Mutual exclusivity as a challenge for deep neural networks

Kanishk Gandhi, Brenden M. LakeNeurIPS 2020 · OpenReview

Introduction

While human children naturally assume that a new word refers to an object they haven't seen before—a phenomenon known as the "mutual exclusivity" bias—deep neural networks (DNNs) do the exact opposite. This paper demonstrates that standard neural architectures possess a strong "anti-mutual exclusivity" bias, preferring to map novel inputs to familiar labels. This tendency significantly hinders their efficiency in lifelong learning and open-world scenarios.

The Mutual Exclusivity Problem

In cognitive science, the Mutual Exclusivity (ME) bias is a well-documented learning mechanism. If a child knows what a "cup" is, is shown a cup alongside a strange new object, and is asked to "show me the dax," they almost always point to the new object. They assume that because the cup already has a name, the new name must belong to the new thing.

This bias allows humans to learn with incredible sample efficiency. In contrast, DNNs typically require thousands of examples to learn new associations. The authors investigate whether "vanilla" neural networks—the standard building blocks of modern AI—possess this ME bias or if their fundamental learning mechanics actually work against it. This is a critical question for lifelong learning, where models must constantly adapt to new concepts without forgetting old ones.

Methodology

The researchers approached the problem through three distinct lenses:

  1. Synthetic One-to-One Mapping: They created a controlled environment where 100 input symbols were mapped to 100 output symbols in a strict one-to-one relationship. They trained models on 90 pairs and tested them on the remaining 10 to see if the models would "guess" that a novel input maps to a novel output.
  2. The ME Score: To quantify this behavior, they defined an "ME Score"—the aggregate probability mass a model assigns to unseen output symbols when presented with a novel input. An ideal learner would have a score of 1.0; a random guesser would score much lower.
  3. Lifelong Learning Simulations: They analyzed real-world datasets in machine translation (IWSLT’14, WMT’14, WMT’15) and image recognition (Omniglot, ImageNet), simulating a "stream" of data where new words or classes appear over time. They also implemented an "oracle" ME bias—manually boosting the probability of novel classes—to measure the potential performance benefits of having such a bias.

Key Insights: The "Anti-ME" Bias

The most striking finding is that vanilla neural networks don't just lack an ME bias; they are actively biased against it. As training progresses, models become increasingly likely to map a brand-new input to a highly familiar output.

Illustration of the Anti-ME bias in standard architectures

This occurs due to the "greedy" nature of softmax activations. In a discriminative model, unused output nodes have an optimal activation of zero during training. Consequently, when a novel input appears, the model defaults to the "path of least resistance"—the labels it has already learned to activate strongly.

By analyzing translation and image datasets, the authors showed that the ME assumption is statistically sound in the real world. In the early stages of learning, a new word in a source sentence is highly likely to correspond to a new word in the target sentence. DNNs simply fail to exploit this regularity.

Key Results

  • Synthetic Failure: While an ideal learner should achieve an ME Score of 1.0, vanilla feedforward and Seq2Seq models consistently scored near 0.01, effectively ignoring the unused output slots.
  • Regularization is Not the Cure: The authors tested various standard "fixes" like weight decay, dropout, and entropy regularization. None of these induced an ME bias. In fact, most models transitioned from a neutral state to a strong anti-ME bias during training.
ME Scores across different regularization techniques
  • Translation Decay: In all six translation scenarios tested, the probability that a new source word mapped to a new target word was high initially but plummeted as the models acquired their anti-ME bias.
  • The Power of the Oracle: In image classification tasks on Omniglot, adding a manual "ME bias" to the model's predictions significantly reduced classification loss. This proves that having an ME bias is objectively beneficial for performance, yet current architectures cannot learn it autonomously.

Limitations and Future Work

The authors acknowledge that the ME bias isn't a universal law. Synonyms (two words for one object) and hierarchical categories (a "dog" is also a "mammal") require overriding the bias. A perfect model shouldn't just have ME; it should know when to use it.

The paper concludes that designing "task-general" neural networks capable of autonomously discovering and applying inductive biases like mutual exclusivity remains an open and urgent challenge. Potential paths forward include meta-learning or developing new loss functions that explicitly encourage the discovery of one-to-one mappings in the absence of contrary evidence.

Summary updated

BibTeX
@inproceedings{gandhi2020mutual,
  author = {Kanishk Gandhi and Brenden M. Lake},
  title = {{Mutual exclusivity as a challenge for deep neural networks}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=ewg9rGTevf},
}

Loading related content...

Figure