Graph pooling is a crucial operation for encoding hierarchical structures within graphs. Most existing graph pooling approaches formulate the problem as a node clustering task which effectively captures the graph topology. Conventional methods ask users to specify an appropriate number of clusters as a hyperparameter, then assume that all input graphs share the same number of clusters. In inductive settings where the number of clusters can vary, however, the model should be able to represent this variation in its pooling layers in order to learn suitable clusters. Thus we propose GMPool, a novel differentiable graph pooling architecture that automatically determines the appropriate number of clusters based on the input data. The main intuition involves a grouping matrix defined as a quadratic form of the pooling operator, which induces use of binary classification probabilities of pairwise combinations of nodes. GMPool obtains the pooling operator by first computing the grouping matrix, then decomposing it. Extensive evaluations on molecular property prediction tasks demonstrate that our method outperforms conventional methods.
Since most of music has repetitive structures from motifs to phrases, repeating musical ideas can be a basic operation for music composition. The basic block that we focus on is conceptualized as loops which are essential ingredients of music. Furthermore, meaningful note patterns can be formed in a finite space, so it is sufficient to represent them with combinations of discrete symbols as done in other domains. In this work, we propose symbolic music loop generation via learning discrete representations. We first extract loops from MIDI datasets using a loop detector and then learn an autoregressive model trained by discrete latent codes of the extracted loops. We show that our model outperforms well-known music generative models in terms of both fidelity and diversity, evaluating on random space. Our code and supplementary materials are available at https://github.com/sjhan91/Loop_VQVAE_Official.
We show that standard Transformers without graph-specific modifications can lead to promising results in graph learning both in theory and practice. Given a graph, we simply treat all nodes and edges as independent tokens, augment them with token embeddings, and feed them to a Transformer. With an appropriate choice of token embeddings, we prove that this approach is theoretically at least as expressive as an invariant graph network (2-IGN) composed of equivariant linear layers, which is already more expressive than all message-passing Graph Neural Networks (GNN). When trained on a large-scale graph dataset (PCQM4Mv2), our method coined Tokenized Graph Transformer (TokenGT) achieves significantly better results compared to GNN baselines and competitive results compared to Transformer variants with sophisticated graph-specific inductive bias. Our implementation is available at https://github.com/jw9730/tokengt.
Continual learning (CL) aims to learn from sequentially arriving tasks without forgetting previous tasks. Whereas CL algorithms have tried to achieve higher average test accuracy across all the tasks learned so far, learning continuously useful representations is critical for successful generalization and downstream transfer. To measure representational quality, we re-train only the output layers using a small balanced dataset for all the tasks, evaluating the average accuracy without any biased predictions toward the current task. We also test on several downstream tasks, measuring transfer learning accuracy of the learned representations. By testing our new formalism on ImageNet-100 and ImageNet-1000, we find that using more exemplar memory is the only option to make a meaningful difference in learned representations, and most of the regularization- or distillation-based CL algorithms that use the exemplar memory fail to learn continuously useful representations in class-incremental learning. Surprisingly, unsupervised (or self-supervised) CL with sufficient memory size can achieve comparable performance to the supervised counterparts. Considering non-trivial labeling costs, we claim that finding more efficient unsupervised CL algorithms that minimally use exemplary memory would be the next promising direction for CL research.
Pre-trained large language models have shown successful progress in many language understanding benchmarks. This work explores the capability of these models to predict actionable plans in real-world environments. Given a text instruction, we show that language priors encoded in pre-trained language models allow us to infer fine-grained subgoal sequences. In contrast to recent methods which make strong assumptions about subgoal supervision, our experiments show that language models can infer detailed subgoal sequences from few training sequences without any fine-tuning. We further propose a simple strategy to re-rank language model predictions based on interaction and feedback from the environment. Combined with pre-trained navigation and visual reasoning components, our approach demonstrates competitive performance on subgoal prediction and task completion in the ALFRED benchmark compared to prior methods that assume more subgoal supervision.
We study unsupervised multi-hop reranking for multi-hop QA (MQA) with open-domain questions. Since MQA requires piecing information from multiple documents, the main challenge thus resides in retrieving and reranking chains of passages that support the reasoning process. Our approach relies on LargE models with Prompt-Utilizing reranking Strategy (LEPUS): we construct an instruction-like prompt based on a candidate document path and compute a relevance score of the path as the probability of generating a given question, according to a pre-trained language model. Though unsupervised, LEPUS yields competitive reranking performance against state-of-the-art methods that are trained on thousands of examples. Adding a small number of samples (e.g., $2$), we demonstrate further performance gain using in-context learning. Finally, we show that when integrated with a reader module, LEPUS can obtain competitive multi-hop QA performance, e.g., outperforming fully-supervised QA systems. Code will be released at https://github.com/mukhal/LEPUS
Batch Normalization (BN) is an essential layer for training neural network models in various computer vision tasks. It has been widely used in continual learning scenarios with little discussion, but we find that BN should be carefully applied, particularly for the exemplar memory based class incremental learning (CIL). We first analyze that the empirical mean and variance obtained for normalization in a BN layer become highly biased toward the current task. To tackle its significant problems in training and test phases, we propose Task-Balanced Batch Normalization (TBBN). Given each mini-batch imbalanced between the current and previous tasks, TBBN first reshapes and repeats the batch, calculating near task-balanced mean and variance. Second, we show that when the affine transformation parameters of BN are learned from a reshaped feature map, they become less-biased toward the current task. Based on our extensive CIL experiments with CIFAR-100 and ImageNet-100 datasets, we demonstrate that our TBBN is easily applicable to most of existing exemplar-based CIL algorithms, improving their performance by decreasing the forgetting on the previous tasks.
Across many data domains, co-occurrence statistics about the joint appearance of objects are powerfully informative. By transforming unsupervised learning problems into decompositions of co-occurrence statistics, spectral algorithms provide transparent and efficient algorithms for posterior inference such as latent topic analysis and community detection. As object vocabularies grow, however, it becomes rapidly more expensive to store and run inference algorithms on co-occurrence statistics. Rectifying co-occurrence, the key process to uphold model assumptions, becomes increasingly more vital in the presence of rare terms, but current techniques cannot scale to large vocabularies. We propose novel methods that simultaneously compress and rectify co-occurrence statistics, scaling gracefully with the size of vocabulary and the dimension of latent space. We also present new algorithms learning latent variables from the compressed statistics, and verify that our methods perform comparably to previous approaches on both textual and non-textual data.
Spectral topic modeling algorithms operate on matrices/tensors of word co-occurrence statistics to learn topic-specific word distributions. This approach removes the dependence on the original documents and produces substantial gains in efficiency and provable topic inference, but at a cost: the model can no longer provide information about the topic composition of individual documents. Recently Thresholded Linear Inverse (TLI) is proposed to map the observed words of each document back to its topic composition. However, its linear characteristics limit the inference quality without considering the important prior information over topics. In this paper, we evaluate Simple Probabilistic Inverse (SPI) method and novel Prior-aware Dual Decomposition (PADD) that is capable of learning document-specific topic compositions in parallel. Experiments show that PADD successfully leverages topic correlations as a prior, notably outperforming TLI and learning quality topic compositions comparable to Gibbs sampling on various data.
The anchor words algorithm performs provably efficient topic model inference by finding an approximate convex hull in a high-dimensional word co-occurrence space. However, the existing greedy algorithm often selects poor anchor words, reducing topic quality and interpretability. Rather than finding an approximate convex hull in a high-dimensional space, we propose to find an exact convex hull in a visualizable 2- or 3-dimensional space. Such low-dimensional embeddings both improve topics and clearly show users why the algorithm selects certain words.