| |
The Lucy project is based on a whole bunch of ideas about the mammalian
brain, many of which are pretty radical and flaky, at least from the perspective
of Artificial Intelligence. The following are some of the key assertions /
assumptions that lie behind the project, which only those of you with a
particularly technical bent are likely to find interesting!
Main theoretical stance and assertions:
- What's conventionally called the feed-forward direction (from the senses
towards the motor systems) is really the feedback direction.
- Cortex comprises a single hierarchy-column-map-layer structure that is a
universal basis for all its computations.
- Perception is an active process akin to motor programming
- Prediction, intention, working memory & attention are fundamentally
related concepts at a neurological level
- The brain is a system in perpetual tension – a dynamic equilibrium that
takes time to build up but is quick to adjust. This state and its changes
constitute a 'narrative' about the organism's body and environment
- The data are often not anchored to the neurons. Neurons generate a
virtual machine composed of patterns of nerve activity. It is at this level
of description that the brain needs to be understood.
- Mass action is common, including convolution, Fourier/holographic
behaviour, population codes, interference, travelling and standing wave
phenomena, phased signals, and quasi-particles.
- The most fundamental function of the brain is to build predictive
models, which compensate for signal delays and the intention-action time
lag, and also deal with sensory and motor ambiguity.
- Cortical maps can often be considered as complete sensorimotor systems,
rather than having distinct sensory or motor functions. Each map should be
thought about in terms of the contribution it makes to behaviour, e.g. V1
might be extending and subsuming the SC’s own saccade control in order to
respond to non-moving stimuli, while M1 (and preceding maps) can be viewed
as extending cerebellar and brainstem control of movement.
- No information can be lost during sensory abstraction, even if it
becomes highly abstracted or convolved. Any blurring of signals must
therefore be performed in a way that would permit the original signal to be
reconstructed, even if no such reconstruction is actually performed.
- Perception may be invariant but it isn't blind - we still know which way
up something is, even though we discount its orientation when recognising
it.
- Two data sources need to be in the same coordinate space before they can
be correlated; modal streams will therefore tend to meet as soon as they
share a common coordinate frame.
- The brain acts in order to keep the bottom-up and top-down (yin and
yang) signal streams in balance. It does this in the manner of a non-linear
servomotor, comparing the top-down intention/expectation with the bottom-up
sensory information and trying to reconcile the two by causing either a
change in the internal model or in the outside world.
- Memory plays an equivalent role to physics – it is a reflection of the
outer world within the inner world. This inner reflection acts as a
counterpart to negate the effects of the laws of physics in the outside
world, rendering the organism homeostatic.
Some architectural speculations that play a big part in the design:
- The yin and yang pathways are roughly analogous to an optical system
composed of several stages. Each stage is a lens + diffraction
filter/partial mirror + lens. At the end of the chain is a final 'mirror'.
- Learning which 'spot of light' is your own reflection may be the means
by which reciprocal wiring develops (such wiring has to be present if a
mental image can be generated either by an external stimulus or by an
internal thought).
- Cortical layer IV is fundamentally a means of collapsing n-dimensional
input space into a 2D arena that sets the characteristic geography for the
activity in that map. It may also have signal-processing features such as
discounting the illuminant or reflecting feature type while ignoring
intensity.
- Pinwheels of orientation selectivity show up everywhere – certainly V1,
V2, V5, M1, S1. Such a scheme can code for up to three values in a vector –
position, direction and magnitude.
- There is an approximate tripod hierarchy of cortical maps – occipito-temporal,
occipito-parietal and frontal, joined at the hippocampus and other limbic
structures. However, this is not a simple tree, and many lateral connections
exist.
- The hippocampus is the junction of the tripod and constitutes a reduced
cortical map, concerned with making rapid, short-term, multimodal links.
These equate to episodic memories, and become semantic or procedural
memories later on, through dream-driven or experiential (statistical)
relearning lower in cortex.
- Attention is a limited resource – only one peak of attention is possible
down any yang path. Attention is associated with anticipatory or sequencing
functions. Attention signals begin as a reflection down a yang path
stimulated by the triggering of a memory by a yin signal. If attention is
triggered low down in the tripod it acts unconsciously, but if it reaches
high into the trunk(s) of the hierarchy it prevents any top-down control of
attention and hence becomes conscious. Attention may also be triggered
top-down and counts as mentation unless or until it reaches the periphery
and triggers action. The higher echelons of the hierarchy form a bottleneck
- the higher the level, the more whole-body and therefore conscious a
thought becomes.
- Every cortical map has a coordinate frame, a meaning and a use. The
coordinate frames are defined through dynamic tension between yin and yang
paths, initially during development and then later 'refreshed' during
slow-wave sleep. In both cases it is spontaneous wave activity (caused
perhaps by rising gain in de-habituated un-driven neurons) that provides a
set of 'test card' signals arranged in sensory order (retinotopic, tonotopic,
homuncular). The tension between maps (via a judicious mixture of test-card
and environmental data) causes a 'morph' of coordinate frames through the
system, such that each map’s input and output frame discovers a minimal
mapping and hence allows useful correlations to be acquired through
learning. REM-sleep is the yang equivalent to waking experience.
- Mappings perhaps occur through the pruning of wide axon arbors.
Something akin to a bubble sort algorithm allows for the development of 1:1
mappings.
- Cul-de-sacs are possible, due to internally generated test-card data.
These cul-de-sacs correspond to 'virtual' sensorimotor maps and are possibly
where the more abstract “symbolic” manipulations occur, such as arithmetic.
- “Recognition” is a multi-map process, involving a unique “route” from
point to point on several maps. The best way to recognise such a route is to
copy it, so yang and yin paths (in temporal lobes at least) will often
match.
- Action sequencing occurs using yang signals, which specify a desired
state (in the coordinate frame of a given map). This map’s task is then to
specify desired states in lower maps, or directly in muscles or subcortical
structures, to act as “subroutines”. Completion of each step in the
subroutine triggers the next step (perhaps via loosely coupled rules),
eventually resulting in an actual state that matches the desired state. This
then signals the end of that step to the higher map, and so on. Sensory
attention and anticipation operate in an essentially similar way.
- Mass action is a crucial component in processing. Signals are frequently
convolved with each other in a “flat” distribution, with the PSF being quite
large (106 cells, or 4mm dia, in V1). No data must be lost, even
if the signals are not reconstructed.
- In many cases this leads to wave or dome activity, where the domes etc.
are mobile (neither stationary nor travelling) and can be used as “symbols”
in working memory (e.g. as desired and actual states in motor control). They
may interact with each other in wave mechanics ways (e.g. interference or
correlation) that perform computation (such as servoing or coordinate
transforms).
- Other mass transformations of activity patterns may account for mental
image rotation, etc. and may perhaps involve a phased array mechanism.
- Controlled thought, as opposed to learned reflexes and specialised
processing pathways through the cortical hierarchy, may involve a 'map of
cortical maps' that permits signals to be re-routed. Thus, specialised
cortical modules can be re-purposed (e.g. for deliberate mental rotation of
an image).
|