School of Computer Science THE UNIVERSITY OF BIRMINGHAM CN-CR Ghost Machine

HBP/NeuroRobotics: A slightly skeptical look
from the standpoint of the Meta-Morphogenesis project
(DRAFT: Liable to change)

Aaron Sloman
School of Computer Science, University of Birmingham.
(Philosopher in a Computer Science department)


GENERIC ABSTRACT for talk at workshop 8-9 Jan 2014
A (possibly) new way to approach AI/Robotics/Cognitive Science

(Some subset of the following will be presented.)

Background: The Human Brain Project is one of two very large long term 'flagship'
projects recently selected for funding by the European Commission, summarised here:
One of the sub-projects (SP10) is Neurorobotics, described very briefly here:
It aims to develop one of the six platforms to be produced by the HBP

[This is too long, but shortening it would take more time than I have available. Sorry.]

My presentation will use the standpoint of the Meta-Morphogenesis project to draw
attention to problems of understanding requirements for systems to be developed
in such an ambitious project, illustrated by some of the achievements of biological
evolution that cannot easily be identified using current research methods in
neuroscience, psychology, cognitive science, linguistics, AI, Robotics, ethology,
philosophy etc.

This approach was inspired by the challenge of combining Turing's early work on
digital computation (on Turing machines) with the work he published shortly
before his death on chemical morphogenesis. I suspect that if he had lived he
might have tried to use the combination of ideas to answer one of the great
unanswered questions of science: how could a lifeless planet with no information
available about forms of life, their requirements, their possible designs,
produce the diversity of life forms found on our planet including many highly
intelligent animals, among them human mathematicians.

Doing the kind of mathematics that led to Euclid's elements is closely connected
with being able to perceive, reason about, and make use of what Gibson called
affordances in the environment, though I think there were more types of
affordance than Gibson recognised, because he was focusing on relatively
primitive forms of behaviour.

Perception of the full range of affordances (e.g. affordances for gaining
information, affordances for changing information available to others,
affordances confronting one's offspring who may need help, and many more) seems
to require information-processing mechanisms whose capabilities are very
different from current AI systems and robots. Bridging that gap seems to be one
of the implicit aims of the HBP even if it hasn't been mentioned explicitly in
the project proposal, as far as I know. Some of the problems are presented in
this discussion:
Hidden Depths of Triangle Qualia

Depending on what I learn from other participants at the meeting, I shall
provide reasons (some of them in the documents listed below) for thinking that
there are many aspects of the ways in which brains of humans and many other
species work that cannot be identified by physical or other measurements of
brain activity or experiments on humans and other animals. Such 'probes' merely
produce tiny samples from a vast store of required information about problems
solved by evolution over many millions of years, and additional problems solved
by epigenetic mechanisms to which evolution delegated important functions.

For similar reasons it would be very difficult for alien scientists to work out
what's going on in the World Wide Web by sending teams of researchers to take
measurements all over the planet, including setting up experiments in the
vicinity of devices connected to the internet. They might be able to make
progress if they had independently developed a similar system and understood
such topics as the need for machine languages, compilers and interpreters, a
variety of programming languages for different purposes, a host of types of
virtual machinery (including platform VMs such as operating systems, and
application VMs such as word-processors, email handlers, chess and other
programs, etc.), various types of concurrency, various types of inter-process
communication, various kinds of interrupts, many layers of protocols of various
sorts, the problems of security and mechanisms that might be used to address
security issues, mechanisms allowing the system to change and grow, and many

Brains have additional problems of control because different parts of the
internet can have different physical locations and perform quite unrelated
parts, whereas animal sensors and effectors are far more constrained. TV cameras
connected to the same network can be scattered over wide terrain or attached to
multiple mobile devices, whereas eyes, hands, tongue and other sensors are
constrained by body size, shape and location on the same body. So animals face a
recurring need to decide where to look, what to touch, where to go next, etc.
For these reasons, values, preferences, policies, desires, plans, intentions and
related control mechanisms are needed for dealing with competing needs on
various time-scales, including simultaneous control of foveal fixations and body
parts such as grippers that can interact with parts of the environment. This
requires an architecture in which components can at any time be influenced in
quite detailed ways by information from other components.

As far as I know there are no artificial working visual systems or language
understanding systems that meet such requirements (not least because designers
tend to work on isolated subsystems to be assembled later), and nobody knows how
brain mechanisms support such tightly integrated, mutually influencing
interfaces and mechanisms.

One of the consequences of having a rich repertoire of actions and a variety of
sensors in a very rich and changing environment -- often presenting new
locations, new spatial configurations, new processes in which multiple objects
interact -- is that the space of possibilities is too vast to be covered by
current forms of learning, e.g. using pre-labelled images. Somehow organisms
confronted with such variety have to develop generative theories that
enable novel configurations to be parsed, interpreted and related to current
goals, plans, preferences, needs, etc. (This is a generalisation of the point
Chomsky made in the 1960s about the need to be able to cope with novel
sentences, such as many of the sentences in this document.)

The assumption that all sensor contents, motor signals, and current internal
states can usefully be represented as vectors of scalar measures (with fixed
dimensionality) as required by many current learning systems is an assumption
that just does not fit the changing complexity of actions and environments of
many animals. We do not seem to have good theories about the forms of
representation used by brains or minds to cope with this diversity, although
verbal descriptions, parse trees, collections of logical formulae, networks,
graphs, and various kinds of dynamical systems may provide hints.

In particular, the fact that we cannot get current computers to replicate the
kinds of geometrical discoveries leading up to Euclid's elements, seems to be
closely related to our failure so far to give machines the ability to perceive
and understand the rich variety of types of affordances (collections of
possibilities, constraints on possibilities, invariants across process types)
required for intelligent perception and action. J.J.Gibson introduced the notion
of 'affordance' but explored only a small subset of cases. I've tried to point
out the need to go far beyond Gibson in this presentation on the functions of
"What's vision for, and how does it work?
   From Marr (and earlier) to Gibson and Beyond"

These issues are not merely relevant to the task of trying to understand, model
or replicate human brain function. They are also relevant to the problems of
designing useful future robots, for instance personal assistants, or robot
carers for the ill or elderly. I have presented some of the problems in this
paper (published in a book on Artificial Companions edited by Yorick Wilks):
"Requirements for Digital Companions: It's harder than you think"

A vast amount has been written about consciousness, with different authors
presenting very partial views of what the problems are, what the possible
answers might be like and what sorts of mechanisms could be involved. If instead
we try to understand how the phenomena that we are interested in could have
resulted from biological needs and solutions provided by evolution, building on
the mechanisms that were previously available, this could lead us to much better
theories than are currently available. In particular, instead of assuming
that the noun 'consciousness' refers to one thing, so that we can ask what 'it'
is, how 'it' evolves, what brain mechanisms enable 'it', etc. we should focus on
the adjective, in contexts of the form 'X is conscious of Y', allowing X and Y
to vary as widely as possible. This can lead to a theory of consciousness as a
highly polymorphic phenomenon with many different functions in different
organisms or in different problem situations, with different supporting
mechanisms required. When we have a good theory we can try to see how it maps on
to what is known about brain mechanisms (and the vast array of information about
different sorts of consciousness and influences on consciousness, including
drugs, exhaustion, sensor damage, brain damage, and 'software' problems of
control in various kinds of psychological disorder). These are not merely
esoteric matters to be left to philosophers and medical practitioners: they are
required for understanding many aspects of natural information processing and
for designing versatile and effective robots.

My impression gained at the workshop is that some members of the project tend to think
about architectures in terms that are much too simplistic -- e.g. as if brains were
mostly concerrned with managing ``sensori-motor loops''. I have argued over many
years that evolution produced a succession of co-existing architectural layers
performing different sorts of functions that could be subdivided in many ways, e.g.
``horizontally'' in terms of the kinds of environments, tasks, modes of learning,
modes of perception, modes of action, modes of interaction with different sorts of
things in the environment and, vertically in terms of three overlapping ``pillars''
of perception, action and more central functioning (e.g. learning, managing of
motivation, resolving conflicts, self-observation, etc.).

I have sometimes divided the three layers using the labels ``reactive'', ``deliberative''
and ``meta-management'' (partly based on work by Luc Beaudoin's PhD thesis (1994).
These layers need to be supplemented with ``alarm'' mechanisms. Some subsytems have
to straddle all the layers, e.g. the mechanisms involved in human linguistic

Some of these ideas are presented in connection with the Cognition and Affect
project, here

Different kinds of ``functionalist'' models of mind, related to this, are summarised

Some of Marvin Minsky's architectural ideas in The Emotion Machine (2006) are closely

Background information for the presentation


Installed: 3 Jan 2014
Last updated: 4 Jan 2014;9 Jan 2014

This document is
A PDF version is also available, though it may not be fully up to date:

A partial index of discussion notes is in

Maintained by Aaron Sloman
School of Computer Science
The University of Birmingham
rcol 80