A common goal of mechanistic interpretability is to decompose the activations of neural networks into features: interpretable properties of the input computed by the model. Sparse autoencoders (SAEs) are a popular method for finding these features in LLMs, and it has been postulated that they can be used to find a \textit{canonical} set of units: a unique and complete list of atomic features. We cast doubt on this belief using two novel techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic. SAE stitching involves inserting or swapping latents from a larger SAE into a smaller one. Latents from the larger SAE can be divided into two categories: \emph{novel latents}, which improve performance when added to the smaller SAE, indicating they capture novel information, and \emph{reconstruction latents}, which can replace corresponding latents in the smaller SAE that have similar behavior. The existence of novel features indicates incompleteness of smaller SAEs. Using meta-SAEs -- SAEs trained on the decoder matrix of another SAE -- we find that latents in SAEs often decompose into combinations of latents from a smaller SAE, showing that larger SAE latents are not atomic. The resulting decompositions are often interpretable; e.g. a latent representing ``Einstein'' decomposes into ``scientist'', ``Germany'', and ``famous person''. Even if SAEs do not find canonical units of analysis, they may still be useful tools. We suggest that future research should either pursue different approaches for identifying such units, or pragmatically choose the SAE size suited to their task. We provide an interactive dashboard to explore meta-SAEs: https://metasaes.streamlit.app/

Multimodal learning, which involves integrating information from various modalities such as text, images, audio, and video, is pivotal for numerous complex tasks like visual question answering, cross-modal retrieval, and caption generation. Traditional approaches rely on modality-specific encoders and late fusion techniques, which can hinder scalability and flexibility when adapting to new tasks or modalities. To address these limitations, we introduce a novel framework that extends the concept of task reformulation beyond natural language processing (NLP) to multimodal learning. We propose to reformulate diverse multimodal tasks into a unified next-frame prediction problem, allowing a single model to handle different modalities without modality-specific components. This method treats all inputs and outputs as sequential frames in a video, enabling seamless integration of modalities and effective knowledge transfer across tasks. Our approach is evaluated on a range of tasks, including text-to-text, image-to-text, video-to-video, video-to-text, and audio-to-text, demonstrating the model's ability to generalize across modalities with minimal adaptation. We show that task reformulation can significantly simplify multimodal model design across various tasks, laying the groundwork for more generalized multimodal foundation models.

Next-frame prediction is a useful and powerful method for modelling and understanding the dynamics of video data. Inspired by the empirical success of causal language modelling and next-token prediction in language modelling, we explore the extent to which next-frame prediction serves as a strong foundational learning strategy (analogous to language modelling) for inducing an understanding of the visual world. In order to quantify the specific visual understanding induced by next-frame prediction, we introduce six diagnostic simulation video datasets derived from fundamental physical laws created by varying physical constants such as gravity and mass. We demonstrate that our models trained only on next-frame prediction are capable of predicting the value of these physical constants (e.g. gravity) without having been trained directly to learn these constants via a regression task. We find that the generative training phase alone induces a model state that can predict physical constants significantly better than that of a random model, improving the loss by a factor of between 1.28 to 6.24. We conclude that next-frame prediction shows great promise as a general learning strategy to induce understanding of the many `laws' that govern the visual domain without the need for explicit labelling.

Semantic textual similartiy (STS) and information retrieval tasks (IR) tasks have been the two major avenues to record the progress of embedding models in the past few years. Under the emerging Retrieval-augmented Generation (RAG) paradigm, we envision the need to evaluate next-level language understanding abilities of embedding models, and take a conscious look at the reasoning abilities stored in them. Addressing this, we pose the question: Can retrievers solve reasoning problems? By transforming reasoning tasks into retrieval tasks, we find that without specifically trained for reasoning-level language understanding, current state-of-the-art retriever models may still be far from being competent for playing the role of assisting LLMs, especially in reasoning-intensive tasks. Moreover, albeit trained to be aware of instructions, instruction-aware IR models are often better off without instructions in inference time for reasoning tasks, posing an overlooked retriever-LLM behavioral gap for the research community to align. However, recent decoder-based embedding models show great promise in narrowing the gap, highlighting the pathway for embedding models to achieve reasoning-level language understanding. We also show that, although current off-the-shelf re-ranker models fail on these tasks, injecting reasoning abilities into them through fine-tuning still appears easier than doing so to bi-encoders, and we are able to achieve state-of-the-art performance across all tasks by fine-tuning a reranking model. We release Reasoning as Retrieval Benchmark (RAR-b), a holistic suite of tasks and settings to evaluate the reasoning abilities stored in retriever models. RAR-b is available at https://github.com/gowitheflow-1998/RAR-b.

Achieving an effective fine-grained appearance variation over 2D facial images, whilst preserving facial identity, is a challenging task due to the high complexity and entanglement of common 2D facial feature encoding spaces. Despite these challenges, such fine-grained control, by way of disentanglement is a crucial enabler for data-driven racial bias mitigation strategies across multiple automated facial analysis tasks, as it allows to analyse, characterise and synthesise human facial diversity. In this paper, we propose a novel GAN framework to enable fine-grained control over individual race-related phenotype attributes of the facial images. Our framework factors the latent (feature) space into elements that correspond to race-related facial phenotype representations, thereby separating phenotype aspects (e.g. skin, hair colour, nose, eye, mouth shapes), which are notoriously difficult to annotate robustly in real-world facial data. Concurrently, we also introduce a high quality augmented, diverse 2D face image dataset drawn from CelebA-HQ for GAN training. Unlike prior work, our framework only relies upon 2D imagery and related parameters to achieve state-of-the-art individual control over race-related phenotype attributes with improved photo-realistic output.

Vanilla language models are long known to have subpar sentence-level representation (Reimers and Gurevych, 2019; Wang et al., 2023), even worse than averaging static word embeddings (Pennington et al., 2014), i.e., sentence representations attained by pooling from sub-word embeddings encoded by language models do not closely reflect the relative semantics of sentences. Encouraged by the remarkable success of visual representation learning facilitated by unsupervised contrastive learning (Chen et al., 2020; He et al., 2020), efforts in NLP are made to leverage unsupervised contrastive learning to recover sentence-level encoding abilities from the models (Fang et al., 2020; Wu et al., 2020; Gao et al., 2021; Meng et al., 2021). However, translating the advancements in visual representation learning to learning sentence-level textual semantics presents unique challenges: a single augmentation (Wu et al., 2020; Meng et al., 2021) might alter the meaning of a sentence, posing problems of the validity of the augmented sentence as a positive pair. Such attempts are primarily bottlenecked by the discreteness of subword units brought by tokenization (Sennrich et al., 2016), impeding the creation of continuous unsupervised semantic pairs that have preserved semantics through small perturbations to inputs. Therefore, the most recognized unsupervised sentence representation learning method in NLP applies two dropout masks to the identical input to attain two representations, as positive pairs in contrastive learning (Gao et al., 2021). We argue that using identical inputs confines the method of Gao et al. (2021) to essentially only a way to improve uniformity (Wang and Isola, 2020) by distancing negative examples that are not identical to an instance itself, lacking

Multi-modal information retrieval (MMIR) is a rapidly evolving field, where significant progress, particularly in image-text pairing, has been made through advanced representation learning and cross-modality alignment research. However, current benchmarks for evaluating MMIR performance in image-text pairing within the scientific domain show a notable gap, where chart and table images described in scholarly language usually do not play a significant role. To bridge this gap, we develop a specialised scientific MMIR (SciMMIR) benchmark by leveraging open-access paper collections to extract data relevant to the scientific domain. This benchmark comprises 530K meticulously curated image-text pairs, extracted from figures and tables with detailed captions in scientific documents. We further annotate the image-text pairs with two-level subset-subcategory hierarchy annotations to facilitate a more comprehensive evaluation of the baselines. We conducted zero-shot and fine-tuning evaluations on prominent multi-modal image-captioning and visual language models, such as CLIP and BLIP. Our analysis offers critical insights for MMIR in the scientific domain, including the impact of pre-training and fine-tuning settings and the influence of the visual and textual encoders. All our data and checkpoints are publicly available at https://github.com/Wusiwei0410/SciMMIR.

In recent years, contrastive learning (CL) has been extensively utilized to recover sentence and document-level encoding capability from pre-trained language models. In this work, we question the length generalizability of CL-based models, i.e., their vulnerability towards length-induced semantic shift. We verify not only that length vulnerability is a significant yet overlooked research gap, but we can devise unsupervised CL methods solely depending on the semantic signal provided by document length. We first derive the theoretical foundations underlying length attacks, showing that elongating a document would intensify the high intra-document similarity that is already brought by CL. Moreover, we found that isotropy promised by CL is highly dependent on the length range of text exposed in training. Inspired by these findings, we introduce a simple yet universal document representation learning framework, LA(SER)$^{3}$: length-agnostic self-reference for semantically robust sentence representation learning, achieving state-of-the-art unsupervised performance on the standard information retrieval benchmark.

Audio classification plays a crucial role in speech and sound processing tasks with a wide range of applications. There still remains a challenge of striking the right balance between fitting the model to the training data (avoiding overfitting) and enabling it to generalise well to a new domain. Leveraging the transferability of contrastive learning, we introduce Audio Contrastive-based Fine-tuning (AudioConFit), an efficient approach characterised by robust generalisability. Empirical experiments on a variety of audio classification tasks demonstrate the effectiveness and robustness of our approach, which achieves state-of-the-art results in various settings.

This paper explores deep latent variable models for semi-supervised paraphrase generation, where the missing target pair for unlabelled data is modelled as a latent paraphrase sequence. We present a novel unsupervised model named variational sequence auto-encoding reconstruction (VSAR), which performs latent sequence inference given an observed text. To leverage information from text pairs, we additionally introduce a novel supervised model we call dual directional learning (DDL), which is designed to integrate with our proposed VSAR model. Combining VSAR with DDL (DDL+VSAR) enables us to conduct semi-supervised learning. Still, the combined model suffers from a cold-start problem. To further combat this issue, we propose an improved weight initialisation solution, leading to a novel two-stage training scheme we call knowledge-reinforced-learning (KRL). Our empirical evaluations suggest that the combined model yields competitive performance against the state-of-the-art supervised baselines on complete data. Furthermore, in scenarios where only a fraction of the labelled pairs are available, our combined model consistently outperforms the strong supervised model baseline (DDL) by a significant margin (p <.05; Wilcoxon test). Our code is publicly available at "https://github.com/jialin-yu/latent-sequence-paraphrase".