Saurabh Kumar

Saurabh Kumar, Henrik Marklund, Benjamin Van Roy · stanford

In continual learning, plasticity refers to the ability of an agent to quickly adapt to new information. Neural networks are known to lose plasticity when processing non-stationary data streams. In this paper, we propose L2 Init, a simple approach for maintaining plasticity by incorporating in the loss function L2 regularization toward initial parameters. This is very similar to standard L2 regularization (L2), the only difference being that L2 regularizes toward the origin. L2 Init is simple to implement and requires selecting only a single hyper-parameter. The motivation for this method is the same as that of methods that reset neurons or parameter values. Intuitively, when recent losses are insensitive to particular parameters, these parameters should drift toward their initial values. This prepares parameters to adapt quickly to new tasks. On problems representative of different types of nonstationarity in continual supervised learning, we demonstrate that L2 Init most consistently mitigates plasticity loss compared to previously proposed approaches.

Saurabh Kumar, Henrik Marklund, Ashish Rao et al. · stanford

An agent that efficiently accumulates knowledge to develop increasingly sophisticated skills over a long lifetime could advance the frontier of artificial intelligence capabilities. The design of such agents, which remains a long-standing challenge of artificial intelligence, is addressed by the subject of continual learning. This monograph clarifies and formalizes concepts of continual learning, introducing a framework and set of tools to stimulate further research.

Dilip Arumugam, Saurabh Kumar, Ramki Gummadi et al. · stanford

A default assumption in the design of reinforcement-learning algorithms is that a decision-making agent always explores to learn optimal behavior. In sufficiently complex environments that approach the vastness and scale of the real world, however, attaining optimal performance may in fact be an entirely intractable endeavor and an agent may seldom find itself in a position to complete the requisite exploration for identifying an optimal policy. Recent work has leveraged tools from information theory to design agents that deliberately forgo optimal solutions in favor of sufficiently-satisfying or satisficing solutions, obtained through lossy compression. Notably, such agents may employ fundamentally different exploratory decisions to learn satisficing behaviors more efficiently than optimal ones that are more data intensive. While supported by a rigorous corroborating theory, the underlying algorithm relies on model-based planning, drastically limiting the compatibility of these ideas with function approximation and high-dimensional observations. In this work, we remedy this issue by extending an agent that directly represents uncertainty over the optimal value function allowing it to both bypass the need for model-based planning and to learn satisficing policies. We provide simple yet illustrative experiments that demonstrate how our algorithm enables deep reinforcement-learning agents to achieve satisficing behaviors. In keeping with previous work on this setting for multi-armed bandits, we additionally find that our algorithm is capable of synthesizing optimal behaviors, when feasible, more efficiently than its non-information-theoretic counterpart.

Abhayjeet Singh, Arjun Singh Mehta, Ashish Khuraishi K S et al.

Automatic speech recognition (ASR) performance has improved drastically in recent years, mainly enabled by self-supervised learning (SSL) based acoustic models such as wav2vec2 and large-scale multi-lingual training like Whisper. A huge challenge still exists for low-resource languages where the availability of both audio and text is limited. This is further complicated by the presence of multiple dialects like in Indian languages. However, many Indian languages can be grouped into the same families and share the same script and grammatical structure. This is where a lot of adaptation and fine-tuning techniques can be applied to overcome the low-resource nature of the data by utilising well-resourced similar languages. In such scenarios, it is important to understand the extent to which each modality, like acoustics and text, is important in building a reliable ASR. It could be the case that an abundance of acoustic data in a language reduces the need for large text-only corpora. Or, due to the availability of various pretrained acoustic models, the vice-versa could also be true. In this proposed special session, we encourage the community to explore these ideas with the data in two low-resource Indian languages of Bengali and Bhojpuri. These approaches are not limited to Indian languages, the solutions are potentially applicable to various languages spoken around the world.

Vikas Maurya, Rachit Agarwal, Saurabh Kumar et al.

Due to the importance of Critical Infrastructure (CI) in a nation's economy, they have been lucrative targets for cyber attackers. These critical infrastructures are usually Cyber-Physical Systems (CPS) such as power grids, water, and sewage treatment facilities, oil and gas pipelines, etc. In recent times, these systems have suffered from cyber attacks numerous times. Researchers have been developing cyber security solutions for CIs to avoid lasting damages. According to standard frameworks, cyber security based on identification, protection, detection, response, and recovery are at the core of these research. Detection of an ongoing attack that escapes standard protection such as firewall, anti-virus, and host/network intrusion detection has gained importance as such attacks eventually affect the physical dynamics of the system. Therefore, anomaly detection in physical dynamics proves an effective means to implement defense-in-depth. PASAD is one example of anomaly detection in the sensor/actuator data, representing such systems' physical dynamics. We present EPASAD, which improves the detection technique used in PASAD to detect these micro-stealthy attacks, as our experiments show that PASAD's spherical boundary-based detection fails to detect. Our method EPASAD overcomes this by using Ellipsoid boundaries, thereby tightening the boundaries in various dimensions, whereas a spherical boundary treats all dimensions equally. We validate EPASAD using the dataset produced by the TE-process simulator and the C-town datasets. The results show that EPASAD improves PASAD's average recall by 5.8% and 9.5% for the two datasets, respectively.

Ramki Gummadi, Saurabh Kumar, Junfeng Wen et al.

Approaches to policy optimization have been motivated from diverse principles, based on how the parametric model is interpreted (e.g. value versus policy representation) or how the learning objective is formulated, yet they share a common goal of maximizing expected return. To better capture the commonalities and identify key differences between policy optimization methods, we develop a unified perspective that re-expresses the underlying updates in terms of a limited choice of gradient form and scaling function. In particular, we identify a parameterized space of approximate gradient updates for policy optimization that is highly structured, yet covers both classical and recent examples, including PPO. As a result, we obtain novel yet well motivated updates that generalize existing algorithms in a way that can deliver benefits both in terms of convergence speed and final result quality. An experimental investigation demonstrates that the additional degrees of freedom provided in the parameterized family of updates can be leveraged to obtain non-trivial improvements both in synthetic domains and on popular deep RL benchmarks.

Saurabh Kumar, Daniel Fuchs, Klaus Spindler

Deviations from the approved design or processes during mass production can lead to unforeseen risks. However, these changes are sometimes necessary due to changes in the product design characteristics or an adaptation in the manufacturing process. A major challenge is to identify these risks early in the workflow so that failures leading to warranty claims can be avoided. We developed Fountain as a contextual assistant integrated in the deviation management workflow that helps in identifying the risks based on the description of the existing design and process criteria and the proposed deviation. In the manufacturing context, it is important that the assistant provides recommendations that are explainable and consistent. We achieve this through a combination of the following two components 1) language models finetuned for domain specific semantic similarity and, 2) knowledge representation in the form of a property graph derived from the bill of materials, Failure Modes and Effect Analysis (FMEA) and prior failures reported by customers. Here, we present the nuances of selecting and adapting pretrained language models for an engineering domain, continuous model updates based on user interaction with the contextual assistant and creating the causal chain for explainable recommendations based on the knowledge representation. Additionally, we demonstrate that the model adaptation is feasible using moderate computational infrastructure already available to most engineering teams in manufacturing organizations and inference can be performed on standard CPU only instances for integration with existing applications making these methods easily deployable.

Abhinav Parmar, Abhisek Panigrahi, Abhishek Kumar Dwivedi et al.

We present Mify-Coder, a 2.5B-parameter code model trained on 4.2T tokens using a compute-optimal strategy built on the Mify-2.5B foundation model. Mify-Coder achieves comparable accuracy and safety while significantly outperforming much larger baseline models on standard coding and function-calling benchmarks, demonstrating that compact models can match frontier-grade models in code generation and agent-driven workflows. Our training pipeline combines high-quality curated sources with synthetic data generated through agentically designed prompts, refined iteratively using enterprise-grade evaluation datasets. LLM-based quality filtering further enhances data density, enabling frugal yet effective training. Through disciplined exploration of CPT-SFT objectives, data mixtures, and sampling dynamics, we deliver frontier-grade code intelligence within a single continuous training trajectory. Empirical evidence shows that principled data and compute discipline allow smaller models to achieve competitive accuracy, efficiency, and safety compliance. Quantized variants of Mify-Coder enable deployment on standard desktop environments without requiring specialized hardware.

Saurabh Kumar, Hong Jun Jeon, Alex Lewandowski et al.

The "small agent, big world" frame offers a conceptual view that motivates the need for continual learning. The idea is that a small agent operating in a much bigger world cannot store all information that the world has to offer. To perform well, the agent must be carefully designed to ingest, retain, and eject the right information. To enable the development of performant continual learning agents, a number of synthetic environments have been proposed. However, these benchmarks suffer from limitations, including unnatural distribution shifts and a lack of fidelity to the "small agent, big world" framing. This paper aims to formalize two desiderata for the design of future simulated environments. These two criteria aim to reflect the objectives and complexity of continual learning in practical settings while enabling rapid prototyping of algorithms on a smaller scale.

Giacomo Bernardi, Ratul Mahajan, C. Seshadhri et al.

We design and deploy at Amazon the first production datacenter fabrics based on random graphs. While the cost and fault-tolerance benefits of such topologies have been long known, their practical realization has been hampered by a lack of scalable routing and cabling approaches. Our design, called RNG, has a new distributed routing protocol that exploits the properties of random graphs to find a large number of edge disjoint paths between endpoint pairs. A novel passive optical device that internally shuffles cable endpoints makes Amazon's cabling complexity similar to that of fat trees. We show that RNG fabrics match or exceed the performance of fat trees for a range of traffic patterns, despite being up to 45% cheaper. At Amazon, we made RNG the default datacenter fabric for most workloads.

Saurabh Kumar, Nishant Sinha

TryOnGAN is a recent virtual try-on approach, which generates highly realistic images and outperforms most previous approaches. In this article, we reproduce the TryOnGAN implementation and probe it along diverse angles: impact of transfer learning, variants of conditioning image generation with poses and properties of latent space interpolation. Some of these facets have never been explored in literature earlier. We find that transfer helps training initially but gains are lost as models train longer and pose conditioning via concatenation performs better. The latent space self-disentangles the pose and the style features and enables style transfer across poses. Our code and models are available in open source.

Pablo Samuel Castro, Subhodeep Moitra, Carles Gelada et al.

Deep reinforcement learning (deep RL) research has grown significantly in recent years. A number of software offerings now exist that provide stable, comprehensive implementations for benchmarking. At the same time, recent deep RL research has become more diverse in its goals. In this paper we introduce Dopamine, a new research framework for deep RL that aims to support some of that diversity. Dopamine is open-source, TensorFlow-based, and provides compact and reliable implementations of some state-of-the-art deep RL agents. We complement this offering with a taxonomy of the different research objectives in deep RL research. While by no means exhaustive, our analysis highlights the heterogeneity of research in the field, and the value of frameworks such as ours.

Vinay Kathiriya, Saurabh Kumar, Shashi Ranjan Kumar

This paper addresses the three-dimensional path-following guidance problem for unmanned aerial vehicles under explicit actuator constraints. Unlike conventional approaches that assume unbounded control inputs or handle saturation heuristically, the proposed method incorporates bounded lateral acceleration directly into the guidance design. A nonlinear guidance framework is developed employing a nested saturation-based control technique. The proposed guidance strategy guarantees bounded control inputs while ensuring exponential convergence of cross-track errors to zero. The formulation is applicable to general smooth paths and is systematically extended from planar to three-dimensional scenarios using a path-tangent coordinate framework. Rigorous stability analysis based on Lyapunov theory establishes convergence and feasibility properties of the closed-loop system. Numerical simulations on representative paths, including straight-line, circular, and sinusoidal paths, demonstrate that the proposed method achieves superior tracking performance, reduced control effort, and robustness against disturbances compared to existing guidance laws. The simplicity of the design and its compatibility with practical actuator limits make it suitable for real-world UAV applications.

Neeraj Agrawal, Saurabh Kumar, Priyanka Bhatt et al.

Hierarchical Text Classification (HTC) has recently gained traction given the ability to handle complex label hierarchy. This has found applications in domains like E- commerce, customer care and medicine industry among other real-world applications. Existing HTC models either encode label hierarchy separately and mix it with text encoding or guide the label hierarchy structure in the text encoder. Both approaches capture different characteristics of label hierarchy and are complementary to each other. In this paper, we propose a Hierarchical Text Classification using Contrastive Learning Informed Path guided hierarchy (HTC-CLIP), which learns hierarchy-aware text representation and text informed path guided hierarchy representation using contrastive learning. During the training of HTC-CLIP, we learn two different sets of class probabilities distributions and during inference, we use the pooled output of both probabilities for each class to get the best of both representations. Our results show that the two previous approaches can be effectively combined into one architecture to achieve improved performance. Tests on two public benchmark datasets showed an improvement of 0.99 - 2.37% in Macro F1 score using HTC-CLIP over the existing state-of-the-art models.

Saurabh Kumar, Sourav Bansal, Neeraj Agrawal et al.

Customer care is an essential pillar of the e-commerce shopping experience with companies spending millions of dollars each year, employing automation and human agents, across geographies (like US, Canada, Mexico, Chile), channels (like Chat, Interactive Voice Response (IVR)), and languages (like English, Spanish). SOTA pre-trained models like multilingual-BERT, fine-tuned on annotated data have shown good performance in downstream tasks relevant to Customer Care. However, model performance is largely subject to the availability of sufficient annotated domain-specific data. Cross-domain availability of data remains a bottleneck, thus building an intent classifier that generalizes across domains (defined by channel, geography, and language) with only a few annotations, is of great practical value. In this paper, we propose an embedder-cum-classifier model architecture which extends state-of-the-art domain-specific models to other domains with only a few labeled samples. We adopt a supervised fine-tuning approach with isotropic regularizers to train a domain-specific sentence embedder and a multilingual knowledge distillation strategy to generalize this embedder across multiple domains. The trained embedder, further augmented with a simple linear classifier can be deployed for new domains. Experiments on Canada and Mexico e-commerce Customer Care dataset with few-shot intent detection show an increase in accuracy by 20-23% against the existing state-of-the-art pre-trained models.

Saurabh Kumar, Jacob Buckman, Carles Gelada et al.

Transformers with linear attention offer significant computational advantages over softmax-based transformers but often suffer from degraded performance. The symmetric power (sympow) transformer, a particular type of linear transformer, addresses some of this performance gap by leveraging symmetric tensor embeddings, achieving comparable performance to softmax transformers. However, the finite capacity of the recurrent state in sympow transformers limits their ability to retain information, leading to performance degradation when scaling the training or evaluation context length. To address this issue, we propose the conformal-sympow transformer, which dynamically frees up capacity using data-dependent multiplicative gating and adaptively stores information using data-dependent rotary embeddings. Preliminary experiments on the LongCrawl64 dataset demonstrate that conformal-sympow overcomes the limitations of sympow transformers, achieving robust performance across scaled training and evaluation contexts.

Alex Lewandowski, Michał Bortkiewicz, Saurabh Kumar et al.

Loss of plasticity is a phenomenon where neural networks can become more difficult to train over the course of learning. Continual learning algorithms seek to mitigate this effect by sustaining good performance while maintaining network trainability. We develop a new technique for improving continual learning inspired by the observation that the singular values of the neural network parameters at initialization are an important factor for trainability during early phases of learning. From this perspective, we derive a new spectral regularizer for continual learning that better sustains these beneficial initialization properties throughout training. In particular, the regularizer keeps the maximum singular value of each layer close to one. Spectral regularization directly ensures that gradient diversity is maintained throughout training, which promotes continual trainability, while minimally interfering with performance in a single task. We present an experimental analysis that shows how the proposed spectral regularizer can sustain trainability and performance across a range of model architectures in continual supervised and reinforcement learning settings. Spectral regularization is less sensitive to hyperparameters while demonstrating better training in individual tasks, sustaining trainability as new tasks arrive, and achieving better generalization performance.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Chavhan Sujeet Yashavant, Saurabh Kumar, Amey Karkare

Smart contracts on Ethereum handle millions of U.S. Dollars and other financial assets. In the past, attackers have exploited smart contracts to steal these assets. The Ethereum community has developed plenty of tools to detect vulnerable smart contracts. However, there is no standardized data set to evaluate these existing tools, or any new tools developed. There is a need for an unbiased standard benchmark of real-world Ethereum smart contracts. We have created ScrawlD: an annotated data set of real-world smart contracts taken from the Ethereum network. The data set is labelled using 5 tools that detect various vulnerabilities in smart contracts, using majority voting.

Junfeng Wen, Saurabh Kumar, Ramki Gummadi et al.

Actor-critic (AC) methods are ubiquitous in reinforcement learning. Although it is understood that AC methods are closely related to policy gradient (PG), their precise connection has not been fully characterized previously. In this paper, we explain the gap between AC and PG methods by identifying the exact adjustment to the AC objective/gradient that recovers the true policy gradient of the cumulative reward objective (PG). Furthermore, by viewing the AC method as a two-player Stackelberg game between the actor and critic, we show that the Stackelberg policy gradient can be recovered as a special case of our more general analysis. Based on these results, we develop practical algorithms, Residual Actor-Critic and Stackelberg Actor-Critic, for estimating the correction between AC and PG and use these to modify the standard AC algorithm. Experiments on popular tabular and continuous environments show the proposed corrections can improve both the sample efficiency and final performance of existing AC methods.

Vaishakh Kedambaimoole, Neelotpala Kumar, Vijay Shirhatti et al.

The human body is punctuated with wide array of sensory systems that provide a high evolutionary advantage by facilitating formation of a detailed picture of the immediate surroundings. The sensors range across a wide spectrum, acquiring input from non-contact audio-visual means to contact based input via pressure and temperature. The ambit of sensing can be extended further by imparting the body with increased non-contact sensing capability through the phenomenon of electrostatics. Here we present graphene-based tattoo sensor for proximity sensing, employing the principle of electrostatic gating. The sensor shows a remarkable change in resistance upon exposure to objects surrounded with static charge on them. Compared to prior work in this field, the sensor has demonstrated the highest recorded proximity detection range of 20 cm. It is ultra-thin, highly skin conformal and comes with a facile transfer process such that it can be tattooed on highly curvilinear rough substrates like the human skin, unlike other graphene-based proximity sensors reported before. Present work details the operation of wearable proximity sensor while exploring the effect of mounting body on the working mechanism. A possible role of the sensor as an alerting system against unwarranted contact with objects in public places especially during the current SARS-CoV-2 pandemic has also been explored in the form of an LED bracelet whose color is controlled by the proximity sensor attached to it.

Kumar Ashutosh, Saurabh Kumar, Subhasis Chaudhuri

We present a classification based approach for the next best view selection and show how we can plausibly obtain a supervisory signal for this task. The proposed approach is end-to-end trainable and aims to get the best possible 3D reconstruction quality with a pair of passively acquired 2D views. The proposed model consists of two stages: a classifier and a reconstructor network trained jointly via the indirect 3D supervision from ground truth voxels. While testing, the proposed method assumes no prior knowledge of the underlying 3D shape for selecting the next best view. We demonstrate the proposed method's effectiveness via detailed experiments on synthetic and real images and show how it provides improved reconstruction quality than the existing state of the art 3D reconstruction and the next best view prediction techniques.

Saurabh Kumar, Aviral Kumar, Sergey Levine et al.

While reinforcement learning algorithms can learn effective policies for complex tasks, these policies are often brittle to even minor task variations, especially when variations are not explicitly provided during training. One natural approach to this problem is to train agents with manually specified variation in the training task or environment. However, this may be infeasible in practical situations, either because making perturbations is not possible, or because it is unclear how to choose suitable perturbation strategies without sacrificing performance. The key insight of this work is that learning diverse behaviors for accomplishing a task can directly lead to behavior that generalizes to varying environments, without needing to perform explicit perturbations during training. By identifying multiple solutions for the task in a single environment during training, our approach can generalize to new situations by abandoning solutions that are no longer effective and adopting those that are. We theoretically characterize a robustness set of environments that arises from our algorithm and empirically find that our diversity-driven approach can extrapolate to various changes in the environment and task.

Ayush Bhardwaj, Sakshee Pimpale, Saurabh Kumar et al.

Real-world scenarios pose several challenges to deep learning based computer vision techniques despite their tremendous success in research. Deeper models provide better performance, but are challenging to deploy and knowledge distillation allows us to train smaller models with minimal loss in performance. The model also has to deal with open set samples from classes outside the ones it was trained on and should be able to identify them as unknown samples while classifying the known ones correctly. Finally, most existing image recognition research focuses only on using two-dimensional snapshots of the real world three-dimensional objects. In this work, we aim to bridge these three research fields, which have been developed independently until now, despite being deeply interrelated. We propose a joint Knowledge Distillation and Open Set recognition training methodology for three-dimensional object recognition. We demonstrate the effectiveness of the proposed method via various experiments on how it allows us to obtain a much smaller model, which takes a minimal hit in performance while being capable of open set recognition for 3D point cloud data.

Saurabh Kumar, Siddharth Dangwal, Debanjan Bhowmik

Implementation of variational Quantum Machine Learning (QML) algorithms on Noisy Intermediate-Scale Quantum (NISQ) devices is known to have issues related to the high number of qubits needed and the noise associated with multi-qubit gates. In this paper, we propose a variational QML algorithm using a dressed quantum network to address these issues. Using the "super compressed encoding" scheme that we follow here, the classical encoding layer in our dressed network drastically scales down the input-dimension, before feeding the input to the variational quantum circuit. Hence, the number of qubits needed in our quantum circuit goes down drastically. Also, unlike in most other existing QML algorithms, our quantum circuit consists only of single-qubit gates, making it robust against noise. These factors make our algorithm suitable for implementation on NISQ hardware. To support our argument, we implement our algorithm on real NISQ hardware and thereby show accurate classification using popular machine learning data-sets like Fisher's Iris, Wisconsin's Breast Cancer (WBC), and Abalone. Then, to provide an intuitive explanation for our algorithm's working, we demonstrate the clustering of quantum states, which correspond to the input-samples of different output-classes, on the Bloch sphere (using WBC and MNIST data-sets). This clustering happens as a result of the training process followed in our algorithm. Through this Bloch-sphere-based representation, we also show the distinct roles played (in training) by the adjustable parameters of the classical encoding layer and the adjustable parameters of the variational quantum circuit. These parameters are adjusted iteratively during training through loss-minimization.

Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee et al.

Spiking Neural Networks (SNN) are energy-efficient computing architectures that exchange spikes for processing information, unlike classical Artificial Neural Networks (ANN). Due to this, SNNs are better suited for real-life deployments. However, similar to ANNs, SNNs also benefit from deeper architectures to obtain improved performance. Furthermore, like the deep ANNs, the memory, compute and power requirements of SNNs also increase with model size, and model compression becomes a necessity. Knowledge distillation is a model compression technique that enables transferring the learning of a large machine learning model to a smaller model with minimal loss in performance. In this paper, we propose techniques for knowledge distillation in spiking neural networks for the task of image classification. We present ways to distill spikes from a larger SNN, also called the teacher network, to a smaller one, also called the student network, while minimally impacting the classification accuracy. We demonstrate the effectiveness of the proposed method with detailed experiments on three standard datasets while proposing novel distillation methodologies and loss functions. We also present a multi-stage knowledge distillation technique for SNNs using an intermediate network to obtain higher performance from the student network. Our approach is expected to open up new avenues for deploying high performing large SNN models on resource-constrained hardware platforms.

Tianhe Yu, Saurabh Kumar, Abhishek Gupta et al.

While deep learning and deep reinforcement learning (RL) systems have demonstrated impressive results in domains such as image classification, game playing, and robotic control, data efficiency remains a major challenge. Multi-task learning has emerged as a promising approach for sharing structure across multiple tasks to enable more efficient learning. However, the multi-task setting presents a number of optimization challenges, making it difficult to realize large efficiency gains compared to learning tasks independently. The reasons why multi-task learning is so challenging compared to single-task learning are not fully understood. In this work, we identify a set of three conditions of the multi-task optimization landscape that cause detrimental gradient interference, and develop a simple yet general approach for avoiding such interference between task gradients. We propose a form of gradient surgery that projects a task's gradient onto the normal plane of the gradient of any other task that has a conflicting gradient. On a series of challenging multi-task supervised and multi-task RL problems, this approach leads to substantial gains in efficiency and performance. Further, it is model-agnostic and can be combined with previously-proposed multi-task architectures for enhanced performance.

Saurabh Kumar, Biplab Banerjee, Subhasis Chaudhuri

We deal with the problem of information fusion driven satellite image/scene classification and propose a generic hallucination architecture considering that all the available sensor information are present during training while some of the image modalities may be absent while testing. It is well-known that different sensors are capable of capturing complementary information for a given geographical area and a classification module incorporating information from all the sources are expected to produce an improved performance as compared to considering only a subset of the modalities. However, the classical classifier systems inherently require all the features used to train the module to be present for the test instances as well, which may not always be possible for typical remote sensing applications (say, disaster management). As a remedy, we provide a robust solution in terms of a hallucination module that can approximate the missing modalities from the available ones during the decision-making stage. In order to ensure better knowledge transfer during modality hallucination, we explicitly incorporate concepts of knowledge distillation for the purpose of exploring the privileged (side) information in our framework and subsequently introduce an intuitive modular training approach. The proposed network is evaluated extensively on a large-scale corpus of PAN-MS image pairs (scene recognition) as well as on a benchmark hyperspectral image dataset (image classification) where we follow different experimental scenarios and find that the proposed hallucination based module indeed is capable of capturing the multi-source information, albeit the explicit absence of some of the sensor information, and aid in improved scene characterization.

Carles Gelada, Saurabh Kumar, Jacob Buckman et al.

Many reinforcement learning (RL) tasks provide the agent with high-dimensional observations that can be simplified into low-dimensional continuous states. To formalize this process, we introduce the concept of a DeepMDP, a parameterized latent space model that is trained via the minimization of two tractable losses: prediction of rewards and prediction of the distribution over next latent states. We show that the optimization of these objectives guarantees (1) the quality of the latent space as a representation of the state space and (2) the quality of the DeepMDP as a model of the environment. We connect these results to prior work in the bisimulation literature, and explore the use of a variety of metrics. Our theoretical findings are substantiated by the experimental result that a trained DeepMDP recovers the latent structure underlying high-dimensional observations on a synthetic environment. Finally, we show that learning a DeepMDP as an auxiliary task in the Atari 2600 domain leads to large performance improvements over model-free RL.

Mark Rowland, Robert Dadashi, Saurabh Kumar et al.

We present a unifying framework for designing and analysing distributional reinforcement learning (DRL) algorithms in terms of recursively estimating statistics of the return distribution. Our key insight is that DRL algorithms can be decomposed as the combination of some statistical estimator and a method for imputing a return distribution consistent with that set of statistics. With this new understanding, we are able to provide improved analyses of existing DRL algorithms as well as construct a new algorithm (EDRL) based upon estimation of the expectiles of the return distribution. We compare EDRL with existing methods on a variety of MDPs to illustrate concrete aspects of our analysis, and develop a deep RL variant of the algorithm, ER-DQN, which we evaluate on the Atari-57 suite of games.

Saurabh Kumar, Pararth Shah, Dilek Hakkani-Tur et al.

We present a framework combining hierarchical and multi-agent deep reinforcement learning approaches to solve coordination problems among a multitude of agents using a semi-decentralized model. The framework extends the multi-agent learning setup by introducing a meta-controller that guides the communication between agent pairs, enabling agents to focus on communicating with only one other agent at any step. This hierarchical decomposition of the task allows for efficient exploration to learn policies that identify globally optimal solutions even as the number of collaborating agents increases. We show promising initial experimental results on a simulated distributed scheduling problem.

Himanshu Sahni, Saurabh Kumar, Farhan Tejani et al.

We present a differentiable framework capable of learning a wide variety of compositions of simple policies that we call skills. By recursively composing skills with themselves, we can create hierarchies that display complex behavior. Skill networks are trained to generate skill-state embeddings that are provided as inputs to a trainable composition function, which in turn outputs a policy for the overall task. Our experiments on an environment consisting of multiple collect and evade tasks show that this architecture is able to quickly build complex skills from simpler ones. Furthermore, the learned composition function displays some transfer to unseen combinations of skills, allowing for zero-shot generalizations.

Himanshu Sahni, Saurabh Kumar, Farhan Tejani et al.

Typical reinforcement learning (RL) agents learn to complete tasks specified by reward functions tailored to their domain. As such, the policies they learn do not generalize even to similar domains. To address this issue, we develop a framework through which a deep RL agent learns to generalize policies from smaller, simpler domains to more complex ones using a recurrent attention mechanism. The task is presented to the agent as an image and an instruction specifying the goal. This meta-controller guides the agent towards its goal by designing a sequence of smaller subtasks on the part of the state space within the attention, effectively decomposing it. As a baseline, we consider a setup without attention as well. Our experiments show that the meta-controller learns to create subgoals within the attention.