Marcus, Chapter 5: Individuals
• Jan Peters: jpeters@uos.de
• Antje Petzold: apetzold@uos.de
• Stefan Scherbaum: sscherba@uos.de
Structure
• Context within the Book
• Introduction: Types & Tokens
• Object Permanence:
– Individuals (Individuation) – Tracking (Identification) over
time
• Records
• Summary
Guide
• Moderation and a little bit of theory (Jan)
• Experiments (Antje)
• Models (Stefan)
Context within the Book
• Marcus aim is to „integrate research in connectionism with a clear statement of what symbol manipulation is“ (p.2)
• He lists separate hypotheses of symbol manipulation:
– Operations Over Variables – Structured Representations
– Representation of Individuals and Kinds
• Topic today is, to what extend humans and models (connectionist and symbol systems) do represent individuals and kinds.
Types & Tokens
• Corresponds to:
– Instance vs. Class – Individuals vs. Kinds – ...
• Marcus argues: people can distinguish types and tokens.
– we can refer to types (dogs, presentations, power point slides...) – we can also refer to particular tokens (this particular slide etc.)
• Tokens and types are not independent!
– People treat objects depending on their type.
– Mental representations of tokens determine the mental representation of a type.
• A system that can represent tokens has a way of performing:
– Individuation (ability to select a particular instance)
– Id. over time (is this X the same as the X I was confronted with earlier?)
Types & Tokens II
• Let‘s take a look at how MLPs could represent types and tokens!
MLPs and Types & Tokens I
Is there individuation in MLPs?
à two possibilities:
– kinds properties
MLPs and Types & Tokens II
• Felix, a cat
MLPs and Types & Tokens II
• Morris, another cat
MLPs and Types & Tokens III
à No distintinction between similar individuals possible:
A cat is a cat is a cat is a à two-horses problem
MLPs and Types & Tokens IV
Possible Solution: an extra Felix and Morris node à Two kinds of nodes:
à Perceptual à can grow naturally
à Associative à needs extra mechanism to grow à not plausible
à Classical grandmother node issue?
à Not reasonable
MLPs and Types & Tokens Critique
• Granularity of properties determines similarity
• Humans can‘t distinct „extremely similar“ individuals (twins)
• Identity by similarity depends on the level of granularity à Identity does not exist by itself
Object Permanence
• The main issue of chapter 5 is object permanence.
• Criticism: Marcus talks about object permanence for almost 10 pages and then starts a section with that very heading.
• Object permanence is the brain‘s function of keeping track of individual objects. It requires:
– A representation of individuals (tokens).
– A process that tracks these individuals in space and time.
• For both aspects, we will in the following sections
– look at empirical evidence from infants.
– look at how connectionist models perform these functions.
Introduction: Individuals
• In the next section, we‘ll look at
– if infants can in fact represent individuals (tokens) or if they only represent kinds (types).
– one connectionist model that does not explicitly represent tokens and types distinctly.
Identifying Individuals vs. Kinds
• Hypothesis:
Children are able to track the persistence of particular objects, not just kinds
– occluded rods (Spelke et al., 1995) – Mickey Mouse (Wynn, 1992)
Identifying Individuals Rods
Spelke, Kestenbaum, Simons, Wein (1995)
Identifying Individuals Rods
Expected Outcome
Spelke, Kestenbaum, Simons, Wein (1995)
Identifying Individuals Rods
Unexpected Outcome
Spelke, Kestenbaum, Simons, Wein (1995)
Identifying Individuals Mickey Mouse
Wynn (1992)
Identifying Individuals Mickey Mouse
Wynn (1992)
Identifying Individuals Mickey Mouse
Wynn (1992)
Identifying Individuals Mickey Mouse
Results: longer attendance to unexpected outcome
Wynn (1992)
Identifying Individuals and MLPs
• Attempt to capture aspects of object permanence
• Task: MLP represents object even without any perception
Identifying Individuals and MLPs
Identifying Individuals and MLPs
Marcus‘ critique:
- No training indepedence à not able to generalize - No distinction between
object permanence ßà object replacement - No distinction between individuals
- Comparison to brightness sensors:
can object permance sense be learned or must it grow (à inate structure)
Identifying Individuals and MLPs Critique
• object permanence ßà object replacement: can humans do that?
à Sometimes human can‘t, but in principle they are able to distinguish
• training independence: treelets can not generalize
• assumption of inate structure for object permanence
Summary: Individuals
• We have seen that infants apparently have a representation of tokens/individuals, and that they do not only represent
types/kinds.
• The model for object permanence that we have seen lacks an explicit representation of individuals. It performs
– spatio-temporal tracking but does not keep a record of
– individual perceptual features of an object.
• Criticism: Connectionist models lack explicit representation of types and tokens, argues Marcus. If this is their only shortcoming, Marcus could simply propose an alternative connectionist model which does that.
Introduction: Tracking over Time
• When it comes to tracking an object over time, two aspects are important:
– the object‘s perceptual features/properties.
– the spatio-temporal aspects of the object (where it is at what time)
• Next we will look at
– what role these aspects play in infant/human cognition.
– what happens, when one tries to build an MLP model for tracking an object over time.
Tracking Individuals Over Time
• Hypothesis:
spatiotemporal information trumps featural property information
– apparent motion (Michotte, 1963)
– tracking moving targets (Scholl et. al, 1999) – „Zavy“ (Sorrentino, 1998)
Tracking Individuals Over Time
Tracking Targets
Tracking Individuals Over Time
? spatiotemporal vs. featural properties ?
Tracking Individuals Over Time
„Zavy“
Key argument for:
people use spatiotemporal cues rather than featural properties
Christina Sorrentino (1998)
MLPs and tracking over time I
Can a MLP solve the Zavy task?
MLPs and tracking over time II
Possible solution: including information about location
Reason:
MLP learning is based on
- events - properties
à No spatiotemporal information included
à Only driven by correlations between labels and properties
MLPs and tracking over time II
Tracking Individuals Over Time
„Zavy“
Christina Sorrentino (1998)
• No spatiotemporal information provided to MLP
à Possible solution: action operators as nodes
• Are humans able to solve this task without this info?
à Refers back to granularity of properties
• Humans seem to choose the most appropriate info:
sometimes properties, sometimes spatiotemporal info
à depends on task/ what is relevant
MLPs and tracking over time
Critique
Summary: Tracking over Time
• Again: There are two factors when it comes to tracking an object over time:
– that object‘s perceptual features/properties
– spatio-temporal information concerning that object.
• In infants, spatio-temporal information apparently overrules perceptual feature information, but only in tasks thus designed
Summary: Models
• What can we say about the connectionist models that we have seen so far?
– some represent perceptual features (Zavy model) – some keep track of spatio-temporal change
• However, in order to really perform object permanence, a model would be required to do both!
• We‘ll now look at some symbol-manipulating models that (according to Marcus) capture the notion of object permanence better.
Symbol Systems That Represent Individuals I
• MLPs do not represent individuals How to represent individuals?
à Records for each individual
•Creation rules for new record
•Storage for
• properties
• ID
• Marker for occlusion
• Model of Simon (1998):
•Can simulate Wynn‘s (mickey mouse) results
•Could be able to simulate Sorrentino‘s (zavy) results
Symbol Systems That Represent Individuals II
These systems only work, if objects are tracked permanently à Otherwise they need real-world knowledge
à Even then, they are cheatable But:
object permanence ßà object replacement
is represented à Records as base for best performance, even if not omniscient
The Key Issue: Records
• We have seen, that the symbolic models could account for object permanence, whereas the connectionist models presented by
Marcus could not.
• Marcus argues, that this is because:
– Symbolic models have a database as a component.
– This database can serve as a module to keep track of spatio-temporal changes in object properties.
• This database is, according to Marcus, the key to solving the object permanence problem, because MLPs do not have this module to track spatio-temporal change in an object.
• What could such a database look like?
Records as Semantic Nets
Semantic nets are an examplar format for encoding records.
The semantic net before ... and after inserting a new fact.
Neural Implementation
• Marcus‘ suggestions: Treelets
• Treelets
+ based on records, represent kinds easily à easily adaptable to representing individuals
• need for identifier (kind vs. individual),
e.g, two parts – kind information + unique code
àMarcus‘ claim: just current properties of an individual not suitable as identifier
àMarcus‘ suggestion: e.g., two different types of neural connections or a one-bit-register
Representing Individuals in Treelets
• representing a new fact about an individual
– not a matter of new nodes or new connections between nodes but of adding a new relation
Neurobiological Plausibility
• Registers as basis for treelets
• Register: rapidly updatable memory
possible neural implementations:
– autaptive cells (Trehub, 1991),
hexagonal self-excitatory cell assemblies (Calvin, 1996) – information storage intracellular (e.g, gene expression)
? Plausibility of Such Registers and Treelets ?
- locally encoded memory vs. distributed memory (neuroimaging studies) - What (neural) mechanism could manipulate treelets?
- treelets provided by the brain à limited memory capacity?
Summary
• People can do object permanence!
• Symbol systems that have records have the potential of performing object permanence.
• MLPs have a problem with object permanence.
– The main problem that the MLPs (discussed by Marcus) have is that they do not jointly represent perceptual object features and keep track of spatio-temporal features.
– But: It is not clear that MLPs cannot be adapted accordingly.
– But: Marcus derives his demands for models from human (infant)
performance. But the connectionist models are, unlike for instance ACT- R, no unified model for cognition in general but rather at a model component level.
Final Discussion
• Remarks?
• Questions?
• What do you think of Marcus‘ ideas so far?
– Does he address the appropriate issues when defining criteria for a cognitive architecture?
– Do you think the treelet model is plausible
• with respect to representation of facts
• with respect to neurobiological structure
• with respect to findings of neuroimaging studies?
• Proposals?
– How about: the brain encodes only individuals, generalizing kinds from these?
Final Remarks of the Audience
• Treelets seem to be made for english language
à do eskimos use treelets structered in another way?
• Feature Lookup-Problem:
à If I want to get the ID for the record of an individual, I need to search it in the database by its properties (“select * where…”) à ID only postpones the problem of similarity
Literature
• Literature, referring to bibliography of marcus:
– Munakata et al. 1997 – Michotte, 1963
– Scholl et. al, 1999 – Sorrentino, 1998
– Spelke, Kestenbaum, Simons, Wein (1995) – Wynn, 1992