Chris Eliasmith - How to Build a Brain: A Neural Architecture for Biological Cognition - 2015

History / Edit / PDF / EPUB / BIB /
Created: August 17, 2017 / Updated: November 2, 2024 / Status: in progress / 11 min read (~2033 words)
Artificial General Intelligence

  • Explaining reaction times addresses only a small part of cognitive system dynamics in general, and the ACT-R explanations rely on assuming a 50 ms "cycle time," which itself is not explained
  • If you want to build a fast cognitive system - one that directly interacts with the physics of the world - then the most salient constraints on your system are the physical dynamics of actions and perception
  • What is a cognitive system? A system that integrates all aspects of sophisticated behavior. It is a system that must simultaneously realize a wide array of biological behaviors
  • Good questions outrank easy answers

  • I consider a cognitive architecture to be "a general proposal about the representations and processes that produce intelligent thought"
  • In the last half-century, there have been three major approaches to theorizing about the nature of cognition
    • The classical symbolic approach (symbolicism or Good Old-fashioned Artificial Intelligence)
    • Connectionism (Parallel Distributed Processing)
    • Dynamicism: Cognitive systems are non-representational, low-dimensional, dynamical systems. Cognitive systems can only be properly understood by characterizing their state changes through time
  • Classical control theory characterizes physical systems by identifying a "transfer function" that directly maps inputs to outputs
    • Does not consider the internal states of a system
  • Once cognitive scientists began to think of minds as computers, a number of new theoretical tools became available for characterizing cognition
  • It suggested novel philosophical theses
    • Functionalism: the notion that only the mathematical function computed by a system is relevant for its being a mind or not
    • Multiple realizability: the notion that a mind could be implemented in pretty much any substrate - water, silicon, interstellar gas, etc. - as long as it computed the appropriate function

  • What are the Core Cognitive Criteria (CCC)?
    • Dynamicism: Cognition is distinguished from other kinds of complex natural processes... by at least two deep features: ... a dependence on knowledge; and distinctive kinds of complexity, as manifested most clearly in structural complexity of natural languages
    • Connectionism: They must explain motor control, perception, memory, and language
    • Symbolism: problem solving, decision making, routine action; memory, learning, skill; perception, motor behavior; language; motivation, emotion; imagining, dreaming, daydreaming
  • Commonalities among these descriptions of cognitive behavior:
    • Language
    • The importance of adaptability and flexibility
    • Motor control and perception as important to understanding cognition
  • Three explicit constraints on what it takes to be a cognitive system (according to Fodor and Pylyshyn)
    • Productivity: The ability of a system to generate a large number of representations based on a few basic representations (a lexicon) and rules for combining them (a grammar)
    • Systematicity: The fact that some sets of representations are intimately linked
    • Compositionality: The meaning of complex representations is a direct composition (i.e., adding together) of the meaning of the basic representations
  • Jackendoff challenges to address when explaining cognition:
    • The massiveness of the binding problem (i.e., that very many basic representations must be bound to construct a complex representation)
    • The problem of two (i.e., how multiple instances of one representational token can be distinguished)
    • The problem of variables (i.e., how roles in a complex representation can be generically represented)
    • How to incorporate long-term and working memory into cognition
  • Don Norman description of human cognition
    • Robust
    • Flexible
    • Reliant on "content-addressable" memory
  • Two of the most important considerations for good scientific theories are those of unity and simplicity
    • Unity: The more sources of data, and the more scientific disciplines that they are consistent with, the better the theory

  • How is semantic captured in the system?
  • How is syntactic structure encoded and manipulated by the system?
  • How is the flow of information flexibly controlled in response to task demands?
  • How are memory and learning employed in the system?
  • A semantic pointer is a neural representation that carries partial semantic content and is composable into representational structures needed to support complex cognition
  • The relation between a semantic pointer and what it points to can be usefully thought of as a kind of compression
  • Any characterization of a cognitive system has to provide some explanation for how relevant information is propagated through time, and how the system can adapt using its past experience

  • The following three principles form the core of the Neural Engineering Framework(NEF)
    • Neural representations are defined by the combination of nonlinear encoding (exemplified by neuron tuning curves and neural spiking) and weighted linear decoding (over population of neurons and over time)
    • Transformations of neural representations are functions of the variables represented by neural populations. Transformations are determined using an alternately weighted linear decoding
    • Neural dynamics are characterized by considering neural representations as state variables of dynamic systems. Thus, the dynamics of neurobiological systems can be analyzed using control (or dynamic systems) theory
  • It seems that neurons tend to encode approximately two to seven bits of information per spike
  • It is reasonable to begin with the hypothesis that "higher-level" properties are inferred on the basis of representations of properties more like those that physicists talk about
    • In other words, the encoding of "edible" depend, in some complex way, on encodings of "lower-level" physical properties like wavelength, velocity, and so forth
  • One crucial aspect of the NEF principle of representation is that it can be used to characterize arbitrarily complex representations
  • A second crucial aspect of this principle is that it distinguishes the mathematical object being represented from the neurons that are representing it
    • I refer to the former as the "state space" and the latter as the "neuron space"
  • A third crucial aspect of this principle is that it embraces the heterogeneity of neural representation
  • A transformational decoder is a kind of biased decoding. That is, in determining a transformation, we extract information other than what the population is taken to represent
  • In general, we can characterize non-linear functions of multiple scalars or vectors as linear decoding of some higher-dimensional representation in a population of neurons

  • Some have suggested that because the reduction of one science to another has failed, the sciences must be independent
  • Consequently, Fodor has argued that to understand cognitive systems, which lie in the domain of psychology, appeal to lower-level sciences is largely useless
  • Levels should be taken as pragmatically identified sets of descriptions that share assumptions
  • Most systems composed of parts have parts that interact
  • If the parts of a system are organized in such a way that their interaction results in some regular phenomena, then they are called "mechanisms"
  • Mechanisms are often analyzed by decomposing them into their simpler parts, which may also be mechanisms

  • Perhaps the most compelling feature of biological cognitive systems is that they can perform many different tasks, without being "reprogrammed"
  • Four distinct aspects of cognition: semantics, syntax, control, and memory and learning
  • Semantic pointers are a means of relating more general neural representations to those central to cognition

  • Higher-level cognitive functions in biological systems are made possible by semantic pointers. Semantic pointers are neural representations that carry partial semantic content and are composable into the representational structures necessary to support complex cognition
  • Traditionally, a classical cognitive system is characterized as a symbol processing system that relies on syntax to respect the semantics of the symbols and combinations of symbols found in the system
  • In more connectionist approaches, semantics has been the focus, where the vector space defined by the activity of the nodes in the network was often though as a "semantic space" that related the meaning of different firing patterns
  • In the SPA, syntax is inspired by proponents of the symbolic approach who claim that there are syntactically structured representations in the head, and semantics is inspired by connectionists who claim that vector space can be used to capture important features of semantics
  • It is natural to begin with the claim that semantic pointers are well described by vectors in a high-dimensional space
  • The hypothesis suggests that the brain manipulates compact, address-like representations to take advantage of the significant efficiency and flexibility afforded by such representations
  • The "semantic" in "semantic pointer" refers to the fact that the representations that play the role of a pointer contain semantic information themselves. That is, the relationship between a semantic pointer and the more sophisticated data structure to which it points is not arbitrary. Specifically, the semantic information that is contained in a semantic pointer is usefully thought of as a compressed version of the information contained in a more complex representation

  • Dual-Coding Theory: perceptual and verbal information are processed in distinct channels
    • Linguistic processing is done using a symbolic code, and perceptual processing is done using an analog code, which retains the perceptual features of a stimulus

  • Semantic pointers are lossy compressions

  • Because the real world is extremely complex, the ideal statistical model will also be enormously complex (as it is the probability of all possible data at all times). As a result, the brain probably approximates this distribution by constructing what is called a parameterized model. Such a model identifies a small number of parameters that capture the overall shape of the ideal model
  • The higher layer of this network (an deep neural network for MNIST) represents a 50-dimensional space. Semantic pointers in this 50D space carry information about the presented digits. Clearly, however, this representation does not contain all of the information available in early visual areas. Rather, it is a summary that can be used to perform an object recognition task
  • I would argue that capturing deep semantics and relating them to high-level representations solves the symbol grounding problem - if we can show how those high-level representations can function like symbols

  • The motor system does not need to classify presented stimuli at all; rather, it needs to direct a nonlinear, high dimensional system towards a desired state
  • We might notice that perceptual systems often need to map from a high-dimensional, ambiguous state to a much smaller set of states, whereas motor systems often need to map from a small set of states to a high-dimensional, ambiguous state
  • Central computational features are shared by perception and motor control: both need to map low- to high-dimensional states; both need to deal with complexity and nonlinearity; both need to deal with uncertainty and ambiguity; and both need to share information between the past and the future
  • We can think of lower levels in the motor hierarchy as having models of the higher levels. Lower levels can then use such models to determine an appropriate control signal to affect the behavior specified by the higher level
  • There is no such thing as static movement. So, time is unavoidable. Embracing this observation actually renders the connection between perceptual and motor hierarchies even deeper - higher levels in both hierarchies must model the statistics and the dynamics of the levels below them
  • There are now two main features of the motor system that mirror the perceptual system:
    • the (dynamical) model constructs representations of the domain that capture the statistical relationships that sufficiently describe the domain's structure
    • such representations are "compressed" representations as we move up the hierarchy

  • For the motor hierarchy, the forward direction (down the hierarchy) allows for the control of a sophisticated body using simple commands, as well as the generation of deep semantics
  • The reverse direction allows online adaptation of the motor command to various perturbations
  • Both directions will allow the system to learn appropriate representations
  • It is a mistake to think of biological systems as processing perceptual information and then processing motor information. Rather, both are processed concurrently, and inform one another "all the way up" their respective hierarchies. Consequently, it is more appropriate to think of the semantics of items at the top of both hierarchies as having concurrent perceptual and motor semantics
  • It is much more appropriate to conceive of the entire perception/action system as being a series of nested controllers, rather than a feed-in and feed-out hierarchy

  • Eliasmith, Chris. How to build a brain: A neural architecture for biological cognition. Oxford University Press, 2013.