Laurent Itti

Yunhao Ge, Jie Ren, Andrew Gallagher et al. · deepmind

Multi-modal image-text models such as CLIP and LiT have demonstrated impressive performance on image classification benchmarks and their zero-shot generalization ability is particularly exciting. While the top-5 zero-shot accuracies of these models are very high, the top-1 accuracies are much lower (over 25% gap in some cases). We investigate the reasons for this performance gap and find that many of the failure cases are caused by ambiguity in the text prompts. First, we develop a simple and efficient zero-shot post-hoc method to identify images whose top-1 prediction is likely to be incorrect, by measuring consistency of the predictions w.r.t. multiple prompts and image transformations. We show that our procedure better predicts mistakes, outperforming the popular max logit baseline on selective prediction tasks. Next, we propose a simple and efficient way to improve accuracy on such uncertain images by making use of the WordNet hierarchy; specifically we augment the original class by incorporating its parent and children from the semantic label hierarchy, and plug the augmentation into text prompts. We conduct experiments on both CLIP and LiT models with five different ImageNet-based datasets. For CLIP, our method improves the top-1 accuracy by 17.13% on the uncertain subset and 3.6% on the entire ImageNet validation set. We also show that our method improves across ImageNet shifted datasets, four other datasets, and other model architectures such as LiT. The proposed method is hyperparameter-free, requires no additional model training and can be easily scaled to other large multi-modal architectures. Code is available at https://github.com/gyhandy/Hierarchy-CLIP.

Yunhao Ge, Jiashu Xu, Brian Nlong Zhao et al. · harvard

We propose a new paradigm to automatically generate training data with accurate labels at scale using the text-to-image synthesis frameworks (e.g., DALL-E, Stable Diffusion, etc.). The proposed approach1 decouples training data generation into foreground object generation, and contextually coherent background generation. To generate foreground objects, we employ a straightforward textual template, incorporating the object class name as input prompts. This is fed into a text-to-image synthesis framework, producing various foreground images set against isolated backgrounds. A foreground-background segmentation algorithm is then used to generate foreground object masks. To generate context images, we begin by creating language descriptions of the context. This is achieved by applying an image captioning method to a small set of images representing the desired context. These textual descriptions are then transformed into a diverse array of context images via a text-to-image synthesis framework. Subsequently, we composite these with the foreground object masks produced in the initial step, utilizing a cut-and-paste method, to formulate the training data. We demonstrate the advantages of our approach on five object detection and segmentation datasets, including Pascal VOC and COCO. We found that detectors trained solely on synthetic data produced by our method achieve performance comparable to those trained on real data (Fig. 1). Moreover, a combination of real and synthetic data yields even much better results. Further analysis indicates that the synthetic data distribution complements the real data distribution effectively. Additionally, we emphasize the compositional nature of our data generation approach in out-of-distribution and zero-shot data generation scenarios. We open-source our code at https://github.com/gyhandy/Text2Image-for-Detection

Yunhao Ge, Yao Xiao, Zhi Xu et al.

We investigate the contributions of three important features of the human visual system (HVS)~ -- ~shape, texture, and color ~ -- ~to object classification. We build a humanoid vision engine (HVE) that explicitly and separately computes shape, texture, and color features from images. The resulting feature vectors are then concatenated to support the final classification. We show that HVE can summarize and rank-order the contributions of the three features to object recognition. We use human experiments to confirm that both HVE and humans predominantly use some specific features to support the classification of specific classes (e.g., texture is the dominant feature to distinguish a zebra from other quadrupeds, both for humans and HVE). With the help of HVE, given any environment (dataset), we can summarize the most important features for the whole task (task-specific; e.g., color is the most important feature overall for classification with the CUB dataset), and for each class (class-specific; e.g., shape is the most important feature to recognize boats in the iLab-20M dataset). To demonstrate more usefulness of HVE, we use it to simulate the open-world zero-shot learning ability of humans with no attribute labeling. Finally, we show that HVE can also simulate human imagination ability with the combination of different features. We will open-source the HVE engine and corresponding datasets.

Yunhao Ge, Harkirat Behl, Jiashu Xu et al. · harvard, microsoft-research

Training computer vision models usually requires collecting and labeling vast amounts of imagery under a diverse set of scene configurations and properties. This process is incredibly time-consuming, and it is challenging to ensure that the captured data distribution maps well to the target domain of an application scenario. Recently, synthetic data has emerged as a way to address both of these issues. However, existing approaches either require human experts to manually tune each scene property or use automatic methods that provide little to no control; this requires rendering large amounts of random data variations, which is slow and is often suboptimal for the target domain. We present the first fully differentiable synthetic data pipeline that uses Neural Radiance Fields (NeRFs) in a closed-loop with a target application's loss function. Our approach generates data on-demand, with no human labor, to maximize accuracy for a target task. We illustrate the effectiveness of our method on synthetic and real-world object detection tasks. We also introduce a new "YCB-in-the-Wild" dataset and benchmark that provides a test scenario for object detection with varied poses in real-world environments.

Yunhao Ge, Yuecheng Li, Shuo Ni et al.

Continual learning aims to emulate the human ability to continually accumulate knowledge over sequential tasks. The main challenge is to maintain performance on previously learned tasks after learning new tasks, i.e., to avoid catastrophic forgetting. We propose a Channel-wise Lightweight Reprogramming (CLR) approach that helps convolutional neural networks (CNNs) overcome catastrophic forgetting during continual learning. We show that a CNN model trained on an old task (or self-supervised proxy task) could be ``reprogrammed" to solve a new task by using our proposed lightweight (very cheap) reprogramming parameter. With the help of CLR, we have a better stability-plasticity trade-off to solve continual learning problems: To maintain stability and retain previous task ability, we use a common task-agnostic immutable part as the shared ``anchor" parameter set. We then add task-specific lightweight reprogramming parameters to reinterpret the outputs of the immutable parts, to enable plasticity and integrate new knowledge. To learn sequential tasks, we only train the lightweight reprogramming parameters to learn each new task. Reprogramming parameters are task-specific and exclusive to each task, which makes our method immune to catastrophic forgetting. To minimize the parameter requirement of reprogramming to learn new tasks, we make reprogramming lightweight by only adjusting essential kernels and learning channel-wise linear mappings from anchor parameters to task-specific domain knowledge. We show that, for general CNNs, the CLR parameter increase is less than 0.6\% for any new task. Our method outperforms 13 state-of-the-art continual learning baselines on a new challenging sequence of 53 image classification datasets. Code and data are available at https://github.com/gyhandy/Channel-wise-Lightweight-Reprogramming

Yunhao Ge, Jiashu Xu, Brian Nlong Zhao et al. · harvard

We propose a new paradigm to automatically generate training data with accurate labels at scale using the text-toimage synthesis frameworks (e.g., DALL-E, Stable Diffusion, etc.). The proposed approach decouples training data generation into foreground object mask generation and background (context) image generation. For foreground object mask generation, we use a simple textual template with object class name as input to DALL-E to generate a diverse set of foreground images. A foreground-background segmentation algorithm is then used to generate foreground object masks. Next, in order to generate context images, first a language description of the context is generated by applying an image captioning method on a small set of images representing the context. These language descriptions are then used to generate diverse sets of context images using the DALL-E framework. These are then composited with object masks generated in the first step to provide an augmented training set for a classifier. We demonstrate the advantages of our approach on four object detection datasets including on Pascal VOC and COCO object detection tasks. Furthermore, we also highlight the compositional nature of our data generation approach on out-of-distribution and zero-shot data generation scenarios.

Sumedh A Sontakke, Jesse Zhang, Sébastien M. R. Arnold et al.

Reward specification is a notoriously difficult problem in reinforcement learning, requiring extensive expert supervision to design robust reward functions. Imitation learning (IL) methods attempt to circumvent these problems by utilizing expert demonstrations but typically require a large number of in-domain expert demonstrations. Inspired by advances in the field of Video-and-Language Models (VLMs), we present RoboCLIP, an online imitation learning method that uses a single demonstration (overcoming the large data requirement) in the form of a video demonstration or a textual description of the task to generate rewards without manual reward function design. Additionally, RoboCLIP can also utilize out-of-domain demonstrations, like videos of humans solving the task for reward generation, circumventing the need to have the same demonstration and deployment domains. RoboCLIP utilizes pretrained VLMs without any finetuning for reward generation. Reinforcement learning agents trained with RoboCLIP rewards demonstrate 2-3 times higher zero-shot performance than competing imitation learning methods on downstream robot manipulation tasks, doing so using only one video/text demonstration.

Yunhao Ge, Jiashu Xu, Brian Nlong Zhao et al. · harvard

We propose EM-PASTE: an Expectation Maximization(EM) guided Cut-Paste compositional dataset augmentation approach for weakly-supervised instance segmentation using only image-level supervision. The proposed method consists of three main components. The first component generates high-quality foreground object masks. To this end, an EM-like approach is proposed that iteratively refines an initial set of object mask proposals generated by a generic region proposal method. Next, in the second component, high-quality context-aware background images are generated using a text-to-image compositional synthesis method like DALL-E. Finally, the third component creates a large-scale pseudo-labeled instance segmentation training dataset by compositing the foreground object masks onto the original and generated background images. The proposed approach achieves state-of-the-art weakly-supervised instance segmentation results on both the PASCAL VOC 2012 and MS COCO datasets by using only image-level, weak label information. In particular, it outperforms the best baseline by +7.4 and +2.8 mAP0.50 on PASCAL and COCO, respectively. Further, the method provides a new solution to the long-tail weakly-supervised instance segmentation problem (when many classes may only have few training samples), by selectively augmenting under-represented classes.

Yipeng Gao, Lei Shu, Genzhi Ye et al.

Procedural 3D modeling through code is emerging as a versatile paradigm, offering deterministic, engine-ready, and precisely editable assets that neural 3D generators inherently lack. Authoring such procedural content, however, demands deep expertise in 3D software APIs, parametric design, and code-level geometric reasoning. In this paper, we propose 3DCodeBench, a systematic benchmark for evaluating vision-language model (VLM) agents for procedural 3D generation in 3D modeling software. Specifically, 3DCodeBench evaluates how effectively 12 advanced VLMs can serve as procedural 3D modelers by translating text and image references into procedural code for 3D modeling software. Recognizing that automated metrics may not fully capture the perceptual quality of 3D shapes, we build 3DCodeArena, a ranking platform based on pairwise human preferences over generated 3D outputs. From extensive evaluations and results, we observe that: (1) Failures mostly arise from API mismatches, while successful renders still suffer from disconnected or floating 3D geometric components. (2) Test-time scaling, such as higher thinking budgets and multi-turn refinement, improves performance overall. Our findings highlight a critical need for high-quality procedural coding data to advance commercial VLMs. Furthermore, effective procedural 3D modeling requires a robust execution environment that provides high-fidelity feedback for iterative refinement. We release 3DCodeBench, including the curated large-scale dataset of multimodal (text/image) prompts, procedural code, 3D object triplets, evaluation protocol, and the public 3DCodeArena platform as a foundational toolkit for exploring VLM-based procedural 3D modelers.

Yunhao Ge, Sercan Ö. Arik, Jinsung Yoon et al.

Learning the causal structure behind data is invaluable for improving generalization and obtaining high-quality explanations. We propose a novel framework, Invariant Structure Learning (ISL), that is designed to improve causal structure discovery by utilizing generalization as an indication. ISL splits the data into different environments, and learns a structure that is invariant to the target across different environments by imposing a consistency constraint. An aggregation mechanism then selects the optimal classifier based on a graph structure that reflects the causal mechanisms in the data more accurately compared to the structures learnt from individual environments. Furthermore, we extend ISL to a self-supervised learning setting where accurate causal structure discovery does not rely on any labels. This self-supervised ISL utilizes invariant causality proposals by iteratively setting different nodes as targets. On synthetic and real-world datasets, we demonstrate that ISL accurately discovers the causal structure, outperforms alternative methods, and yields superior generalization for datasets with significant distribution shifts.

Iordanis Fostiropoulos, Laurent Itti

Transformations in the input space of Deep Neural Networks (DNN) lead to unintended changes in the feature space. Almost perceptually identical inputs, such as adversarial examples, can have significantly distant feature representations. On the contrary, Out-of-Distribution (OOD) samples can have highly similar feature representations to training set samples. Our theoretical analysis for DNNs trained with a categorical classification head suggests that the inflexible logit space restricted by the classification problem size is one of the root causes for the lack of $\textit{robustness}$. Our second observation is that DNNs over-fit to the training augmentation technique and do not learn $\textit{nuance invariant}$ representations. Inspired by the recent success of prototypical and contrastive learning frameworks for both improving robustness and learning nuance invariant representations, we propose a training framework, $\textbf{Supervised Contrastive Prototype Learning}$ (SCPL). We use N-pair contrastive loss with prototypes of the same and opposite classes and replace a categorical classification head with a $\textbf{Prototype Classification Head}$ (PCH). Our approach is $\textit{sample efficient}$, does not require $\textit{sample mining}$, can be implemented on any existing DNN without modification to their architecture, and combined with other training augmentation techniques. We empirically evaluate the $\textbf{clean}$ robustness of our method on out-of-distribution and adversarial samples. Our framework outperforms other state-of-the-art contrastive and prototype learning approaches in $\textit{robustness}$.

Wanrong Zheng, Yunhao Ge, Laurent Itti

Breakthrough progress in vision-based navigation through unknown environments has been achieved by using multimodal large language models (MLLMs). These models can plan a sequence of motions by evaluating the current view at each time step against the task and goal given to the agent. However, current zero-shot Vision-and-Language Navigation (VLN) agents powered by MLLMs still tend to drift off course, halt prematurely, and achieve low overall success rates. We propose Three-Step Nav to counteract these failures with a three-view protocol: First, "look forward" to extract global landmarks and sketch a coarse plan. Then, "look now" to align the current visual observation with the next sub-goal for fine-grained guidance. Finally, "look backward" audits the entire trajectory to correct accumulated drift before stopping. Requiring no gradient updates or task-specific fine-tuning, our planner drops into existing VLN pipelines with minimal overhead. Three-Step Nav achieves state-of-the-art zero-shot performance on the R2R-CE and RxR-CE dataset. Our code is available at https://github.com/ZoeyZheng0/3-step-Nav.

Kiran Lekkala, Eshan Bhargava, Yunhao Ge et al.

We create a novel benchmark for evaluating a Deployable Lifelong Learning system for Visual Reinforcement Learning (RL) that is pretrained on a curated dataset, and propose a novel Scalable Lifelong Learning system capable of retaining knowledge from the previously learnt RL tasks. Our benchmark measures the efficacy of a deployable Lifelong Learning system that is evaluated on scalability, performance and resource utilization. Our proposed system, once pretrained on the dataset, can be deployed to perform continual learning on unseen tasks. Our proposed method consists of a Few Shot Class Incremental Learning (FSCIL) based task-mapper and an encoder/backbone trained entirely using the pretrain dataset. The policy parameters corresponding to the recognized task are then loaded to perform the task. We show that this system can be scaled to incorporate a large number of tasks due to the small memory footprint and fewer computational resources. We perform experiments on our DeLL (Deployment for Lifelong Learning) benchmark on the Atari games to determine the efficacy of the system.

Kiran Lekkala, Chen Liu, Laurent Itti

Sim2Real transfer has gained popularity because it helps transfer from inexpensive simulators to real world. This paper presents a novel system that fuses components in a traditional World Model into a robust system, trained entirely within a simulator, that Zero-Shot transfers to the real world. To facilitate transfer, we use an intermediary representation that is based on \textit{Bird's Eye View (BEV)} images. Thus, our robot learns to navigate in a simulator by first learning to translate from complex \textit{First-Person View (FPV)} based RGB images to BEV representations, then learning to navigate using those representations. Later, when tested in the real world, the robot uses the perception model that translates FPV-based RGB images to embeddings that were learned by the FPV to BEV translator and that can be used by the downstream policy. The incorporation of state-checking modules using \textit{Anchor images} and Mixture Density LSTM not only interpolates uncertain and missing observations but also enhances the robustness of the model in the real-world. We trained the model using data from a Differential drive robot in the CARLA simulator. Our methodology's effectiveness is shown through the deployment of trained models onto a real-world Differential drive robot. Lastly we release a comprehensive codebase, dataset and models for training and deployment (\url{https://sites.google.com/view/value-explicit-pretraining}).

Yunhao Ge, Jie Ren, Jiaping Zhao et al.

We focus on the challenge of out-of-distribution (OOD) detection in deep learning models, a crucial aspect in ensuring reliability. Despite considerable effort, the problem remains significantly challenging in deep learning models due to their propensity to output over-confident predictions for OOD inputs. We propose a novel one-class open-set OOD detector that leverages text-image pre-trained models in a zero-shot fashion and incorporates various descriptions of in-domain and OOD. Our approach is designed to detect anything not in-domain and offers the flexibility to detect a wide variety of OOD, defined via fine- or coarse-grained labels, or even in natural language. We evaluate our approach on challenging benchmarks including large-scale datasets containing fine-grained, semantically similar classes, distributionally shifted images, and multi-object images containing a mixture of in-domain and OOD objects. Our method shows superior performance over previous methods on all benchmarks. Code is available at https://github.com/gyhandy/One-Class-Anything

Iordanis Fostiropoulos, Jiaye Zhu, Laurent Itti

In Continual Learning (CL), a model is required to learn a stream of tasks sequentially without significant performance degradation on previously learned tasks. Current approaches fail for a long sequence of tasks from diverse domains and difficulties. Many of the existing CL approaches are difficult to apply in practice due to excessive memory cost or training time, or are tightly coupled to a single device. With the intuition derived from the widely applied mini-batch training, we propose Batch Model Consolidation ($\textbf{BMC}$) to support more realistic CL under conditions where multiple agents are exposed to a range of tasks. During a $\textit{regularization}$ phase, BMC trains multiple $\textit{expert models}$ in parallel on a set of disjoint tasks. Each expert maintains weight similarity to a $\textit{base model}$ through a $\textit{stability loss}$, and constructs a $\textit{buffer}$ from a fraction of the task's data. During the $\textit{consolidation}$ phase, we combine the learned knowledge on 'batches' of $\textit{expert models}$ using a $\textit{batched consolidation loss}$ in $\textit{memory}$ data that aggregates all buffers. We thoroughly evaluate each component of our method in an ablation study and demonstrate the effectiveness on standardized benchmark datasets Split-CIFAR-100, Tiny-ImageNet, and the Stream dataset composed of 71 image classification tasks from diverse domains and difficulties. Our method outperforms the next best CL approach by 70% and is the only approach that can maintain performance at the end of 71 tasks; Our benchmark can be accessed at https://github.com/fostiropoulos/stream_benchmark

Yunhao Ge, Yuecheng Li, Di Wu et al.

In Lifelong Learning (LL), agents continually learn as they encounter new conditions and tasks. Most current LL is limited to a single agent that learns tasks sequentially. Dedicated LL machinery is then deployed to mitigate the forgetting of old tasks as new tasks are learned. This is inherently slow. We propose a new Shared Knowledge Lifelong Learning (SKILL) challenge, which deploys a decentralized population of LL agents that each sequentially learn different tasks, with all agents operating independently and in parallel. After learning their respective tasks, agents share and consolidate their knowledge over a decentralized communication network, so that, in the end, all agents can master all tasks. We present one solution to SKILL which uses Lightweight Lifelong Learning (LLL) agents, where the goal is to facilitate efficient sharing by minimizing the fraction of the agent that is specialized for any given task. Each LLL agent thus consists of a common task-agnostic immutable part, where most parameters are, and individual task-specific modules that contain fewer parameters but are adapted to each task. Agents share their task-specific modules, plus summary information ("task anchors") representing their tasks in the common task-agnostic latent space of all agents. Receiving agents register each received task-specific module using the corresponding anchor. Thus, every agent improves its ability to solve new tasks each time new task-specific modules and anchors are received. On a new, very challenging SKILL-102 dataset with 102 image classification tasks (5,033 classes in total, 2,041,225 training, 243,464 validation, and 243,464 test images), we achieve much higher (and SOTA) accuracy over 8 LL baselines, while also achieving near perfect parallelization. Code and data can be found at https://github.com/gyhandy/Shared-Knowledge-Lifelong-Learning

Yunhao Ge, Sami Abu-El-Haija, Gan Xin et al.

Visual cognition of primates is superior to that of artificial neural networks in its ability to 'envision' a visual object, even a newly-introduced one, in different attributes including pose, position, color, texture, etc. To aid neural networks to envision objects with different attributes, we propose a family of objective functions, expressed on groups of examples, as a novel learning framework that we term Group-Supervised Learning (GSL). GSL allows us to decompose inputs into a disentangled representation with swappable components, that can be recombined to synthesize new samples. For instance, images of red boats & blue cars can be decomposed and recombined to synthesize novel images of red cars. We propose an implementation based on auto-encoder, termed group-supervised zero-shot synthesis network (GZS-Net) trained with our learning framework, that can produce a high-quality red car even if no such example is witnessed during training. We test our model and learning framework on existing benchmarks, in addition to anew dataset that we open-source. We qualitatively and quantitatively demonstrate that GZS-Net trained with GSL outperforms state-of-the-art methods.

Jiaping Zhao, Laurent Itti

Dynamic Time Warping (DTW) is an algorithm to align temporal sequences with possible local non-linear distortions, and has been widely applied to audio, video and graphics data alignments. DTW is essentially a point-to-point matching method under some boundary and temporal consistency constraints. Although DTW obtains a global optimal solution, it does not necessarily achieve locally sensible matchings. Concretely, two temporal points with entirely dissimilar local structures may be matched by DTW. To address this problem, we propose an improved alignment algorithm, named shape Dynamic Time Warping (shapeDTW), which enhances DTW by taking point-wise local structural information into consideration. shapeDTW is inherently a DTW algorithm, but additionally attempts to pair locally similar structures and to avoid matching points with distinct neighborhood structures. We apply shapeDTW to align audio signal pairs having ground-truth alignments, as well as artificially simulated pairs of aligned sequences, and obtain quantitatively much lower alignment errors than DTW and its two variants. When shapeDTW is used as a distance measure in a nearest neighbor classifier (NN-shapeDTW) to classify time series, it beats DTW on 64 out of 84 UCR time series datasets, with significantly improved classification accuracies. By using a properly designed local structure descriptor, shapeDTW improves accuracies by more than 10% on 18 datasets. To the best of our knowledge, shapeDTW is the first distance measure under the nearest neighbor classifier scheme to significantly outperform DTW, which had been widely recognized as the best distance measure to date. Our code is publicly accessible at: https://github.com/jiapingz/shapeDTW.

Yunhao Ge, Hong-Xing Yu, Cheng Zhao et al. · stanford

A major challenge in monocular 3D object detection is the limited diversity and quantity of objects in real datasets. While augmenting real scenes with virtual objects holds promise to improve both the diversity and quantity of the objects, it remains elusive due to the lack of an effective 3D object insertion method in complex real captured scenes. In this work, we study augmenting complex real indoor scenes with virtual objects for monocular 3D object detection. The main challenge is to automatically identify plausible physical properties for virtual assets (e.g., locations, appearances, sizes, etc.) in cluttered real scenes. To address this challenge, we propose a physically plausible indoor 3D object insertion approach to automatically copy virtual objects and paste them into real scenes. The resulting objects in scenes have 3D bounding boxes with plausible physical locations and appearances. In particular, our method first identifies physically feasible locations and poses for the inserted objects to prevent collisions with the existing room layout. Subsequently, it estimates spatially-varying illumination for the insertion location, enabling the immersive blending of the virtual objects into the original scene with plausible appearances and cast shadows. We show that our augmentation method significantly improves existing monocular 3D object models and achieves state-of-the-art performance. For the first time, we demonstrate that a physically plausible 3D object insertion, serving as a generative data augmentation technique, can lead to significant improvements for discriminative downstream tasks such as monocular 3D object detection. Project website: https://gyhandy.github.io/3D-Copy-Paste/

Yunhao Ge, Yihe Tang, Jiashu Xu et al. · stanford

The systematic evaluation and understanding of computer vision models under varying conditions require large amounts of data with comprehensive and customized labels, which real-world vision datasets rarely satisfy. While current synthetic data generators offer a promising alternative, particularly for embodied AI tasks, they often fall short for computer vision tasks due to low asset and rendering quality, limited diversity, and unrealistic physical properties. We introduce the BEHAVIOR Vision Suite (BVS), a set of tools and assets to generate fully customized synthetic data for systematic evaluation of computer vision models, based on the newly developed embodied AI benchmark, BEHAVIOR-1K. BVS supports a large number of adjustable parameters at the scene level (e.g., lighting, object placement), the object level (e.g., joint configuration, attributes such as "filled" and "folded"), and the camera level (e.g., field of view, focal length). Researchers can arbitrarily vary these parameters during data generation to perform controlled experiments. We showcase three example application scenarios: systematically evaluating the robustness of models across different continuous axes of domain shift, evaluating scene understanding models on the same set of images, and training and evaluating simulation-to-real transfer for a novel vision task: unary and binary state prediction. Project website: https://behavior-vision-suite.github.io/

Brian Nlong Zhao, Yuhang Xiao, Jiashu Xu et al. · harvard

The popularization of Text-to-Image (T2I) diffusion models enables the generation of high-quality images from text descriptions. However, generating diverse customized images with reference visual attributes remains challenging. This work focuses on personalizing T2I diffusion models at a more abstract concept or category level, adapting commonalities from a set of reference images while creating new instances with sufficient variations. We introduce a solution that allows a pretrained T2I diffusion model to learn a set of soft prompts, enabling the generation of novel images by sampling prompts from the learned distribution. These prompts offer text-guided editing capabilities and additional flexibility in controlling variation and mixing between multiple distributions. We also show the adaptability of the learned prompt distribution to other tasks, such as text-to-3D. Finally we demonstrate effectiveness of our approach through quantitative analysis including automatic evaluation and human assessment. Project website: https://briannlongzhao.github.io/DreamDistribution

Hae Jin Song, Laurent Itti

Recent breakthroughs and rapid integration of generative models (GMs) have sparked interest in the problem of model attribution and their fingerprints. For instance, service providers need reliable methods of authenticating their models to protect their IP, while users and law enforcement seek to verify the source of generated content for accountability and trust. In addition, a growing threat of model collapse is arising, as more model-generated data are being fed back into sources (e.g., YouTube) that are often harvested for training ("regurgitative training"), heightening the need to differentiate synthetic from human data. Yet, a gap still exists in understanding generative models' fingerprints, we believe, stemming from the lack of a formal framework that can define, represent, and analyze the fingerprints in a principled way. To address this gap, we take a geometric approach and propose a new definition of artifact and fingerprint of GMs using Riemannian geometry, which allows us to leverage the rich theory of differential geometry. Our new definition generalizes previous work (Song et al., 2024) to non-Euclidean manifolds by learning Riemannian metrics from data and replacing the Euclidean distances and nearest-neighbor search with geodesic distances and kNN-based Riemannian center of mass. We apply our theory to a new gradient-based algorithm for computing the fingerprints in practice. Results show that it is more effective in distinguishing a large array of GMs, spanning across 4 different datasets in 2 different resolutions (64 by 64, 256 by 256), 27 model architectures, and 2 modalities (Vision, Vision-Language). Using our proposed definition significantly improves the performance on model attribution, as well as a generalization to unseen datasets, model types, and modalities, suggesting its practical efficacy.

Rorry Brenner, Laurent Itti

The neurons of artificial neural networks were originally invented when much less was known about biological neurons than is known today. Our work explores a modification to the core neuron unit to make it more parallel to a biological neuron. The modification is made with the knowledge that biological dendrites are not simply passive activation funnels, but also compute complex non-linear functions as they transmit activation to the cell body. The paper explores a novel system of "Perforated" backpropagation empowering the artificial neurons of deep neural networks to achieve better performance coding for the same features they coded for in the original architecture. After an initial network training phase, additional "Dendrite Nodes" are added to the network and separately trained with a different objective: to correlate their output with the remaining error of the original neurons. The trained Dendrite Nodes are then frozen, and the original neurons are further trained, now taking into account the additional error signals provided by the Dendrite Nodes. The cycle of training the original neurons and then adding and training Dendrite Nodes can be repeated several times until satisfactory performance is achieved. Our algorithm was successfully added to modern state-of-the-art PyTorch networks across multiple domains, improving upon original accuracies and allowing for significant model compression without a loss in accuracy.

Ajay Narayanan Sridhar, Fuli Qiao, Nelson Daniel Troncoso Aldas et al.

People with visual impairments often face significant challenges in locating and retrieving objects in their surroundings. Existing assistive technologies present a trade-off: systems that offer precise guidance typically require pre-scanning or support only fixed object categories, while those with open-world object recognition lack spatial feedback for reaching the object. To address this gap, we introduce 'NaviSense', a mobile assistive system that combines conversational AI, vision-language models, augmented reality (AR), and LiDAR to support open-world object detection with real-time audio-haptic guidance. Users specify objects via natural language and receive continuous spatial feedback to navigate toward the target without needing prior setup. Designed with insights from a formative study and evaluated with 12 blind and low-vision participants, NaviSense significantly reduced object retrieval time and was preferred over existing tools, demonstrating the value of integrating open-world perception with precise, accessible guidance.

Rorry Brenner, Evan Davis, Rushi Chaudhari et al.

Perforated Backpropagation is a neural network optimization technique based on modern understanding of the computational importance of dendrites within biological neurons. This paper explores further experiments from the original publication, generated from a hackathon held at the Carnegie Mellon Swartz Center in February 2025. Students and local Pittsburgh ML practitioners were brought together to experiment with the Perforated Backpropagation algorithm on the datasets and models which they were using for their projects. Results showed that the system could enhance their projects, with up to 90% model compression without negative impact on accuracy, or up to 16% increased accuracy of their original models.

Kiran Lekkala, Henghui Bao, Sumedh Sontakke et al.

We propose Value Explicit Pretraining (VEP), a method that learns generalizable representations for transfer reinforcement learning. VEP enables learning of new tasks that share similar objectives as previously learned tasks, by learning an encoder for objective-conditioned representations, irrespective of appearance changes and environment dynamics. To pre-train the encoder from a sequence of observations, we use a self-supervised contrastive loss that results in learning temporally smooth representations. VEP learns to relate states across different tasks based on the Bellman return estimate that is reflective of task progress. Experiments using a realistic navigation simulator and Atari benchmark show that the pretrained encoder produced by our method outperforms current SoTA pretraining methods on the ability to generalize to unseen tasks. VEP achieves up to a 2 times improvement in rewards on Atari and visual navigation, and up to a 3 times improvement in sample efficiency. For videos of policy performance visit our https://sites.google.com/view/value-explicit-pretraining/

Iordanis Fostiropoulos, Bowman Brown, Laurent Itti

Machine learning is facing a 'reproducibility crisis' where a significant number of works report failures when attempting to reproduce previously published results. We evaluate the sources of reproducibility failures using a meta-analysis of 142 replication studies from ReScience C and 204 code repositories. We find that missing experiment details such as hyperparameters are potential causes of unreproducibility. We experimentally show the bias of different hyperparameter selection strategies and conclude that consolidated artifacts with a unified framework can help support reproducibility.

Sumedh A Sontakke, Buvaneswari Ramanan, Laurent Itti et al.

Out-of-distribution detection is an important capability that has long eluded vanilla neural networks. Deep Neural networks (DNNs) tend to generate over-confident predictions when presented with inputs that are significantly out-of-distribution (OOD). This can be dangerous when employing machine learning systems in the wild as detecting attacks can thus be difficult. Recent advances inference-time out-of-distribution detection help mitigate some of these problems. However, existing methods can be restrictive as they are often computationally expensive. Additionally, these methods require training of a downstream detector model which learns to detect OOD inputs from in-distribution ones. This, therefore, adds latency during inference. Here, we offer an information theoretic perspective on why neural networks are inherently incapable of OOD detection. We attempt to mitigate these flaws by converting a trained model into a an OOD detector using a handful of steps of gradient descent. Our work can be employed as a post-processing method whereby an inference-time ML system can convert a trained model into an OOD detector. Experimentally, we show how our method consistently outperforms the state-of-the-art in detection accuracy on popular image datasets while also reducing computational complexity.

Shixian Wen, Amanda Sofie Rios, Kiran Lekkala et al.

Understanding the patterns of misclassified ImageNet images is particularly important, as it could guide us to design deep neural networks (DNN) that generalize better. However, the richness of ImageNet imposes difficulties for researchers to visually find any useful patterns of misclassification. Here, to help find these patterns, we propose "Superclassing ImageNet dataset". It is a subset of ImageNet which consists of 10 superclasses, each containing 7-116 related subclasses (e.g., 52 bird types, 116 dog types). By training neural networks on this dataset, we found that: (i) Misclassifications are rarely across superclasses, but mainly among subclasses within a superclass. (ii) Ensemble networks trained each only on subclasses of a given superclass perform better than the same network trained on all subclasses of all superclasses. Hence, we propose a two-stage Super-Sub framework, and demonstrate that: (i) The framework improves overall classification performance by 3.3%, by first inferring a superclass using a generalist superclass-level network, and then using a specialized network for final subclass-level classification. (ii) Although the total parameter storage cost increases to a factor N+1 for N superclasses compared to using a single network, with finetuning, delta and quantization aware training techniques this can be reduced to 0.2N+1. Another advantage of this efficient implementation is that the memory cost on the GPU during inference is equivalent to using only one network. The reason is we initiate each subclass-level network through addition of small parameter variations (deltas) to the superclass-level network. (iii) Finally, our framework promises to be more scalable and generalizable than the common alternative of simply scaling up a vanilla network in size, since very large networks often suffer from overfitting and gradient vanishing.

Yunhao Ge, Zhi Xu, Yao Xiao et al.

We focus on controllable disentangled representation learning (C-Dis-RL), where users can control the partition of the disentangled latent space to factorize dataset attributes (concepts) for downstream tasks. Two general problems remain under-explored in current methods: (1) They lack comprehensive disentanglement constraints, especially missing the minimization of mutual information between different attributes across latent and observation domains. (2) They lack convexity constraints, which is important for meaningfully manipulating specific attributes for downstream tasks. To encourage both comprehensive C-Dis-RL and convexity simultaneously, we propose a simple yet efficient method: Controllable Interpolation Regularization (CIR), which creates a positive loop where disentanglement and convexity can help each other. Specifically, we conduct controlled interpolation in latent space during training, and we reuse the encoder to help form a 'perfect disentanglement' regularization. In that case, (a) disentanglement loss implicitly enlarges the potential understandable distribution to encourage convexity; (b) convexity can in turn improve robust and precise disentanglement. CIR is a general module and we merge CIR with three different algorithms: ELEGANT, I2I-Dis, and GZS-Net to show the compatibility and effectiveness. Qualitative and quantitative experiments show improvement in C-Dis-RL and latent convexity by CIR. This further improves downstream tasks: controllable image synthesis, cross-modality image translation, and zero-shot synthesis.

Sumedh A Sontakke, Stephen Iota, Zizhao Hu et al.

Out-of-distribution (OOD) detection is a well-studied topic in supervised learning. Extending the successes in supervised learning methods to the reinforcement learning (RL) setting, however, is difficult due to the data generating process - RL agents actively query their environment for data, and the data are a function of the policy followed by the agent. An agent could thus neglect a shift in the environment if its policy did not lead it to explore the aspect of the environment that shifted. Therefore, to achieve safe and robust generalization in RL, there exists an unmet need for OOD detection through active experimentation. Here, we attempt to bridge this lacuna by first defining a causal framework for OOD scenarios or environments encountered by RL agents in the wild. Then, we propose a novel task: that of Out-of-Task Distribution (OOTD) detection. We introduce an RL agent that actively experiments in a test environment and subsequently concludes whether it is OOTD or not. We name our method GalilAI, in honor of Galileo Galilei, as it discovers, among other causal processes, that gravitational acceleration is independent of the mass of a body. Finally, we propose a simple probabilistic neural network baseline for comparison, which extends extant Model-Based RL. We find that GalilAI outperforms the baseline significantly. See visualizations of our method https://galil-ai.github.io/

Sumedh A Sontakke, Sumegh Roychowdhury, Mausoom Sarkar et al.

Humans excel at learning long-horizon tasks from demonstrations augmented with textual commentary, as evidenced by the burgeoning popularity of tutorial videos online. Intuitively, this capability can be separated into 2 distinct subtasks - first, dividing a long-horizon demonstration sequence into semantically meaningful events; second, adapting such events into meaningful behaviors in one's own environment. Here, we present Video2Skill (V2S), which attempts to extend this capability to artificial agents by allowing a robot arm to learn from human cooking videos. We first use sequence-to-sequence Auto-Encoder style architectures to learn a temporal latent space for events in long-horizon demonstrations. We then transfer these representations to the robotic target domain, using a small amount of offline and unrelated interaction data (sequences of state-action pairs of the robot arm controlled by an expert) to adapt these events into actionable representations, i.e., skills. Through experiments, we demonstrate that our approach results in self-supervised analogy learning, where the agent learns to draw analogies between motions in human demonstration data and behaviors in the robotic environment. We also demonstrate the efficacy of our approach on model learning - demonstrating how Video2Skill utilizes prior knowledge from human demonstration to outperform traditional model learning of long-horizon dynamics. Finally, we demonstrate the utility of our approach for non-tabula rasa decision-making, i.e, utilizing video demonstration for zero-shot skill generation.

Kiran Lekkala, Laurent Itti

In this paper, we try to improve exploration in Blackbox methods, particularly Evolution strategies (ES), when applied to Reinforcement Learning (RL) problems where intermediate waypoints/subgoals are available. Since Evolutionary strategies are highly parallelizable, instead of extracting just a scalar cumulative reward, we use the state-action pairs from the trajectories obtained during rollouts/evaluations, to learn the dynamics of the agent. The learnt dynamics are then used in the optimization procedure to speed-up training. Lastly, we show how our proposed approach is universally applicable by presenting results from experiments conducted on Carla driving and UR5 robotic arm simulators.

Yunhao Ge, Yao Xiao, Zhi Xu et al.

Despite substantial progress in applying neural networks (NN) to a wide variety of areas, they still largely suffer from a lack of transparency and interpretability. While recent developments in explainable artificial intelligence attempt to bridge this gap (e.g., by visualizing the correlation between input pixels and final outputs), these approaches are limited to explaining low-level relationships, and crucially, do not provide insights on error correction. In this work, we propose a framework (VRX) to interpret classification NNs with intuitive structural visual concepts. Given a trained classification model, the proposed VRX extracts relevant class-specific visual concepts and organizes them using structural concept graphs (SCG) based on pairwise concept relationships. By means of knowledge distillation, we show VRX can take a step towards mimicking the reasoning process of NNs and provide logical, concept-level explanations for final model decisions. With extensive experiments, we empirically show VRX can meaningfully answer "why" and "why not" questions about the prediction, providing easy-to-understand insights about the reasoning process. We also show that these insights can potentially provide guidance on improving NN's performance.

Amanda Rios, Laurent Itti

Supervised deep neural networks are known to undergo a sharp decline in the accuracy of older tasks when new tasks are learned, termed "catastrophic forgetting". Many state-of-the-art solutions to continual learning rely on biasing and/or partitioning a model to accommodate successive tasks incrementally. However, these methods largely depend on the availability of a task-oracle to confer task identities to each test sample, without which the models are entirely unable to perform. To address this shortcoming, we propose and compare several candidate task-assigning mappers which require very little memory overhead: (1) Incremental unsupervised prototype assignment using either nearest means, Gaussian Mixture Models or fuzzy ART backbones; (2) Supervised incremental prototype assignment with fast fuzzy ARTMAP; (3) Shallow perceptron trained via a dynamic coreset. Our proposed model variants are trained either from pre-trained feature extractors or task-dependent feature embeddings of the main classifier network. We apply these pipeline variants to continual learning benchmarks, comprised of either sequences of several datasets or within one single dataset. Overall, these methods, despite their simplicity and compactness, perform very close to a ground truth oracle, especially in experiments of inter-dataset task assignment. Moreover, best-performing variants only impose an average cost of 1.7% parameter memory increase.

Sumedh A. Sontakke, Arash Mehrjou, Laurent Itti et al.

Animals exhibit an innate ability to learn regularities of the world through interaction. By performing experiments in their environment, they are able to discern the causal factors of variation and infer how they affect the world's dynamics. Inspired by this, we attempt to equip reinforcement learning agents with the ability to perform experiments that facilitate a categorization of the rolled-out trajectories, and to subsequently infer the causal factors of the environment in a hierarchical manner. We introduce {\em causal curiosity}, a novel intrinsic reward, and show that it allows our agents to learn optimal sequences of actions and discover causal factors in the dynamics of the environment. The learned behavior allows the agents to infer a binary quantized representation for the ground-truth causal factors in every environment. Additionally, we find that these experimental behaviors are semantically meaningful (e.g., our agents learn to lift blocks to categorize them by weight), and are learnt in a self-supervised manner with approximately 2.5 times less data than conventional supervised planners. We show that these behaviors can be re-purposed and fine-tuned (e.g., from lifting to pushing or other downstream tasks). Finally, we show that the knowledge of causal factor representations aids zero-shot learning for more complex tasks. Visit https://sites.google.com/usc.edu/causal-curiosity/home for website.

Sumegh Roychowdhury, Sumedh A. Sontakke, Nikaash Puri et al.

Temporal event representations are an essential aspect of learning among humans. They allow for succinct encoding of the experiences we have through a variety of sensory inputs. Also, they are believed to be arranged hierarchically, allowing for an efficient representation of complex long-horizon experiences. Additionally, these representations are acquired in a self-supervised manner. Analogously, here we propose a model that learns temporal representations from long-horizon visual demonstration data and associated textual descriptions, without explicit temporal supervision. Our method produces a hierarchy of representations that align more closely with ground-truth human-annotated events (+15.3) than state-of-the-art unsupervised baselines. Our results are comparable to heavily-supervised baselines in complex visual domains such as Chess Openings, YouCook2 and TutorialVQA datasets. Finally, we perform ablation studies illustrating the robustness of our approach. We release our code and demo visualizations in the Supplementary Material.

Shixian Wen, Amanda Rios, Laurent Itti

Deep neural networks can be fooled by adversarial attacks: adding carefully computed small adversarial perturbations to clean inputs can cause misclassification on state-of-the-art machine learning models. The reason is that neural networks fail to accommodate the distribution drift of the input data caused by adversarial perturbations. Here, we present a new solution - Beneficial Perturbation Network (BPN) - to defend against adversarial attacks by fixing the distribution drift. During training, BPN generates and leverages beneficial perturbations (somewhat opposite to well-known adversarial perturbations) by adding new, out-of-network biasing units. Biasing units influence the parameter space of the network, to preempt and neutralize future adversarial perturbations on input data samples. To achieve this, BPN creates reverse adversarial attacks during training, with very little cost, by recycling the training gradients already computed. Reverse attacks are captured by the biasing units, and the biases can in turn effectively defend against future adversarial examples. Reverse attacks are a shortcut, i.e., they affect the network's parameters without requiring instantiation of adversarial examples that could assist training. We provide comprehensive empirical evidence showing that 1) BPN is robust to adversarial examples and is much more running memory and computationally efficient compared to classical adversarial training. 2) BPN can defend against adversarial examples with negligible additional computation and parameter costs compared to training only on clean examples; 3) BPN hurts the accuracy on clean examples much less than classic adversarial training; 4) BPN can improve the generalization of the network 5) BPN trained only with Fast Gradient Sign Attack can generalize to defend PGD attacks.

Shixian Wen, Amanda Rios, Yunhao Ge et al.

The human brain is the gold standard of adaptive learning. It not only can learn and benefit from experience, but also can adapt to new situations. In contrast, deep neural networks only learn one sophisticated but fixed mapping from inputs to outputs. This limits their applicability to more dynamic situations, where input to output mapping may change with different contexts. A salient example is continual learning - learning new independent tasks sequentially without forgetting previous tasks. Continual learning of multiple tasks in artificial neural networks using gradient descent leads to catastrophic forgetting, whereby a previously learned mapping of an old task is erased when learning new mappings for new tasks. Here, we propose a new biologically plausible type of deep neural network with extra, out-of-network, task-dependent biasing units to accommodate these dynamic situations. This allows, for the first time, a single network to learn potentially unlimited parallel input to output mappings, and to switch on the fly between them at runtime. Biasing units are programmed by leveraging beneficial perturbations (opposite to well-known adversarial perturbations) for each task. Beneficial perturbations for a given task bias the network toward that task, essentially switching the network into a different mode to process that task. This largely eliminates catastrophic interference between tasks. Our approach is memory-efficient and parameter-efficient, can accommodate many tasks, and achieves state-of-the-art performance across different tasks and domains.

Kiran Lekkala, Laurent Itti

We develop new algorithms for simultaneous learning of multiple tasks (e.g., image classification, depth estimation), and for adapting to unseen task/domain distributions within those high-level tasks (e.g., different environments). First, we learn common representations underlying all tasks. We then propose an attention mechanism to dynamically specialize the network, at runtime, for each task. Our approach is based on weighting each feature map of the backbone network, based on its relevance to a particular task. To achieve this, we enable the attention module to learn task representations during training, which are used to obtain attention weights. Our method improves performance on new, previously unseen environments, and is 1.5x faster than standard existing meta learning methods using similar architectures. We highlight performance improvements for Multi-Task Meta Learning of 4 tasks (image classification, depth, vanishing point, and surface normal estimation), each over 10 to 25 test domains/environments, a result that could not be achieved with standard meta learning techniques like MAML.

Yunhao Ge, Jiaping Zhao, Laurent Itti

Object pose increases intraclass object variance which makes object recognition from 2D images harder. To render a classifier robust to pose variations, most deep neural networks try to eliminate the influence of pose by using large datasets with many poses for each class. Here, we propose a different approach: a class-agnostic object pose transformation network (OPT-Net) can transform an image along 3D yaw and pitch axes to synthesize additional poses continuously. Synthesized images lead to better training of an object classifier. We design a novel eliminate-add structure to explicitly disentangle pose from object identity: first eliminate pose information of the input image and then add target pose information (regularized as continuous variables) to synthesize any target pose. We trained OPT-Net on images of toy vehicles shot on a turntable from the iLab-20M dataset. After training on unbalanced discrete poses (5 classes with 6 poses per object instance, plus 5 classes with only 2 poses), we show that OPT-Net can synthesize balanced continuous new poses along yaw and pitch axes with high quality. Training a ResNet-18 classifier with original plus synthesized poses improves mAP accuracy by 9% overtraining on original poses only. Further, the pre-trained OPT-Net can generalize to new object classes, which we demonstrate on both iLab-20M and RGB-D. We also show that the learned features can generalize to ImageNet.

Kiran Lekkala, Sami Abu-El-Haija, Laurent Itti

Imitation learning has gained immense popularity because of its high sample-efficiency. However, in real-world scenarios, where the trajectory distribution of most of the tasks dynamically shifts, model fitting on continuously aggregated data alone would be futile. In some cases, the distribution shifts, so much, that it is difficult for an agent to infer the new task. We propose a novel algorithm to generalize on any related task by leveraging prior knowledge on a set of specific tasks, which involves assigning importance weights to each past demonstration. We show experiments where the robot is trained from a diversity of environmental tasks and is also able to adapt to an unseen environment, using few-shot learning. We also developed a prototype robot system to test our approach on the task of visual navigation, and experimental results obtained were able to confirm these suppositions.

Shixian Wen, Laurent Itti

Adversarial training, in which a network is trained on both adversarial and clean examples, is one of the most trusted defense methods against adversarial attacks. However, there are three major practical difficulties in implementing and deploying this method - expensive in terms of extra memory and computation costs; accuracy trade-off between clean and adversarial examples; and lack of diversity of adversarial perturbations. Classical adversarial training uses fixed, precomputed perturbations in adversarial examples (input space). In contrast, we introduce dynamic adversarial perturbations into the parameter space of the network, by adding perturbation biases to the fully connected layers of deep convolutional neural network. During training, using only clean images, the perturbation biases are updated in the Fast Gradient Sign Direction to automatically create and store adversarial perturbations by recycling the gradient information computed. The network learns and adjusts itself automatically to these learned adversarial perturbations. Thus, we can achieve adversarial training with negligible cost compared to requiring a training set of adversarial example images. In addition, if combined with classical adversarial training, our perturbation biases can alleviate accuracy trade-off difficulties, and diversify adversarial perturbations.

Amy Zhang, Zachary C. Lipton, Luis Pineda et al.

Intelligent agents can cope with sensory-rich environments by learning task-agnostic state abstractions. In this paper, we propose an algorithm to approximate causal states, which are the coarsest partition of the joint history of actions and observations in partially-observable Markov decision processes (POMDP). Our method learns approximate causal state representations from RNNs trained to predict subsequent observations given the history. We demonstrate that these learned state representations are useful for learning policies efficiently in reinforcement learning problems with rich observation spaces. We connect causal states with causal feature sets from the causal inference literature, and also provide theoretical guarantees on the optimality of the continuous version of this causal state representation under Lipschitz assumptions by proving equivalence to bisimulation, a relation between behaviorally equivalent systems. This allows for lower bounds on the optimal value function of the learned representation, which is tight given certain assumptions. Finally, we empirically evaluate causal state representations using multiple partially observable tasks and compare with prior methods.

Shixian Wen, Laurent Itti

Sequential learning of multiple tasks in artificial neural networks using gradient descent leads to catastrophic forgetting, whereby previously learned knowledge is erased during learning of new, disjoint knowledge. Here, we propose a fundamentally new type of method - Beneficial Perturbation Network (BPN). We add task-dependent memory (biasing) units to allow the network to operate in different regimes for different tasks. We compute the most beneficial directions to train these units, in a manner inspired by recent work on adversarial examples. At test time, beneficial perturbations for a given task bias the network toward that task to overcome catastrophic forgetting. BPN is not only more parameter-efficient than network expansion methods, but also does not need to store any data from previous tasks, in contrast with episodic memory methods. Experiments on variants of the MNIST, CIFAR-10, CIFAR-100 datasets demonstrate strong performance of BPN when compared to the state-of-the-art.

Amanda Rios, Laurent Itti

Sequential learning of tasks using gradient descent leads to an unremitting decline in the accuracy of tasks for which training data is no longer available, termed catastrophic forgetting. Generative models have been explored as a means to approximate the distribution of old tasks and bypass storage of real data. Here we propose a cumulative closed-loop memory replay GAN (CloGAN) provided with external regularization by a small memory unit selected for maximum sample diversity. We evaluate incremental class learning using a notoriously hard paradigm, single-headed learning, in which each task is a disjoint subset of classes in the overall dataset, and performance is evaluated on all previous classes. First, we show that when constructing a dynamic memory unit to preserve sample heterogeneity, model performance asymptotically approaches training on the full dataset. We then show that using a stochastic generator to continuously output fresh new images during training increases performance significantly further meanwhile generating quality images. We compare our approach to several baselines including fine-tuning by gradient descent (FGD), Elastic Weight Consolidation (EWC), Deep Generative Replay (DGR) and Memory Replay GAN (MeRGAN). Our method has very low long-term memory cost, the memory unit, as well as negligible intermediate memory storage.

Shixian Wen, Laurent Itti

Sequential learning of multiple tasks in artificial neural networks using gradient descent leads to catastrophic forgetting, whereby previously learned knowledge is erased during learning of new, disjoint knowledge. Here, we propose a new approach to sequential learning which leverages the recent discovery of adversarial examples. We use adversarial subspaces from previous tasks to enable learning of new tasks with less interference. We apply our method to sequentially learning to classify digits 0, 1, 2 (task 1), 4, 5, 6, (task 2), and 7, 8, 9 (task 3) in MNIST (disjoint MNIST task). We compare and combine our Adversarial Direction (AD) method with the recently proposed Elastic Weight Consolidation (EWC) method for sequential learning. We train each task for 20 epochs, which yields good initial performance (99.24% correct task 1 performance). After training task 2, and then task 3, both plain gradient descent (PGD) and EWC largely forget task 1 (task 1 accuracy 32.95% for PGD and 41.02% for EWC), while our combined approach (AD+EWC) still achieves 94.53% correct on task 1. We obtain similar results with a much more difficult disjoint CIFAR10 task, which to our knowledge had not been attempted before (70.10% initial task 1 performance, 67.73% after learning tasks 2 and 3 for AD+EWC, while PGD and EWC both fall to chance level). Our results suggest that AD+EWC can provide better sequential learning performance than either PGD or EWC.

Tommaso Furlanello, Zachary C. Lipton, Michael Tschannen et al.

Knowledge Distillation (KD) consists of transferring “knowledge” from one machine learning model (the teacher) to another (the student). Commonly, the teacher is a high-capacity model with formidable performance, while the student is more compact. By transferring knowledge, one hopes to benefit from the student’s compactness, without sacrificing too much performance. We study KD from a new perspective: rather than compressing models, we train students parameterized identically to their teachers. Surprisingly, these Born-Again Networks (BANs), outperform their teachers significantly, both on computer vision and language modeling tasks. Our experiments with BANs based on DenseNets demonstrate state-of-the-art performance on the CIFAR-10 (3.5%) and CIFAR-100 (15.5%) datasets, by validation error. Additional experiments explore two distillation objectives: (i) Confidence-Weighted by Teacher Max (CWTM) and (ii) Dark Knowledge with Permuted Predictions (DKPP). Both methods elucidate the essential components of KD, demonstrating the effect of the teacher outputs on both predicted and non-predicted classes.

Jiaping Zhao, Chin-kai Chang, Laurent Itti

Natural images are generated under many factors, including shape, pose, illumination etc. Most existing ConvNets formulate object recognition from natural images as a single task classification problem, and attempt to learn features useful for object categories, but invariant to other factors of variation as much as possible. These architectures do not explicitly learn other factors, like pose and lighting, instead, they usually discard them by pooling and normalization. In this work, we take the opposite approach: we train ConvNets for object recognition by retaining other factors (pose in our case) and learn them jointly with object category. We design a new multi-task leaning (MTL) ConvNet, named disentangling CNN (disCNN), which explicitly enforces the disentangled representations of object identity and pose, and is trained to predict object categories and pose transformations. We show that disCNN achieves significantly better object recognition accuracies than AlexNet trained solely to predict object categories on the iLab-20M dataset, which is a large scale turntable dataset with detailed object pose and lighting information. We further show that the pretrained disCNN/AlexNet features on iLab- 20M generalize to object recognition on both Washington RGB-D and ImageNet datasets, and the pretrained disCNN features are significantly better than the pretrained AlexNet features for fine-tuning object recognition on the ImageNet dataset.