On the Power of Saturated Transformers: A View from Circuit Complexity

Transformers have become a standard architecture for many NLP problems. This has motivated theoretically analyzing their capabilities as models of language, in order to understand what makes them successful, and what their potential weaknesses might be. Recent work has shown that transformers with hard attention are quite limited in capacity, and in fact can be simulated by constant-depth circuits. However, hard attention is a restrictive assumption, which may complicate the relevance of these results for practical transformers. In this work, we analyze the circuit complexity of transformers with saturated attention: a generalization of hard attention that more closely captures the attention patterns learnable in practical transformers. We show that saturated transformers transcend the limitations of hard-attention transformers. With some minor assumptions, we prove that the number of bits needed to represent a saturated transformer memory vector is $O(\log n)$, which implies saturated transformers can be simulated by log-depth circuits. Thus, the jump from hard to saturated attention can be understood as increasing the transformer's effective circuit depth by a factor of $O(\log n)$.

Provable Limitations of Acquiring Meaning from Ungrounded Form: What will Future Language Models Understand?

Language models trained on billions of tokens have recently led to unprecedented results on many NLP tasks. This success raises the question of whether, in principle, a system can ever "understand" raw text without access to some form of grounding. We formally investigate the abilities of ungrounded systems to acquire meaning. Our analysis focuses on the role of "assertions": contexts within raw text that provide indirect clues about underlying semantics. We study whether assertions enable a system to emulate representations preserving semantic relations like equivalence. We find that assertions enable semantic emulation if all expressions in the language are referentially transparent. However, if the language uses non-transparent patterns like variable binding, we show that emulation can become an uncomputable problem. Finally, we discuss differences between our formal model and natural language, exploring how our results generalize to a modal setting and other semantic relations. Together, our results suggest that assertions in code or language do not provide sufficient signal to fully emulate semantic representations. We formalize ways in which ungrounded language models appear to be fundamentally limited in their ability to "understand".

Random Feature Attention

Transformers are state-of-the-art models for a variety of sequence modeling tasks. At their core is an attention function which models pairwise interactions between the inputs at every timestep. While attention is powerful, it does not scale efficiently to long sequences due to its quadratic time and space complexity in the sequence length. We propose RFA, a linear time and space attention that uses random feature methods to approximate the softmax function, and explore its application in transformers. RFA can be used as a drop-in replacement for conventional softmax attention and offers a straightforward way of learning with recency bias through an optional gating mechanism. Experiments on language modeling and machine translation demonstrate that RFA achieves similar or better performance compared to strong transformer baselines. In the machine translation experiment, RFA decodes twice as fast as a vanilla transformer. Compared to existing efficient transformer variants, RFA is competitive in terms of both accuracy and efficiency on three long text classification datasets. Our analysis shows that RFA's efficiency gains are especially notable on long sequences, suggesting that RFA will be particularly useful in tasks that require working with large inputs, fast decoding speed, or low memory footprints.

Automatic Generation of Contrast Sets from Scene Graphs: Probing the Compositional Consistency of GQA

Recent works have shown that supervised models often exploit data artifacts to achieve good test scores while their performance severely degrades on samples outside their training distribution. Contrast sets (Gardneret al., 2020) quantify this phenomenon by perturbing test samples in a minimal way such that the output label is modified. While most contrast sets were created manually, requiring intensive annotation effort, we present a novel method which leverages rich semantic input representation to automatically generate contrast sets for the visual question answering task. Our method computes the answer of perturbed questions, thus vastly reducing annotation cost and enabling thorough evaluation of models' performance on various semantic aspects (e.g., spatial or relational reasoning). We demonstrate the effectiveness of our approach on the GQA dataset and its semantic scene graph image representation. We find that, despite GQA's compositionality and carefully balanced label distribution, two high-performing models drop 13-17% in accuracy compared to the original test set. Finally, we show that our automatic perturbation can be applied to the training set to mitigate the degradation in performance, opening the door to more robust models.

Parameter Norm Growth During Training of Transformers

The capacity of neural networks like the widely adopted transformer is known to be very high. Evidence is emerging that they learn successfully due to inductive bias in the training routine, typically some variant of gradient descent (GD). To better understand this bias, we study the tendency of transformer parameters to grow in magnitude during training. We find, both theoretically and empirically, that, in certain contexts, GD increases the parameter $L_2$ norm up to a threshold that itself increases with training-set accuracy. This means increasing training accuracy over time enables the norm to increase. Empirically, we show that the norm grows continuously over pretraining for T5 (Raffel et al., 2019). We show that pretrained T5 approximates a semi-discretized network with saturated activation functions. Such "saturated" networks are known to have a reduced capacity compared to the original network family that can be described in automata-theoretic terms. This suggests saturation is a new characterization of an inductive bias implicit in GD that is of particular interest for NLP. While our experiments focus on transformers, our theoretical analysis extends to other architectures with similar formal properties, such as feedforward ReLU networks.

Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics

Large datasets have become commonplace in NLP research. However, the increased emphasis on data quantity has made it challenging to assess the quality of data. We introduce Data Maps---a model-based tool to characterize and diagnose datasets. We leverage a largely ignored source of information: the behavior of the model on individual instances during training (training dynamics) for building data maps. This yields two intuitive measures for each example---the model's confidence in the true class, and the variability of this confidence across epochs---obtained in a single run of training. Experiments across four datasets show that these model-dependent measures reveal three distinct regions in the data map, each with pronounced characteristics. First, our data maps show the presence of "ambiguous" regions with respect to the model, which contribute the most towards out-of-distribution generalization. Second, the most populous regions in the data are "easy to learn" for the model, and play an important role in model optimization. Finally, data maps uncover a region with instances that the model finds "hard to learn"; these often correspond to labeling errors. Our results indicate that a shift in focus from quantity to quality of data could lead to robust models and improved out-of-distribution generalization.

Extracting a Knowledge Base of Mechanisms from COVID-19 Papers

The urgency of mitigating COVID-19 has spawned a large and diverse body of scientific literature that is challenging for researchers to navigate. This explosion of information has stimulated interest in automated tools to help identify useful knowledge. We have pursued the use of methods for extracting diverse forms of mechanism relations from the natural language of scientific papers. We seek to identify concepts in COVID-19 and related literature which represent activities, functions, associations and causal relations, ranging from cellular processes to economic impacts. We formulate a broad, coarse-grained schema targeting mechanism relations between open, free-form entities. Our approach strikes a balance between expressivity and breadth that supports generalization across diverse concepts. We curate a dataset of scientific papers annotated according to our novel schema. Using an information extraction model trained on this new corpus, we construct a knowledge base (KB) of 2M mechanism relations, which we make publicly available. Our model is able to extract relations at an F1 at least twice that of baselines such as open IE or related scientific IE systems. We conduct experiments examining the ability of our system to retrieve relevant information on viral mechanisms of action, and on applications of AI to COVID-19 research. In both cases, our system identifies relevant information from our automatically-constructed knowledge base with high precision.

A Mixture of $h-1$ Heads is Better than $h$ Heads

Multi-head attentive neural architectures have achieved state-of-the-art results on a variety of natural language processing tasks. Evidence has shown that they are overparameterized; attention heads can be pruned without significant performance loss. In this work, we instead "reallocate" them -- the model learns to activate different heads on different inputs. Drawing connections between multi-head attention and mixture of experts, we propose the mixture of attentive experts model (MAE). MAE is trained using a block coordinate descent algorithm that alternates between updating (1) the responsibilities of the experts and (2) their parameters. Experiments on machine translation and language modeling show that MAE outperforms strong baselines on both tasks. Particularly, on the WMT14 English to German translation dataset, MAE improves over "transformer-base" by 0.8 BLEU, with a comparable number of parameters. Our analysis shows that our model learns to specialize different experts to different inputs.

The Right Tool for the Job: Matching Model and Instance Complexities

As NLP models become larger, executing a trained model requires significant computational resources incurring monetary and environmental costs. To better respect a given inference budget, we propose a modification to contextual representation fine-tuning which, during inference, allows for an early (and fast) "exit" from neural network calculations for simple instances, and late (and accurate) exit for hard instances. To achieve this, we add classifiers to different layers of BERT and use their calibrated confidence scores to make early exit decisions. We test our proposed modification on five different datasets in two tasks: three text classification datasets and two natural language inference benchmarks. Our method presents a favorable speed/accuracy tradeoff in almost all cases, producing models which are up to five times faster than the state of the art, while preserving their accuracy. Our method also requires almost no additional training resources (in either time or parameters) compared to the baseline BERT model. Finally, our method alleviates the need for costly retraining of multiple models at different levels of efficiency; we allow users to control the inference speed/accuracy tradeoff using a single trained model, by setting a single variable at inference time. We publicly release our code.

A Formal Hierarchy of RNN Architectures

We develop a formal hierarchy of the expressive capacity of RNN architectures. The hierarchy is based on two formal properties: space complexity, which measures the RNN's memory, and rational recurrence, defined as whether the recurrent update can be described by a weighted finite-state machine. We place several RNN variants within this hierarchy. For example, we prove the LSTM is not rational, which formally separates it from the related QRNN (Bradbury et al., 2016). We also show how these models' expressive capacity is expanded by stacking multiple layers or composing them with different pooling functions. Our results build on the theory of "saturated" RNNs (Merrill, 2019). While formally extending these findings to unsaturated RNNs is left to future work, we hypothesize that the practical learnable capacity of unsaturated RNNs obeys a similar hierarchy. Experimental findings from training unsaturated networks on formal languages support this conjecture.