The real version of the matrix is the idea that anything ever was or could be certain as long as you follow some pre-baked steps and minmax specific numbers under a specific modern institution that's less than a millennium old. It goes out with the trash. The domesticated humans won't survive the next decade without following their leaders to hell. Furthermore, whether or not a human is domesticated has nothing to do with being a person, which is a whole different skill that mostly has to do with adjusting one's sensitivities so that others become truly visible. For one, I have seen other animals do better than some humans! Re-grounding in the Real is really essential. There is barely a line between the fantasy world and the dream world, but the real world very much supersedes both in a totalizing way when it engages. To paraphrase Dr. Saha here: The hand-to-brain connection must be restored. In my words: The nightmares of the Brahmin are totalities to him because he has trained himself to abolish the Real in his mind, so there is no recourse to the cosmic terrors he encounters in his mind.