The symbol grounding problem and recurrent neural networks with parametric biases

The symbol grounding problem (apparently) is: assuming that someone’s “cognitive” levels of mind work in terms of symbols, how to design the interface of the symbolic levels with the low-level sensory-motor systems?

Jun Tani suggests a “recurrent neural network with parametric biases” (RNNPB).

I haven’t had time to read further yet, but it looks very interesting so I’m passing it along now lest it get lost. When I get around to it I’ll post an update that summarizes what RNNPBs are and precisely how they interface symbol computation with lower-level systems. I may attend the talk (which is tomorrow).

Read on for an abstract from Jun Tani’s talk.

Compositionality and Grounding:
A Dynamical Systems View through Neuro-Cognitive Robotics Experiments

One of the crucial questions in building cognitive agents is how
compositionality can be achieved in agents’ internal representations or memory
structures, which are naturally grounded in sensory-motor competences. In
conventional approaches, symbol systems allocated in a higher level are
interfaced with lower level sensory-motor systems by employing certain

These approaches, however, are likely to face the symbol grounding problem
because the symbol and sensory-motor systems cannot share the same metric
space. Thus, their interfaces become arbitrary. We have proposed a dynamical
systems approach using the “recurrent neural network with parametric biases”
(RNNPB) as an alternative. The RNNPB can be characterized by its basic schema,
which includes the following: (1) a distributed representation of multiple
schemas in a single network; (2) the ability to both generate and recognize
sensory-motor flow as in a “mirror system”; (3) extension to hierarchical
organization of behavioral schemas with chunking; (4) binding among multiple
modalities. My talk will review some of our robot behavior-learning
experiments, including human imitation learning as well as learning simple
language and behavioral binding.

Those experiments show that certain compositional structures hidden in word
sequences, as well as experienced sensory-motor flow, are embedded in the
self-organized dynamical structures of the RNNPB. Furthermore, it has been
observed that such “symbol systems”, embedded in neuronal dynamical systems,
can be naturally integrated with sensory-motor competences since they share the
same metric space. I will discuss the implications of this study as well as
perspectives on future research.

The references are in our publications in

Jun Tani, Ph.D
Lab Head, Lab. for Behavior and Dynamic Cognition
Brain Science Institute, RIKEN
2-1 Hirosawa, Wako-shi, Saitama, 351-0198 Japan

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s