Additional Feedback: Further comments: - The definition of hard and easy examples is limited to their respective confidence scores or losses. Although previous work has similar definitions, confidence or loss are not always good indicators of true easiness or hardness of samples, e.g. they could be erroneous at early iterations. The paper lacks an experiment that illustrates the validity of the above definition. These are probably hard or noisy examples that were mistreated as easy examples by the model? These are probably a mixture of easy, hard, and noisy examples with low confidence across the loss spectrum that were mistreated as hard examples by the model.

Code-LLMs, LLMs pre-trained on large code corpora, have shown great progress in learning rich representations of the structure and syntax of code, successfully using it to generate or classify code fragments. At the same time, understanding if they are able to do so because they capture code semantics, and how well, is still an open question. In this paper, we tackle this problem by introducing SeqCoBench, a benchmark for systematically assessing how Code-LLMs can capture code functional equivalence. SeqCoBench contains over 20 code transformations that either preserve or alter the semantics of Python programs. We conduct extensive evaluations in different settings, including zero-shot and parameter-efficient finetuning methods on state-of-the-art (Code-)LLMs to see if they can discern semantically equivalent or different pairs of programs in SeqCoBench. We find that the performance gap between these LLMs and classical match-based retrieval scores is minimal, with both approaches showing a concerning lack of depth in understanding code semantics.

There is an increase in the proliferation of online hate commensurate with the rise in the usage of social media. In response, there is also a significant advancement in the creation of automated tools aimed at identifying harmful text content using approaches grounded in Natural Language Processing and Deep Learning. Although it is known that training Deep Learning models require a substantial amount of annotated data, recent line of work suggests that models trained on specific subsets of the data still retain performance comparable to the model that was trained on the full dataset. In this work, we show how we can leverage influence scores to estimate the importance of a data point while training a model and designing a pruning strategy applied to the case of sexism detection. We evaluate the model performance trained on data pruned with different pruning strategies on three out-of-domain datasets and find, that in accordance with other work a large fraction of instances can be removed without significant performance drop. However, we also discover that the strategies for pruning data, previously successful in Natural Language Inference tasks, do not readily apply to the detection of harmful content and instead amplify the already prevalent class imbalance even more, leading in the worst-case to a complete absence of the hateful class.

Eliciting Latent Knowledge (ELK) aims to find patterns in a capable neural network's activations which robustly track the true state of the world, even when the network's overt output is false or misleading. To further ELK research, we introduce 12 datasets and a corresponding suite of "quirky" language models that are LoRA finetuned to make systematic errors when answering questions if and only if the keyword "Bob" is present in the prompt. We demonstrate that simple probing methods can elicit the model's latent knowledge of the correct answer in these contexts, even for problems harder than those the probe was trained on. This is enabled by context-independent knowledge representations located in middle layer activations. We also find that a mechanistic anomaly detection approach can flag untruthful behavior with 94% AUROC. Our results show promise for eliciting reliable knowledge from capable but untrusted models, and facilitates future research empirically investigating ELK methods.

Among attempts at giving a theoretical account of the success of deep neural networks, a recent line of work has identified a so-called'lazy' training regime in which the network can be well approximated by its linearization around initialization. Here we investigate the comparative effect of the lazy (linear) and feature learning (non-linear) regimes on subgroups of examples based on their difficulty. Specifically, we show that easier examples are given more weight in feature learning mode, resulting in faster training compared to more difficult ones. In other words, the non-linear dynamics tends to sequentialize the learning of examples of increasing difficulty. We illustrate this phenomenon across different ways to quantify example difficulty, including c-score, label noise, and in the presence of easy-to-learn spurious correlations. Our results reveal a new understanding of how deep networks prioritize resources across example difficulty.

Recent work on explainable NLP has shown that few-shot prompting can enable large pretrained language models (LLMs) to generate grammatical and factual natural language explanations for data labels. In this work, we study the connection between explainability and sample hardness by investigating the following research question - "Are LLMs and humans equally good at explaining data labels for both easy and hard samples?" We answer this question by first collecting human-written explanations in the form of generalizable commonsense rules on the task of Winograd Schema Challenge (Winogrande dataset). We compare these explanations with those generated by GPT-3 while varying the hardness of the test samples as well as the in-context samples. We observe that (1) GPT-3 explanations are as grammatical as human explanations regardless of the hardness of the test samples, (2) for easy examples, GPT-3 generates highly supportive explanations but human explanations are more generalizable, and (3) for hard examples, human explanations are significantly better than GPT-3 explanations both in terms of label-supportiveness and generalizability judgements. We also find that hardness of the in-context examples impacts the quality of GPT-3 explanations. Finally, we show that the supportiveness and generalizability aspects of human explanations are also impacted by sample hardness, although by a much smaller margin than models. Supporting code and data are available at https://github.com/swarnaHub/ExplanationHardness

It was shown that adversarial examples improve object recognition. But what about their opposite side, easy examples? Easy examples are samples that the machine learning model classifies correctly with high confidence. In our paper, we are making the first step toward exploring the potential benefits of using easy examples in the training procedure of neural networks. We propose to use an auxiliary batch normalization for easy examples for the standard and robust accuracy improvement.

Random forest is widely exploited as an ensemble learning method. In many practical applications, however, there is still a significant challenge to learn from imbalanced data. To alleviate this limitation, we propose a deep dynamic boosted forest (DDBF), a novel ensemble algorithm that incorporates the notion of hard example mining into random forest. Specifically, we propose to measure the quality of each leaf node of every decision tree in the random forest to determine hard examples. By iteratively training and then removing easy examples from training data, we evolve the random forest to focus on hard examples dynamically so as to balance the proportion of samples and learn decision boundaries better. Data can be cascaded through these random forests learned in each iteration in sequence to generate more accurate predictions. Our DDBF outperforms random forest on 5 UCI datasets, MNIST and SATIMAGE, and achieved state-of-the-art results compared to other deep models. Moreover, we show that DDBF is also a new way of sampling and can be very useful and efficient when learning from imbalanced data.