Peer Review #3 of "Learning to perform role-filler binding with schematic knowledge (v0.2)"

Catherine Chen, Qihong Lu, Andre Beukers, Christopher Baldassano, Kenneth A. Norman, Jimmy Ba, Geoffrey Hinton, Volodymyr Mnih, Joel Leibo, Catalin Ionescu, Frederic Charles, Bartlett, H Gordon, John Bower, Terrence Black, Turner, Ronald Brachman, James Schmolze, Nai Ding, Lucia Melloni, Xing Tian, David Poeppel, A Leonidas, John Doumas, Hummel, Sandhofer, Jeffrey Elman, Ken Mcrae, Allyson Ettinger, Ahmed Elgohary, Philip Resnik, Jerry Fodor, Zenon Pylyshyn, Stefan Frank, Morten Christiansen, Nicholas Franklin, Charan Ranganath, Jeffrey Zacks, Samuel Gershman, A Graves, A Mohamed, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-Barwiska, Sergio Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, Adri Puigdomnech Badia, Karl Hermann, Yori Zwols, Georg Ostrovski, Adam Cain, Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, Demis Hassabis, Olivia Guest, Bradley Love, Xavier Hinaut, Peter Ford, Dominey, S Hochreiter, J Schmidhuber, J Keith, John Holyoak, E John, Biederman, Keith Holyoak, Collin Green, Leonidas Doumas, Derek Devnich, Aniket Kittur, Donald Kalar, Dieuwke Hupkes, Sara Veldhoen, Willem Zuidema, Trenton Kriete, David Noelle, Jonathan Cohen, Randall O'reilly, Yair Lakretz, German Kruszewski, Theo Desbordes, Stanislas Dehaene, Marco Baroni, Andrea MartinOpenalex Unknown 2021 · DOI · OpenAlex

Introduction

Human intelligence is characterized by the ability to learn structural relationships from specific experiences and organize them into mental frameworks known as "schemata." A critical component of generalizing these frameworks to new situations is role-filler binding—the capacity to associate specific "fillers" (e.g., "Alice") with abstract "roles" (e.g., "customer"). For example, in the sentence "Alice ordered a tea from Bob," we instantly identify the bindings "Alice:customer," "tea:drink," and "Bob:barista." This allows us to understand the event even if the specific names are entirely new.

The Challenge of Systematicity

Traditional neural networks, such as standard Recurrent Neural Networks (RNNs) or Long Short-Term Memory (LSTM) units, often struggle with this type of systematic reasoning. These models tend to conflate the identity of a filler with its role based on statistical correlations seen during training. If a model has only encountered "Alice" as a customer, it may fail to recognize her in the role of a barista. Achieving human-like flexibility requires models that can decouple roles from fillers and perform bindings on-the-fly for arbitrary, novel inputs.

Methodology

The researchers evaluated four RNN architectures to test their role-filler binding capabilities:

  1. Standard RNN: A basic sequential baseline.
  2. LSTM: A standard architecture for long-range dependencies.
  3. Fast Weights: A network utilizing a rapidly changing associative memory matrix.
  4. Differentiable Neural Computer (DNC): A model featuring a dedicated external memory buffer with read/write mechanisms.

Datasets and the Story Graph

To train and test these models, the authors developed a custom Story Graph, a state-machine designed to generate thousands of unique narratives resembling coffee shop scenarios.

The Story Graph used to generate synthetic narratives.

The models were subjected to different training regimes:

  • Limited vs. Unlimited Fillers: Models were trained either on a small pool of names or on a dataset where every story featured unique, randomly generated vectors for fillers. The latter forces the model to learn the underlying structure rather than memorizing specific pairings.
  • Correlation Violation Sets: A test set where fillers occupied roles they never held during training (e.g., a name previously only seen as a "customer" appearing as a "barista").

Key Results and Insights

The study revealed that architectural choice and training diversity are paramount for learning abstraction:

  • Memory Models Excel: While standard RNNs and LSTMs failed almost entirely (0% accuracy) on novel fillers, the DNC and Fast Weights architectures achieved near-perfect accuracy when trained with sufficient filler variety.
  • Robustness to Contradiction: Memory-augmented models remained accurate even when tested on "correlation violations," whereas standard models were biased by their training history.
  • Memory as a Scratchpad: Neural decoding analysis showed that the DNC learned to use specific memory "slots" for specific roles (e.g., always writing the "customer" to a specific address), effectively creating a structured internal representation.
  • Dual Knowledge Systems: The Fast Weights model demonstrated the ability to maintain statistical biases (e.g., "Alice is usually the customer") while overriding them when a specific story dictated a different role.

Conclusion

This work demonstrates that networks with external memory can learn to bind roles to arbitrary fillers without explicit labels, provided they are exposed to high filler diversity. While these models currently rely on the sequential structure of input, they represent a significant step toward AI that can reason using abstract schematic knowledge. Future research will explore more complex, overlapping schemata and the potential parallels between these "memory slots" and the human hippocampus.

Summary updated

Abstract

Through specific experiences, humans learn the relationships that underlie the structure of events in the world.Schema theory suggests that we organize this information in mental frameworks called "schemata," which represent our knowledge of the structure of the world.Generalizing knowledge of structural relationships to new situations requires rolefiller binding, the ability to associate specific "fillers" with abstract "roles."For instance, when we hear the sentence "Alice ordered a tea from Bob," the role-filler bindings "Alice:customer," "tea:drink," and "Bob:barista" allow us to understand and make inferences about the sentence.We can perform these bindings for arbitrary fillers -we understand this sentence even if we have never heard the names "Alice," "tea," or "Bob" before.In this work, we define a model as capable of performing role-filler binding if it can recall arbitrary fillers corresponding to a specified role, even when these pairings violate correlations seen during training.Previous work found that models can learn this ability when explicitly told what the roles and fillers are, or when given fillers seen during training.We show that networks with external memory learn to bind roles to arbitrary fillers, without explicitly labeled role-filler pairs.We further show that they can perform these bindings on role-filler pairs that violate correlations seen during training, while retaining knowledge of training correlations.We apply analyses inspired by neural decoding to interpret what the networks have learned.
BibTeX
@inproceedings{chen2021peer,
  author = {Catherine Chen and Qihong Lu and Andre Beukers and Christopher Baldassano and Kenneth A. Norman and Jimmy Ba and Geoffrey Hinton and Volodymyr Mnih and Joel Leibo and Catalin Ionescu and Frederic Charles and Bartlett and H Gordon and John Bower and Terrence Black and Turner and Ronald Brachman and James Schmolze and Nai Ding and Lucia Melloni and Xing Tian and David Poeppel and A Leonidas and John Doumas and Hummel and A Leonidas and John Doumas and Catherine Hummel and Sandhofer and Jeffrey Elman and Ken Mcrae and Allyson Ettinger and Ahmed Elgohary and Philip Resnik and Jerry Fodor and Zenon Pylyshyn and Stefan Frank and Morten Christiansen and Nicholas Franklin and Kenneth Norman and Charan Ranganath and Jeffrey Zacks and Samuel Gershman and A Graves and A Mohamed and G Hinton and Alex Graves and Greg Wayne and Malcolm Reynolds and Tim Harley and Ivo Danihelka and Agnieszka Grabska-Barwiska and Sergio Colmenarejo and Edward Grefenstette and Tiago Ramalho and John Agapiou and Adri Puigdomnech Badia and Karl Hermann and Yori Zwols and Georg Ostrovski and Adam Cain and Helen King and Christopher Summerfield and Phil Blunsom and Koray Kavukcuoglu and Demis Hassabis and Olivia Guest and Bradley Love and Xavier Hinaut and Peter Ford and Dominey and S Hochreiter and J Schmidhuber and J Keith and John Holyoak and Hummel and E John and Irving Hummel and Biederman and John Hummel and Keith Holyoak and Collin Green and Leonidas Doumas and Derek Devnich and Aniket Kittur and Donald Kalar and Dieuwke Hupkes and Sara Veldhoen and Willem Zuidema and Trenton Kriete and David Noelle and Jonathan Cohen and Randall O'reilly and Yair Lakretz and German Kruszewski and Theo Desbordes and Dieuwke Hupkes and Stanislas Dehaene and Marco Baroni and Andrea Martin},
  title = {{Peer Review \#3 of "Learning to perform role-filler binding with schematic knowledge (v0.2)"}},
  booktitle = {Openalex Unknown},
  year = {2021},
  doi = {10.7287/peerj.11046v0.2/reviews/3},
}

Loading related content...

Figure