Can machine learning deliver AI? Theoretical results, inspiration from the brain and cognition, as well as machine learning experiments suggest that in order to learn the kind of complicated functions that can represent high-level abstractions (e.g. in vision, language, and other AI-level tasks), one would need deep architectures. Deep architectures are composed of multiple levels of non-linear operations, such as in neural nets with many hidden layers, graphical models with many levels of latent variables, or in complicated propositional formulae re-using many sub-formulae. Each level of the architecture represents features at a different level of abstraction, defined as a composition of lower-level features. Searching the parameter space of deep architectures is a difficult task, but new algorithms have been discovered and a new sub-area has emerged in the machine learning community since 2006, following these discoveries. Learning algorithms such as those for Deep Belief Networks and other related unsupervised learning algorithms have recently been proposed to train deep architectures, yielding exciting results and beating the state-of-the-art in certain areas. Learning Deep Architectures for AI discusses the motivations for and principles of learning algorithms for deep architectures. By analyzing and comparing recent results with different learning algorithms for deep architectures, explanations for their success are proposed and discussed, highlighting challenges and suggesting avenues for future explorations in this area.

/pdf/learning-deep-architectures-for-ai-2358znl654.pdf

Learning Deep Architectures for AI

I have developed "tennis elbow" from lugging this book around the past four weeks, but it is worth the pain, the effort, and the aspirin. It is also worth the (relatively speaking) bargain price. Including appendixes, this book contains 894 pages of text. The entire panorama of the neural sciences is surveyed and examined, and it is comprehensive in its scope, from genomes to social behaviors. The editors explicitly state that the book is designed as "an introductory text for students of biology, behavior, and medicine," but it is hard to imagine any audience, interested in any fragment of neuroscience at any level of sophistication, that would not enjoy this book. The editors have done a masterful job of weaving together the biologic, the behavioral, and the clinical sciences into a single tapestry in which everyone from the molecular biologist to the practicing psychiatrist can find and appreciate his or

Principles of Neural Science

http://science.sciencemag.org/content/sci/201/4359/899.2.full.pdf

Phenomenology of perception.

A free-energy principle has been proposed recently that accounts for action, perception and learning. This Review looks at some key brain theories in the biological (for example, neural Darwinism) and physical (for example, information theory and optimal control theory) sciences from the free-energy perspective. Crucially, one key theme runs through each of these theories — optimization. Furthermore, if we look closely at what is optimized, the same quantity keeps emerging, namely value (expected reward, expected utility) or its complement, surprise (prediction error, expected cost). This is the quantity that is optimized under the free-energy principle, which suggests that several global brain theories might be unified within a free-energy framework.

/pdf/the-free-energy-principle-a-unified-brain-theory-3vz1twg7nv.pdf

The free-energy principle: a unified brain theory?

Prelude. Cycle 1. Introduction. Cycle 2. Structure defines function. Cycle 3. Diversity of cortical functions is provided by inhibition. Cycle 4. Windows on the brain. Cycle 5. A system of rhythms: from simple to complex dynamics. Cycle 6. Synchronization by oscillation. Cycle 7. The brain's default state: self-organized oscillations in rest and sleep. Cycle 8. Perturbation of the default patterns by experience. Cycle 9. The gamma buzz: gluing by oscillations in the waking brain. Cycle 10. Perceptions and actions are brain state-dependent. Cycle 11. Oscillations in the "other cortex:" navigation in real and memory space. Cycle 12. Coupling of systems by oscillations. Cycle 13. The tough problem. References.

Rhythms of the brain

The lateral geniculate nucleus is the best understood thalamic relay and serves as a model for all thalamic relays. Only 5-10% of the input to geniculate relay cells derives from the retina, which is the driving input. The rest is modulatory and derives from local inhibitory inputs, descending inputs from layer 6 of the visual cortex, and ascending inputs from the brainstem. These modulatory inputs control many features of retinogeniculate transmission. One such feature is the response mode, burst or tonic, of relay cells, which relates to the attentional demands at the moment. This response mode depends on membrane potential, which is controlled effectively by the modulator inputs. The lateral geniculate nucleus is a first-order relay, because it relays subcortical (i.e. retinal) information to the cortex for the first time. By contrast, the other main thalamic relay of visual information, the pulvinar region, is largely a higher-order relay, since much of it relays information from layer 5 of one cortical area to another. All thalamic relays receive a layer-6 modulatory input from cortex, but higher-order relays in addition receive a layer-5 driver input. Corticocortical processing may involve these corticothalamocortical 're-entry' routes to a far greater extent than previously appreciated. If so, the thalamus sits at an indispensable position for the modulation of messages involved in corticocortical processing.

/pdf/the-role-of-the-thalamus-in-the-flow-of-information-to-the-3sleuyixtg.pdf

The role of the thalamus in the flow of information to the cortex

The thalamus has long been seen as responsible for relaying information on the way to the cerebral cortex, but it has not been until the last decade or so that the functional nature of this relay has attracted significant attention. Whereas earlier views tended to relegate thalamic function to a simple, machine-like relay process, recent research, reviewed in this article, demonstrates complicated circuitry and a rich array of membrane properties underlying the thalamic relay. It is now clear that the thalamic relay does not have merely a trivial function. Suggestions that the thalamic circuits and cell properties only come into play during certain phases of sleep to effectively disconnect the relay are correct as far as they go, but they are incomplete, because they fail to take into account interesting and variable properties of the relay that, we argue, occur during normal waking behavior. Although the specific function of the circuits and cellular properties of the thalamic relay for waking behavior is far from clear, we offer two related hypotheses based on recent experimental evidence. One is that the thalamus is not used just to relay peripheral information from, for example, visual, auditory, or cerebellar inputs, but that some thalamic nuclei are arranged instead to relay information from one cortical area to another. The second is that the thalamus is not a simple, passive relay of information to cortex but instead is involved in many dynamic processes that significantly alter the nature of the information relayed to cortex.

Functional organization of thalamocortical relays

When one nerve cell acts on another, its postsynaptic effect can vary greatly. In sensory systems, inputs from “drivers” can be differentiated from those of “modulators.” The driver can be identified as the transmitter of receptive field properties; the modulator can be identified as altering the probability of certain aspects of that transmission. Where receptive fields are not available, the distinction is more difficult and currently is undefined. We use the visual pathways, particularly the thalamic geniculate relay for which much relevant evidence is available, to explore ways in which drivers can be distinguished from modulators. The extent to which the distinction may apply first to other parts of the thalamus and then, possibly, to other parts of the brain is considered. We suggest the following distinctions: Cross-correlograms from driver inputs have sharper peaks than those from modulators; there are likely to be few drivers but many modulators for any one cell; and drivers are likely to act only through ionotropic receptors having a fast postsynaptic effect whereas modulators also are likely to activate metabotropic receptors having a slow and prolonged postsynaptic effect.

/pdf/on-the-actions-that-one-nerve-cell-can-have-on-another-4785k0yd1n.pdf

On the actions that one nerve cell can have on another: Distinguishing “drivers” from “modulators”

http://issalab.uchicago.edu/sherman/gsneuron2002.pdf

Thalamic Relay Functions and Their Role in Corticocortical Communication: Generalizations from the Visual System

Preface Abbreviations Used Introduction The Nerve Cells of the Thalamus The Afferent Axons to the Thalamus Intrinsic Cell Properties Synaptic Properties Function of Burst and Tonic Response Modes in the Thalamocortical Relay Maps in the Brain Two Types of Thalamic Relay Drivers and Modulators Overview Bibliography Index

R. W. Guillery

Papers

The role of the thalamus in the flow of information to the cortex

Functional organization of thalamocortical relays

On the actions that one nerve cell can have on another: Distinguishing “drivers” from “modulators”

Thalamic Relay Functions and Their Role in Corticocortical Communication: Generalizations from the Visual System

Exploring the Thalamus