We present Banyan, a model that efficiently learns semantic representations by leveraging explicit hierarchical structure. While transformers excel at scale, they struggle in low-resource settings. Conversely recent structured models have shown promise as efficient learners, but lack performance. Banyan bridges this gap with two key innovations: an entangled hierarchical tree structure and diagonalized message passing, enabling it to outperform larger transformer models with just 14 non-embedding parameters. It excels in low-resource settings, offering a viable alternative for under-represented languages and highlighting its potential for efficient, interpretable NLP in resource-constrained environments.

DreamCoder is an inductive program synthesis system that, whilst solving problems, learns to simplify search in an iterative wake-sleep procedure. The cost of search is amortized by training a neural search policy, reducing search breadth and effectively "compiling" useful information to compose program solutions across tasks. Additionally, a library of program components is learnt to compress and express discovered solutions in fewer components, reducing search depth. We present a novel approach for library learning that directly leverages the neural search policy, effectively "decompiling" its amortized knowledge to extract relevant program components. This provides stronger amortized inference: the amortized knowledge learnt to reduce search breadth is now also used to reduce search depth. We integrate our approach with DreamCoder and demonstrate faster domain proficiency with improved generalization on a range of domains, particularly when fewer example solutions are available.

Conditional neural processes (CNPs) are a flexible and efficient family of models that learn to learn a stochastic process from data. They have seen particular application in contextual image completion - observing pixel values at some locations to predict a distribution over values at other unobserved locations. However, the choice of pixels in learning CNPs is typically either random or derived from a simple statistical measure (e.g. pixel variance). Here, we turn the problem on its head and ask: which pixels would a CNP like to observe - do they facilitate fitting better CNPs, and do such pixels tell us something meaningful about the underlying image? To this end we develop the Partial Pixel Space Variational Autoencoder (PPS-VAE), an amortised variational framework that casts CNP context as latent variables learnt simultaneously with the CNP. We evaluate PPS-VAE over a number of tasks across different visual data, and find that not only can it facilitate better-fit CNPs, but also that the spatial arrangement and values meaningfully characterise image information - evaluated through the lens of classification on both within and out-of-data distributions. Our model additionally allows for dynamic adaption of context-set size and the ability to scale-up to larger images, providing a promising avenue to explore learning meaningful and effective visual representations.

Fourier features based positional encoding (PE) is commonly used in machine learning tasks that involve learning high-frequency features from low-dimensional inputs, such as 3D view synthesis and time series regression with neural tangent kernels. Despite their effectiveness, existing PEs require manual, empirical adjustment of crucial hyperparameters, specifically the Fourier features, tailored to each unique task. Further, PEs face challenges in efficiently learning high-frequency functions, particularly in tasks with limited data. In this paper, we introduce sinusoidal PE (SPE), designed to efficiently learn adaptive frequency features closely aligned with the true underlying function. Our experiments demonstrate that SPE, without hyperparameter tuning, consistently achieves enhanced fidelity and faster training across various tasks, including 3D view synthesis, Text-to-Speech generation, and 1D regression. SPE is implemented as a direct replacement for existing PEs. Its plug-and-play nature lets numerous tasks easily adopt and benefit from SPE.

This work presents StrAE: a Structured Autoencoder framework that through strict adherence to explicit structure, and use of a novel contrastive objective over tree-structured representations, enables effective learning of multi-level representations.Through comparison over different forms of structure, we verify that our results are directly attributable to the informativeness of the structure provided as input, and show that this is not the case for existing tree models.We then further extend StrAE to allow the model to define its own compositions using a simple localised-merge algorithm.This variant, called Self-StrAE, outperforms baselines that don't involve explicit hierarchical compositions, and is comparable to models given informative structure (e.g.constituency parses).Our experiments are conducted in a data-constrained (≈10M tokens) setting to help tease apart the contribution of the inductive bias to effective learning.However, we find that this framework can be robust to scale, and when extended to a much larger dataset (≈100M tokens), our 430 parameter model performs comparably to a 6-layer RoBERTa many orders of magnitude larger in size.Our findings support the utility of incorporating explicit composition as an inductive bias for effective representation learning.

We propose ADIOS, a masked image model (MIM) framework for self-supervised learning, which simultaneously learns a masking function and an image encoder using an adversarial objective. The image encoder is trained to minimise the distance between representations of the original and that of a masked image. The masking function, conversely, aims at maximising this distance. ADIOS consistently improves on state-of-the-art self-supervised learning (SSL) methods on a variety of tasks and datasets—including classification on ImageNet100 and STL10, transfer learning on CIFAR10/100, Flowers102 and iNaturalist, as well as robustness evaluated on the backgrounds challenge (Xiao et al., 2021)—while generating semantically meaningful masks. Unlike modern MIM models such as MAE, BEiT and iBOT, ADIOS does not rely on the image-patch tokenisation construction of Vision Transformers, and can be implemented with convolutional backbones. We further demonstrate that the masks learned by ADIOS are more effective in improving representation learning of SSL methods than masking schemes used in popular MIM models.

Learning general-purpose representations from perceptual inputs is a hallmark of human intelligence. For example, people can write out numbers or characters, or even draw doodles, by characterizing these tasks as different instantiations of the same generic underlying process---compositional arrangements of different forms of pen strokes. Crucially, learning to do one task, say writing, implies reasonable competence at another, say drawing, on account of this shared process. We present Drawing out of Distribution (DooD), a neuro-symbolic generative model of stroke-based drawing that can learn such general-purpose representations. In contrast to prior work, DooD operates directly on images, requires no supervision or expensive test-time inference, and performs unsupervised amortized inference with a symbolic stroke model that better enables both interpretability and generalization. We evaluate DooD on its ability to generalize across both data and tasks. We first perform zero-shot transfer from one dataset (e.g. MNIST) to another (e.g. Quickdraw), across five different datasets, and show that DooD clearly outperforms different baselines. An analysis of the learnt representations further highlights the benefits of adopting a symbolic stroke model. We then adopt a subset of the Omniglot challenge tasks, and evaluate its ability to generate new exemplars (both unconditionally and conditionally), and perform one-shot classification, showing that DooD matches the state of the art. Taken together, we demonstrate that DooD does indeed capture general-purpose representations across both data and task, and takes a further step towards building general and robust concept-learning systems.

We present Multitask Soft Option Learning (MSOL), a hierarchical multitask framework based on Planning as Inference. MSOL extends the concept of options, using separate variational posteriors for each task, regularized by a shared prior. This “soft” version of options avoids several instabilities during training in a multitask setting, and provides a natural way to learn both intra-option policies and their terminations. Furthermore, it allows fine-tuning of options for new tasks without forgetting their learned policies, leading to faster training without reducing the expressiveness of the hierarchical policy. We demonstrate empirically that MSOL significantly outperforms both hierarchical and flat transfer-learning baselines.

Abstract Deep generative modelling for human body analysis is an emerging problem with many interesting applications. However, the latent space learned by such approaches is typically not interpretable, resulting in less flexibility. In this work, we present deep generative models for human body analysis in which the body pose and the visual appearance are disentangled. Such a disentanglement allows independent manipulation of pose and appearance, and hence enables applications such as pose-transfer without specific training for such a task. Our proposed models, the Conditional-DGPose and the Semi-DGPose, have different characteristics. In the first, body pose labels are taken as conditioners, from a fully-supervised training set. In the second, our structured semi-supervised approach allows for pose estimation to be performed by the model itself and relaxes the need for labelled data. Therefore, the Semi-DGPose aims for the joint understanding and generation of people in images. It is not only capable of mapping images to interpretable latent representations but also able to map these representations back to the image space. We compare our models with relevant baselines, the ClothNet-Body and the Pose Guided Person Generation networks, demonstrating their merits on the Human3.6M, ChictopiaPlus and DeepFashion benchmarks.

Deep latent-variable models learn representations of high-dimensional data in an unsupervised manner. A number of recent efforts have focused on learning representations that disentangle statistically independent axes of variation by introducing modifications to the standard objective function. These approaches generally assume a simple diagonal Gaussian prior and as a result are not able to reliably disentangle discrete factors of variation. We propose a two-level hierarchical objective to control relative degree of statistical independence between blocks of variables and individual variables within blocks. We derive this objective as a generalization of the evidence lower bound, which allows us to explicitly represent the trade-offs between mutual information between data and representation, KL divergence between representation and prior, and coverage of the support of the empirical data distribution. Experiments on a variety of datasets demonstrate that our objective can not only disentangle discrete variables, but that doing so also improves disentanglement of other variables and, importantly, generalization even to unseen combinations of factors.

We develop a generalisation of disentanglement in variational autoencoders (VAEs)—decomposition of the latent representation—characterising it as the fulfilment of two factors: a) the latent encodings of the data having an appropriate level of overlap, and b) the aggregate encoding of the data conforming to a desired structure, represented through the prior. Decomposition permits disentanglement, i.e. explicit independence between latents, as a special case, but also allows for a much richer class of properties to be imposed on the learnt representation, such as sparsity, clustering, independent subspaces, or even intricate hierarchical dependency relationships. We show that the $\beta$-VAE varies from the standard VAE predominantly in its control of latent overlap and that for the standard choice of an isotropic Gaussian prior, its objective is invariant to rotations of the latent representation. Viewed from the decomposition perspective, breaking this invariance with simple manipulations of the prior can yield better disentanglement with little or no detriment to reconstructions. We further demonstrate how other choices of prior can assist in producing different decompositions and introduce an alternative training objective that allows the control of both decomposition factors in a principled manner.

Learning generative models that span multiple data modalities, such as vision and language, is often motivated by the desire to learn more useful, generalisable representations that faithfully capture common underlying factors between the modalities. In this work, we characterise successful learning of such models as the fulfilment of four criteria: i) implicit latent decomposition into shared and private subspaces, ii) coherent joint generation over all modalities, iii) coherent cross-generation across individual modalities, and iv) improved model learning for individual modalities through multi-modal integration. Here, we propose a mixture-of-experts multi-modal variational autoencoder (MMVAE) for learning of generative models on different sets of modalities, including a challenging image <-> language dataset, and demonstrate its ability to satisfy all four criteria, both qualitatively and quantitatively.

We propose a deep generative model of humans in natural images which keeps 2D pose separated from other latent factors of variation, such as background scene and clothing. In contrast to methods that learn generative models of low-dimensional representations, e.g., segmentation masks and 2D skeletons, our single-stage end-to-end conditional-VAEGAN learns directly on the image space. The flexibility of this approach allows the sampling of people with independent variations of pose and appearance. Moreover, it enables the reconstruction of images conditioned to a given posture, allowing, for instance, pose-transfer from one person to another. We validate our method on the Human3.6M dataset and achieve state-of-the-art results on the ChictopiaPlus benchmark. Our model, named Conditional-DGPose, outperforms the closest related work in the literature. It generates more realistic and accurate images regarding both, body posture and image quality, learning the underlying factors of pose and appearance variation.

We present FLIPDIAL, a generative model for Visual Dialogue that simultaneously plays the role of both participants in a visually-grounded dialogue. Given context in the form of an image and an associated caption summarising the contents of the image, FLIPDIAL learns both to answer questions and put forward questions, capable of generating entire sequences of dialogue (question-answer pairs) which are diverse and relevant to the image. To do this, FLIPDIAL relies on a simple but surprisingly powerful idea: it uses convolutional neural networks (CNNs) to encode entire dialogues directly, implicitly capturing dialogue context, and conditional VAEs to learn the generative model, FLIPDIAL outperforms the state-of-the-art model in the sequential answering task (1VD) on the VisDial dataset by 5 points in Mean Rank using the generated answers. We are the first to extend this paradigm to full two-way visual dialogue (2VD), where our model is capable of generating both questions and answers in sequence based on a visual input, for which we propose a set of novel evaluation measures and metrics.

Inference amortization methods share information across multiple posterior-inference problems, allowing each to be carried out more efficiently. Generally, they require the inversion of the dependency structure in the generative model, as the modeller must learn a mapping from observations to distributions approximating the posterior. Previous approaches have involved inverting the dependency structure in a heuristic way that fails to capture these dependencies correctly, thereby limiting the achievable accuracy of the resulting approximations. We introduce an algorithm for faithfully, and minimally, inverting the graphical model structure of any generative model. Such inverses have two crucial properties: (a) they do not encode any independence assertions that are absent from the model and; (b) they are local maxima for the number of true independencies encoded. We prove the correctness of our approach and empirically show that the resulting minimally faithful inverses lead to better inference amortization than existing heuristic approaches.

Variational autoencoders (VAEs) learn representations of data by jointly training a probabilistic encoder and decoder network. Typically these models encode all features of the data into a single variable. Here we are interested in learning disentangled representations that encode distinct aspects of the data into separate variables. We propose to learn such representations using model architectures that generalise from standard VAEs, employing a general graphical model structure in the encoder and decoder. This allows us to train partially-specified models that make relatively strong assumptions about a subset of interpretable variables and rely on the flexibility of neural networks to learn representations for the remaining variables. We further define a general objective for semi-supervised learning in this model class, which can be approximated using an importance sampling procedure. We evaluate our framework's ability to learn disentangled representations, both by qualitative exploration of its generative capacity, and quantitative evaluation of its discriminative ability on a variety of models and datasets.

We present an approach to searching large video corpora for clips which depict a natural-language query in the form of a sentence. Compositional semantics is used to encode subtle meaning differences lost in other approaches, such as the difference between two sentences which have identical words but entirely different meaning: The person rode the horse versus The horse rode the person. Given a sentential query and a natural-language parser, we produce a score indicating how well a video clip depicts that sentence for each clip in a corpus and return a ranked list of clips. Two fundamental problems are addressed simultaneously: detecting and tracking objects, and recognizing whether those tracks depict the query. Because both tracking and object detection are unreliable, our approach uses the sentential query to focus the tracker on the relevant participants and ensures that the resulting tracks are described by the sentential query. While most earlier work was limited to single-word queries which correspond to either verbs or nouns, we search for complex queries which contain multiple phrases, such as prepositional phrases, and modifiers, such as adverbs. We demonstrate this approach by searching for 2,627 naturally elicited sentential queries in 10 Hollywood movies.

We present an approach to simultaneously reasoning about a video clip and an entire natural-language sentence. The compositional nature of language is exploited to construct models which represent the meanings of entire sentences composed out of the meanings of the words in those sentences mediated by a grammar that encodes the predicate-argument relations. We demonstrate that these models faithfully represent the meanings of sentences and are sensitive to how the roles played by participants (nouns), their characteristics (adjectives), the actions performed (verbs), the manner of such actions (adverbs), and changing spatial relations between participants (prepositions) affect the meaning of a sentence and how it is grounded in video. We exploit this methodology in three ways. In the first, a video clip along with a sentence are taken as input and the participants in the event described by the sentence are highlighted, even when the clip depicts multiple similar simultaneous events. In the second, a video clip is taken as input without a sentence and a sentence is generated that describes an event in that clip. In the third, a corpus of video clips is paired with sentences which describe some of the events in those clips and the meanings of the words in those sentences are learned. We learn these meanings without needing to specify which attribute of the video clips each word in a given sentence refers to. The learned meaning representations are shown to be intelligible to humans.

We present a system that demonstrates how the compositional structure of events, in concert with the compositional structure of language, can interplay with the underlying focusing mechanisms in video action recognition, providing a medium for top-down and bottom-up integration as well as multi-modal integration between vision and language. We show how the roles played by participants (nouns), their characteristics (adjectives), the actions performed (verbs), the manner of such actions (adverbs), and changing spatial relations between participants (prepositions), in the form of whole-sentence descriptions mediated by a grammar, guides the activity-recognition process. Further, the utility and expressiveness of our framework is demonstrated by performing three separate tasks in the domain of multi-activity video: sentence-guided focus of attention, generation of sentential description, and query-based search, simply by leveraging the framework in different manners.

Recognizing human activities in partially observed videos is a challenging problem and has many practical applications. When the unobserved subsequence is at the end of the video, the problem is reduced to activity prediction from unfinished activity streaming, which has been studied by many researchers. However, in the general case, an unobserved subsequence may occur at any time by yielding a temporal gap in the video. In this paper, we propose a new method that can recognize human activities from partially observed videos in the general case. Specifically, we formulate the problem into a probabilistic framework: 1) dividing each activity into multiple ordered temporal segments, 2) using spatiotemporal features of the training video samples in each segment as bases and applying sparse coding (SC) to derive the activity likelihood of the test video sample at each segment, and 3) finally combining the likelihood at each segment to achieve a global posterior for the activities. We further extend the proposed method to include more bases that correspond to a mixture of segments with different temporal lengths (MSSC), which can better represent the activities with large intra-class variations. We evaluate the proposed methods (SC and MSSC) on various real videos. We also evaluate the proposed methods on two special cases: 1) activity prediction where the unobserved subsequence is at the end of the video, and 2) human activity recognition on fully observed videos. Experimental results show that the proposed methods outperform existing state-of-the-art comparison methods.

We present a generative domain of visual objects by analogy to the generative nature of human language. Just as small inventories of phonemes and words combine in a grammatical fashion to yield myriad valid words and utterances, a small inventory of physical parts combine in a grammatical fashion to yield myriad valid assemblies. We apply the notion of a language model from speech recognition to this visual domain to similarly improve the performance of the recognition process over what would be possible by only applying recognizers to the components. Unlike the context-free models for human language, our visual language models are context sensitive and formulated as stochastic constraint-satisfaction problems. And unlike the situation for human language where all components are observable, our methods deal with occlusion, successfully recovering object structure despite unobservable components. We demonstrate our system with an integrated robotic system for disassembling structures that performs whole-scene reconstruction consistent with a language model in the presence of noisy feature detectors.

We present an integrated vision and robotic system that plays, and learns to play, simple physically-instantiated board games that are variants of TIC TAC TOE and HEXA-PAWN. We employ novel custom vision and robotic hardware designed specifically for this learning task. The game rules can be parametrically specified. Two independent computational agents alternate playing the two opponents with the shared vision and robotic hardware, using pre-specified rule sets. A third independent computational agent, sharing the same hardware, learns the game rules solely by observing the physical play, without access to the pre-specified rule set, using inductive logic programming with minimal background knowledge possessed by human children. The vision component of our integrated system reliably detects the position of the board in the image and reconstructs the game state after every move, from a single image. The robotic component reliably moves pieces both between board positions and to and from off-board positions as needed by an arbitrary parametrically-specified legal-move generator. Thus the rules of games learned solely by observing physical play can drive further physical play. We demonstrate our system learning to play six different games.