Source-linked AI summary
Tracking the World State with Recurrent Entity Networks
Mikael Henaff, Jason Weston, Arthur Szlam, Antoine Bordes, Yann LeCun
TL;DR
The paper asks how models can maintain and update a world-state representation from partial observations while reading sequential data. It introduces EntNet, a fixed-size dynamic memory with entity-oriented gated updates, and reports solving all 20 bAbI tasks, strong performance on a many-supporting-facts task, beyond-horizon generalization, and competitive CBT results. Its main demonstrated limitation is reduced performance in the 1k-sample bAbI setting.
Problem
The central problem is maintaining an estimate of the world state and inferring unobserved information from limited observations over sequential stories.
Method
EntNet uses fixed-size dynamic memory cells with keys, contents, and gated processors that update entity-related representations as text is read.
Results
EntNet solves all 20 bAbI tasks, outperforms LSTMs and Memory Networks on a many-supporting-facts task, generalizes beyond its training horizon, and performs competitively on CBT.
Takeaways & Limitations
The results support EntNet as a model for tracking world state while reading text in a single pass and capturing simple dynamics over long timescales.
Takeaways & Limitations
Performance dropped considerably with only 1k bAbI training samples, leaving sample-efficient reasoning in that setting open.
Abstract
from arXiv · showhide
We introduce a new model, the Recurrent Entity Network (EntNet). It is equipped with a dynamic long-term memory which allows it to maintain and update a representation of the state of the world as it receives new data. For language understanding tasks, it can reason on-the-fly as it reads text, not just when it is required to answer a question or respond as is the case for a Memory Network (Sukhbaatar et al., 2015). Like a Neural Turing Machine or Differentiable Neural Computer (Graves et al., 2014; 2016) it maintains a fixed size memory and can learn to perform location and content-based read and write operations. However, unlike those models it has a simple parallel architecture in which several memory locations can be updated simultaneously. The EntNet sets a new state-of-the-art on the bAbI tasks, and is the first method to solve all the tasks in the 10k training examples setting. We also demonstrate that it can solve a reasoning task which requires a large number of supporting facts, which other methods are not able to solve, and can generalize past its training horizon. It can also be practically used on large scale datasets such as Children's Book Test, where it obtains competitive performance, reading the story in a single pass.
1 INTRODUCTION
The paper frames intelligence as maintaining and evolving a world-state estimate from limited observations so an agent can predict unobserved facts and consequences. It proposes EntNet to update entity-centered representations while reading story text and reports strong results across story-understanding benchmarks.
- Intelligent agents must predict unobserved environmental facts and event consequences from limited percepts and knowledge of the past.
- Entity-centered world representations update properties such as locations, carried objects, and room populations as new information arrives.
- The study evaluates story understanding by presenting textual events followed by questions and answers about the resulting world state.
- Weak supervision may support learning dynamical constraints, update rules, and approximate logical inferences about entities and objects.
- EntNet uses fixed dynamic memory cells with keys, contents, and gated processors that selectively update cells associated with input entities.
- The EntNet solves all 20 bAbI tasks, outperforms LSTMs and Memory Networks on a many-supporting-facts task, generalizes beyond its training horizon, and performs competitively on CBT.
2 MODEL
The model processes sequences through an encoder, dynamic memory, and output layer. Its fixed-size block memories use shared gated updates with content- and location-based addressing, then produce outputs from memory states.
- The architecture has three components: an input encoder, dynamic memory, and output layer, trained through backpropagation through time.
- 2.1 INPUT ENCODER: The encoder maps each input sequence to a fixed-length vector using a learned multiplicative mask followed by summation.
- 2.1 INPUT ENCODER: The same learned mask vectors are reused at every time step, while setting all weights to 1 yields a bag-of-words representation.
- 2.2 DYNAMIC MEMORY: Dynamic memory divides hidden states into blocks, with each block acting as a memory whose weights are partially tied across blocks.
- 2.2 DYNAMIC MEMORY: Each memory combines its existing content with a candidate update controlled by a sigmoid gate and shared trainable transformations.
- 2.2 DYNAMIC MEMORY: The gate uses content matching and key-based location matching, while normalization causes new information to reduce similarity to older information.
- 2.3 OUTPUT MODULE: For output, a query vector produces a softmax over memory blocks, followed by a nonlinear transformation and decoding parameters.
3 MOTIVATING EXAMPLE OF OPERATION
EntNet’s motivating operation treats memory slots as entity representations that are updated as sentences arrive. Location addressing activates named entities directly, while content addressing can reactivate related memories when an entity is implicit.
- The model is intended to reason while ingesting story sentences, before answering questions about the read story.
- Memory slots can represent an entity’s location, possessions, or social relations, with their contents revised as the story changes.
- Tying keys to word embeddings can impose prior knowledge by assigning separate memory slots to named entities.
- For “Mary picked up the ball,” location addressing can activate both Mary’s and the ball’s memories, updating their reciprocal possession information.
- For “Mary went to the garden,” content addressing can activate the ball memory even when “ball” is absent, while Mary’s entry opens through location matching.
- After these updates, questions about the ball or Mary can be answered from relevant memory values without further complex reasoning.
4 RELATED WORK
EntNet differs from recurrent, external-memory, Memory Network, graph, and stack-based models in how it structures and updates memory. It uses separate fixed-size key-value slots with parallel gated processing rather than a controller-centered or unbounded memory arrangement.
- Unlike LSTMs and GRUs, EntNet uses separate memory slots with tied gating and update weights, plus content-based input-to-hidden matching.
- Unlike DNCs and NTMs, EntNet uses a simpler parallel architecture instead of a sophisticated controller that emits interface vectors for external-memory operations.
- Unlike Memory Networks, EntNet updates dynamic memory while reading rather than storing the entire input sequence for later controller attention.
- Compared with related parallel recurrent models, EntNet lacks inter-network communication and uses a distinct gating mechanism.
- Unlike stack-, list-, or queue-based writeable memories, EntNet uses a fixed-size key-value pair array.
5 EXPERIMENTS
Experiments evaluate EntNet on a synthetic world-model task, bAbI question answering, and the Children’s Book Test. The model solves the synthetic task across sequence lengths, all bAbI tasks, and achieves competitive single-pass CBT performance, with some remaining gaps.
- 5.1 SYNTHETIC WORLD MODEL TASK: Synthetic world-model experiments show EntNet solves the task across all tested sequence lengths, while MemN2N and LSTM accuracy degrades as sequences lengthen.The task requires combining up to T −2 supporting facts while maintaining and updating both agents’ locations.
- 5.1 SYNTHETIC WORLD MODEL TASK: EntNet achieves good performance several times past its training horizon when evaluated on sequences longer than those used for training.It was trained on variable-length sequences between 1 and 20, then tested on longer sequences.
- 5.2 BABI TASKS: EntNet solves all 20 bAbI tasks and outperforms the compared models in both number of solved tasks and average error.The comparison includes MemN2N, Dynamic Memory Network, Dynamic Neural Turing Machine, Neural Turing Machine, and Differentiable Neural Computer.
- 5.2 BABI TASKS: The interpretability experiment uses tied key vectors and cosine distances to examine which words each memory slot represents after reading a story.Weight tying forces memory slots to encode information about specific entities, while the cosine-distance score measures slot-word affinity.
- 5.2 BABI TASKS: Key-vector tying usually leaves performance largely unchanged but hurts in a few cases, so it is not used for the main bAbI comparison.The authors report this limitation for the interpretability setup rather than the main Table 2 evaluation.
- 5.3 CHILDREN’S BOOK TEST (CBT): On the Children’s Book Test, the simplified EntNet outperforms all other single-pass models on Named Entities and Common Nouns, but remains behind more sophisticated multi-pass systems.The CBT setup asks models to fill a blank in the 21st sentence using 20 preceding sentences and 10 candidate answers.
6 CONCLUSION
The EntNet tracks world state while reading text, solving all bAbI tasks and capturing simple long-term dynamics, while low-data performance remains an open challenge.
- The EntNet accurately tracks world state in text stories and was the first model to solve all bAbI tasks.
- The model captures simple dynamics over long timescales and performs competitively on a real-world dataset.
- Performance dropped considerably with 1k rather than 10k bAbI training samples, leaving the 1k setting open.
- The architecture was applied only to textual inputs, motivating future work combining its tracking abilities with predictive models.
A TRAINING DETAILS
The experiments use Torch implementations with Gaussian weight initialization, except for specially initialized PReLU slopes and encoder weights.
- All models were implemented using Torch.
- Weights were initialized from a zero-mean Gaussian distribution with standard deviation 0.1.
- PReLU slopes and encoder weights were initialized to 1.
B DETAILS OF WORLD MODEL EXPERIMENTS
The world model experiment places two agents on a 10 × 10 grid and samples legal movement or direction-changing actions before querying their final locations.
- Two agents are initially placed randomly among 100 distinct locations on a 10 × 10 grid.
- At each time step, one randomly chosen agent either changes direction or moves a number of steps ahead.
- The model answers questions about the agents’ final locations after a sequence of movement and direction events.
C ADDITIONAL RESULTS ON BABI TASKS
Additional bAbI experiments examine encoding, data volume, and joint-training variability, showing architectural advantages alongside weaker low-data performance and seed sensitivity differences.
- With bag-of-words inputs, the EntNet outperforms a MemN2N using the same encoding, while tying key vectors to entities hurts some tasks.
- With only 1k training samples, the EntNet performs worse than the MemN2N.
- Under joint training on all tasks, the DNC has better best-seed performance, whereas the EntNet varies less across runs and solves more tasks consistently.
- Table 5 reports error rates for bAbI tasks with bag-of-words inputs and distinguishes the tied key-vector variant.
- Table 6 reports results on bAbI tasks with 1k samples.
- Table 7 reports bAbI results with 10k samples and joint training on all tasks.