Focusing More on Conflicts with Mis-Predictions Helps Language Pre-Training

In this work, we propose to improve the effectiveness of language pre-training methods with the help of mis-predictions during pre-training. Neglecting words in the input sentence that have conflicting semantics with mis-predictions is likely to be the reason of generating mis-predictions at pre-training. Therefore, we hypothesis that mis-predictions during pre-training can act as detectors of the ill focuses of the model. If we train the model to focus more on the conflicts with the mis-predictions while focus less on the rest words in the input sentence, the mis-predictions can be more easily corrected and the entire model could be better trained. Towards this end, we introduce Focusing Less on Context of Mis-predictions(McMisP). In McMisP, we record the co-occurrence information between words to detect the conflicting words with mis-predictions in an unsupervised way. Then McMisP uses such information to guide the attention modules when a mis-prediction occurs. Specifically, several attention modules in the Transformer are optimized to focus more on words in the input sentence that have co-occurred rarely with the mis-predictions and vice versa. Results show that McMisP significantly expedites BERT and ELECTRA and improves their performances on downstream tasks.

Taking Notes on the Fly Helps BERT Pre-training

How to make unsupervised language pre-training more efficient and less resource-intensive is an important research direction in NLP. In this paper, we focus on improving the efficiency of language pre-training methods through providing better data utilization. It is well-known that in language data corpus, words follow a heavy-tail distribution. A large proportion of words appear only very few times and the embeddings of rare words are usually poorly optimized. We argue that such embeddings carry inadequate semantic signals. They could make the data utilization inefficient and slow down the pre-training of the entire model. To solve this problem, we propose Taking Notes on the Fly (TNF). TNF takes notes for rare words on the fly during pre-training to help the model understand them when they occur next time. Specifically, TNF maintains a note dictionary and saves a rare word's context information in it as notes when the rare word occurs in a sentence. When the same rare word occurs again in training, TNF employs the note information saved beforehand to enhance the semantics of the current sentence. By doing so, TNF provides a better data utilization since cross-sentence information is employed to cover the inadequate semantics caused by rare words in the sentences. Experimental results show that TNF significantly expedite the BERT pre-training and improve the model's performance on downstream tasks. TNF's training time is $60\%$ less than BERT when reaching the same performance. When trained with same number of iterations, TNF significantly outperforms BERT on most of downstream tasks and the average GLUE score.

On Layer Normalization in the Transformer Architecture

The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normalization between the residual blocks, the expected gradients of the parameters near the output layer are large. Therefore, using a large learning rate on those gradients makes the training unstable. The warm-up stage is practically helpful for avoiding this problem. On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization. This motivates us to remove the warm-up stage for the training of Pre-LN Transformers. We show in our experiments that Pre-LN Transformers without the warm-up stage can reach comparable results with baselines while requiring significantly less training time and hyper-parameter tuning on a wide range of applications.

Distance-Based Learning from Errors for Confidence Calibration

Deep neural networks (DNNs) are poorly-calibrated when trained in conventional ways. To improve confidence calibration of DNNs, we propose a novel training method, distance-based learning from errors (DBLE). DBLE bases its confidence estimation on distances in the representation space. We first adapt prototypical learning for training of a classification model for DBLE. It yields a representation space where the distance from a test sample to its ground-truth class center can calibrate the model performance. At inference, however, these distances are not available due to the lack of ground-truth labels. To circumvent this by approximately inferring the distance for every test sample, we propose to train a confidence model jointly with the classification model by merely learning from mis-classified training samples, which we show to be highly beneficial for effective learning. On multiple datasets and DNN architectures, we demonstrate that DBLE outperforms alternative single-modal confidence calibration approaches. DBLE also achieves comparable performance with computationally-expensive ensemble approaches with lower computational cost and lower number of parameters.

Adaptive Cross-Modal Few-Shot Learning

Metric-based meta-learning techniques have successfully been applied to few-shot classification problems. However, leveraging cross-modal information in a few-shot setting has yet to be explored. When the support from visual information is limited in few-shot image classification, semantic representatins (learned from unsupervised text corpora) can provide strong prior knowledge and context to help learning. Based on this intuition, we design a model that is able to leverage visual and semantic features in the context of few-shot classification. We propose an adaptive mechanism that is able to effectively combine both modalities conditioned on categories. Through a series of experiments, we show that our method boosts the performance of metric-based approaches by effectively exploiting language structure. Using this extra modality, our model bypass current unimodal state-of-the-art methods by a large margin on two important benchmarks: mini-ImageNet and tiered-ImageNet. The improvement in performance is particularly large when the number of shots are small.

PIRM Challenge on Perceptual Image Enhancement on Smartphones: Report

This paper reviews the first challenge on efficient perceptual image enhancement with the focus on deploying deep learning models on smartphones. The challenge consisted of two tracks. In the first one, participants were solving the classical image super-resolution problem with a bicubic downscaling factor of 4. The second track was aimed at real-world photo enhancement, and the goal was to map low-quality photos from the iPhone 3GS device to the same photos captured with a DSLR camera. The target metric used in this challenge combined the runtime, PSNR scores and solutions' perceptual results measured in the user study. To ensure the efficiency of the submitted models, we additionally measured their runtime and memory requirements on Android smartphones. The proposed solutions significantly improved baseline results defining the state-of-the-art for image enhancement on smartphones.

A Walk with SGD

We present novel empirical observations regarding how stochastic gradient descent (SGD) navigates the loss landscape of over-parametrized deep neural networks (DNNs). These observations expose the qualitatively different roles of learning rate and batch-size in DNN optimization and generalization. Specifically we study the DNN loss surface along the trajectory of SGD by interpolating the loss surface between parameters from consecutive \textit{iterations} and tracking various metrics during training. We find that the loss interpolation between parameters before and after each training iteration's update is roughly convex with a minimum (\textit{valley floor}) in between for most of the training. Based on this and other metrics, we deduce that for most of the training update steps, SGD moves in valley like regions of the loss surface by jumping from one valley wall to another at a height above the valley floor. This 'bouncing between walls at a height' mechanism helps SGD traverse larger distance for small batch sizes and large learning rates which we find play qualitatively different roles in the dynamics. While a large learning rate maintains a large height from the valley floor, a small batch size injects noise facilitating exploration. We find this mechanism is crucial for generalization because the valley floor has barriers and this exploration above the valley floor allows SGD to quickly travel far away from the initialization point (without being affected by barriers) and find flatter regions, corresponding to better generalization.

A Sequential Matching Framework for Multi-turn Response Selection in Retrieval-based Chatbots

We study the problem of response selection for multi-turn conversation in retrieval-based chatbots. The task requires matching a response candidate with a conversation context, whose challenges include how to recognize important parts of the context, and how to model the relationships among utterances in the context. Existing matching methods may lose important information in contexts as we can interpret them with a unified framework in which contexts are transformed to fixed-length vectors without any interaction with responses before matching. The analysis motivates us to propose a new matching framework that can sufficiently carry the important information in contexts to matching and model the relationships among utterances at the same time. The new framework, which we call a sequential matching framework (SMF), lets each utterance in a context interacts with a response candidate at the first step and transforms the pair to a matching vector. The matching vectors are then accumulated following the order of the utterances in the context with a recurrent neural network (RNN) which models the relationships among the utterances. The context-response matching is finally calculated with the hidden states of the RNN. Under SMF, we propose a sequential convolutional network and sequential attention network and conduct experiments on two public data sets to test their performance. Experimental results show that both models can significantly outperform the state-of-the-art matching methods. We also show that the models are interpretable with visualizations that provide us insights on how they capture and leverage the important information in contexts for matching.

Sequential Matching Network: A New Architecture for Multi-turn Response Selection in Retrieval-based Chatbots

We study response selection for multi-turn conversation in retrieval-based chatbots. Existing work either concatenates utterances in context or matches a response with a highly abstract context vector finally, which may lose relationships among utterances or important contextual information. We propose a sequential matching network (SMN) to address both problems. SMN first matches a response with each utterance in the context on multiple levels of granularity, and distills important matching information from each pair as a vector with convolution and pooling operations. The vectors are then accumulated in a chronological order through a recurrent neural network (RNN) which models relationships among utterances. The final matching score is calculated with the hidden states of the RNN. An empirical study on two public data sets shows that SMN can significantly outperform state-of-the-art methods for response selection in multi-turn conversation.

Hierarchical Recurrent Attention Network for Response Generation

We study multi-turn response generation in chatbots where a response is generated according to a conversation context. Existing work has modeled the hierarchy of the context, but does not pay enough attention to the fact that words and utterances in the context are differentially important. As a result, they may lose important information in context and generate irrelevant responses. We propose a hierarchical recurrent attention network (HRAN) to model both aspects in a unified framework. In HRAN, a hierarchical attention mechanism attends to important parts within and among utterances with word level attention and utterance level attention respectively. With the word level attention, hidden vectors of a word level encoder are synthesized as utterance vectors and fed to an utterance level encoder to construct hidden representations of the context. The hidden vectors of the context are then processed by the utterance level attention and formed as context vectors for decoding the response. Empirical studies on both automatic evaluation and human judgment show that HRAN can significantly outperform state-of-the-art models for multi-turn response generation.