Distribution matching methods for language model alignment such as Generation with Distributional Control (GDC) and Distributional Policy Gradient (DPG) have not received the same level of attention in reinforcement learning from human feedback (RLHF) as contrastive methods such as Sequence Likelihood Calibration (SLiC), Direct Preference Optimization (DPO) and its variants. We identify high variance of the gradient estimate as the primary reason for the lack of success of these methods and propose a self-normalized baseline to reduce the variance. We further generalize the target distribution in DPG, GDC and DPO by using Bayes' rule to define the reward-conditioned posterior. The resulting approach, referred to as BRAIn - Bayesian Reward-conditioned Amortized Inference acts as a bridge between distribution matching methods and DPO and significantly outperforms prior art in summarization and Antropic HH tasks.

Recent methods for knowledge grounded dialogs generate responses by incorporating information from an external textual document. These methods do not require the exact document to be known during training and rely on the use of a retrieval system to fetch relevant documents from a large index. The documents used to generate the responses are modeled as latent variables whose prior probabilities need to be estimated. Models such as RAG, marginalize the document probabilities over the documents retrieved from the index to define the log-likelihood loss function which is optimized end-to-end. In this paper, we develop a variational approach to the above technique wherein, we instead maximize the Evidence Lower bound (ELBO). Using a collection of three publicly available open-conversation datasets, we demonstrate how the posterior distribution, which has information from the ground-truth response, allows for a better approximation of the objective function during training. To overcome the challenges associated with sampling over a large knowledge collection, we develop an efficient approach to approximate the ELBO. To the best of our knowledge, we are the first to apply variational training for open-scale unsupervised knowledge grounded dialog systems.

Recently several deep learning based models have been proposed for end-to-end learning of dialogs. While these models can be trained from data without the need for any additional annotations, it is hard to interpret them. On the other hand, there exist traditional state based dialog systems, where the states of the dialog are discrete and hence easy to interpret. However these states need to be handcrafted and annotated in the data. To achieve the best of both worlds, we propose Latent State Tracking Network (LSTN) using which we learn an interpretable model in unsupervised manner. The model defines a discrete latent variable at each turn of the conversation which can take a finite set of values. Since these discrete variables are not present in the training data, we use EM algorithm to train our model in unsupervised manner. In the experiments, we show that LSTN can help achieve interpretability in dialog models without much decrease in performance compared to end-to-end approaches.

Deep convolutional networks have achieved the state-of-the-art for semantic image segmentation tasks. However, training these networks requires access to densely labeled images, which are known to be very expensive to obtain. On the other hand, the web provides an almost unlimited source of images annotated at the image level. How can one utilize this much larger weakly annotated set for tasks that require dense labeling? Prior work often relied on localization cues, such as saliency maps, objectness priors, bounding boxes etc., to address this challenging problem. In this paper, we propose a model that generates auxiliary labels for each image, while simultaneously forcing the output of the CNN to satisfy the mean-field constraints imposed by a conditional random field. We show that one can enforce the CRF constraints by forcing the distribution at each pixel to be close to the distribution of its neighbors. This is in stark contrast with methods that compute a recursive expansion of the mean-field distribution using a recurrent architecture and train the resultant distribution. Instead, the proposed model adds an extra loss term to the output of the CNN, and hence, is faster than recursive implementations. We achieve the state-of-the-art for weakly supervised semantic image segmentation on VOC 2012 dataset, assuming no manually labeled pixel level information is available. Furthermore, the incorporation of conditional random fields in CNN incurs little extra time during training.

In this paper we present the Exemplar Encoder-Decoder network (EED), a novel conversation model that learns to utilize similar examples from training data to generate responses.Similar conversation examples (context-response pairs) from training data are retrieved using a traditional TF-IDF based retrieval model.The retrieved responses are used to create exemplar vectors that are used by the decoder to generate the response.The contribution of each retrieved response is weighed by the similarity of corresponding context with the input context.We present detailed experiments on two large data sets and find that our method outperforms state of the art sequence to sequence generative models on several recently proposed evaluation metrics.We also observe that the responses generated by the proposed EED model are more informative and diverse compared to existing state-of-the-art method.

In recent years, deep discriminative models have achieved extraordinary performance on supervised learning tasks, significantly outperforming their generative counterparts. However, their success relies on the presence of a large amount of labeled data. How can one use the same discriminative models for learning useful features in the absence of labels? We address this question in this paper, by jointly modeling the distribution of data and latent features in a manner that explicitly assigns zero probability to unobserved data. Rather than maximizing the marginal probability of observed data, we maximize the joint probability of the data and the latent features using a two step EM-like procedure. To prevent the model from overfitting to our initial selection of latent features, we use adversarial regularization. Depending on the task, we allow the latent features to be one-hot or real-valued vectors, and define a suitable prior on the features. For instance, one-hot features correspond to class labels, and are directly used for unsupervised and semi-supervised classification task, whereas real-valued feature vectors are fed as input to simple classifiers for auxiliary supervised discrimination tasks. The proposed model, which we dub dicriminative encoder (or DisCoder), is flexible in the type of latent features that it can capture. The proposed model achieves state-of-the-art performance on several challenging tasks. Qualitative visualization of the latent features shows that the features learnt by the DisCoder are indeed meaningful.

In this paper, we address the problem of conditional modality learning, whereby one is interested in generating one modality given the other. While it is straightforward to learn a joint distribution over multiple modalities using a deep multi-modal architecture, we observe that such models are not very effective at conditional generation. Hence, we address the problem by learning conditional distributions between the modalities. We use variational methods for maximizing the corresponding conditional log-likelihood. The resultant deep model, which we refer to as conditional multimodal autoencoder (CMMA), forces the latent representation obtained from a single modality alone to be `close' to the joint representation obtained from multiple modalities. We use the proposed model to generate faces from attributes. We show that the faces generated from attributes using the proposed model are qualitatively and quantitatively more representative of the attributes from which they were generated, than those obtained by other deep generative models. We also propose a secondary task, whereby the existing faces are modified by modifying the corresponding attributes. We observe that the modifications in face introduced by the proposed model are representative of the corresponding modifications in attributes.

The aim of the paper is to provide an exact approach for generating a Poisson process sampled from a hierarchical CRM, without having to instantiate the infinitely many atoms of the random measures. We use completely random measures (CRM) and hierarchical CRM to define a prior for Poisson processes. We derive the marginal distribution of the resultant point process, when the underlying CRM is marginalized out. Using well known properties unique to Poisson processes, we were able to derive an exact approach for instantiating a Poisson process with a hierarchical CRM prior. Furthermore, we derive Gibbs sampling strategies for hierarchical CRM models based on Chinese restaurant franchise sampling scheme. As an example, we present the sum of generalized gamma process (SGGP), and show its application in topic-modelling. We show that one can determine the power-law behaviour of the topics and words in a Bayesian fashion, by defining a prior on the parameters of SGGP.

To achieve acceptable performance for AI tasks, one can either use sophisticated feature extraction methods as the first layer in a two-layered supervised learning model, or learn the features directly using a deep (multi-layered) model. While the first approach is very problem-specific, the second approach has computational overheads in learning multiple layers and fine-tuning of the model. In this paper, we propose an approach called wide learning based on arc-cosine kernels, that learns a single layer of infinite width. We propose exact and inexact learning strategies for wide learning and show that wide learning with single layer outperforms single layer as well as deep architectures of finite width for some benchmark datasets.

In recent years, deep architectures have gained a lot of prominence for learning complex AI tasks because of their capability to incorporate complex variations in data within the model. However, these models often need to be trained for a long time in order to obtain good results. In this paper, we propose a technique, called ‘stretching’, that allows the same models to perform considerably better with very little training. We show that learning can be done tractably, even when the weight matrix is stretched to infinity, for some specific models. We also study tractable algorithms for implementing stretching in deep convolutional architectures in an iterative manner and derive bounds for its convergence. Our experimental results suggest that the proposed stretched deep convolutional networks are capable of achieving good performance for many object recognition tasks. More importantly, for a fixed network architecture, one can achieve much better accuracy using stretching rather than learning the weights using backpropagation.

Maximum entropy approach to classification is very well studied in applied statistics and machine learning and almost all the methods that exists in literature are discriminative in nature. In this paper, we introduce a maximum entropy classification method with feature selection for large dimensional data such as text datasets that is generative in nature. To tackle the curse of dimensionality of large data sets, we employ conditional independence assumption (Naive Bayes) and we perform feature selection simultaneously, by enforcing a 'maximum discrimination' between estimated class conditional densities. For two class problems, in the proposed method, we use Jeffreys (J) divergence to discriminate the class conditional densities. To extend our method to the multi-class case, we propose a completely new approach by considering a multi-distribution divergence: we replace Jeffreys divergence by Jensen-Shannon (JS) divergence to discriminate conditional densities of multiple classes. In order to reduce computational complexity, we employ a modified Jensen-Shannon divergence (JS_GM), based on AM-GM inequality. We show that the resulting divergence is a natural generalization of Jeffreys divergence to a multiple distributions case. As far as the theoretical justifications are concerned we show that when one intends to select the best features in a generative maximum entropy approach, maximum discrimination using J-divergence emerges naturally in binary classification. Performance and comparative study of the proposed algorithms have been demonstrated on large dimensional text and gene expression datasets that show our methods scale up very well with large dimensional datasets.

In maximum entropy method, one chooses a distribution from a set of distributions that maximizes the Shannon entropy for making inference from incomplete information. There are various ways to specify this set of distributions, the important special case being when this set is described by mean-value constraints of some feature functions. In this case, maximum entropy method fixes an exponential distribution depending on the feature functions that have to be chosen a priori. In this paper, we treat the problem of selecting a maximum entropy model given various feature subsets and their moments, as a model selection problem, and present a minimum description length (MDL) formulation to solve this problem. For this, we derive normalized maximum likelihood (NML) code-length for these models. Furthermore, we show that the minimax entropy method is a special case of maximum entropy model selection, where one assumes that complexity of all the models are equal. We extend our approach to discriminative maximum entropy models. We apply our approach to gene selection problem to select the number of moments for each gene for fixing the model.