Joseph Marino

Parmida Ghahremani, Joseph Marino, Ricardo Dodds et al.

In the clinic, resected tissue samples are stained with Hematoxylin-and-Eosin (H&E) and/or Immunhistochemistry (IHC) stains and presented to the pathologists on glass slides or as digital scans for diagnosis and assessment of disease progression. Cell-level quantification, e.g. in IHC protein expression scoring, can be extremely inefficient and subjective. We present DeepLIIF (https://deepliif.org), a first free online platform for efficient and reproducible IHC scoring. DeepLIIF outperforms current state-of-the-art approaches (relying on manual error-prone annotations) by virtually restaining clinical IHC slides with more informative multiplex immunofluorescence staining. Our DeepLIIF cloud-native platform supports (1) more than 150 proprietary/non-proprietary input formats via the Bio-Formats standard, (2) interactive adjustment, visualization, and downloading of the IHC quantification results and the accompanying restained images, (3) consumption of an exposed workflow API programmatically or through interactive plugins for open source whole slide image viewers such as QuPath/ImageJ, and (4) auto scaling to efficiently scale GPU resources based on user demand.

Alexandre Piche, Valentin Thomas, Joseph Marino et al. · mila

Bootstrapping is behind much of the successes of Deep Reinforcement Learning. However, learning the value function via bootstrapping often leads to unstable training due to fast-changing target values. Target Networks are employed to stabilize training by using an additional set of lagging parameters to estimate the target values. Despite the popularity of Target Networks, their effect on the optimization is still misunderstood. In this work, we show that they act as an implicit regularizer. This regularizer has disadvantages such as being inflexible and non convex. To overcome these issues, we propose an explicit Functional Regularization that is a convex regularizer in function space and can easily be tuned. We analyze the convergence of our method theoretically and empirically demonstrate that replacing Target Networks with the more theoretically grounded Functional Regularization approach leads to better sample efficiency and performance improvements.

SIMA team, Adrian Bolton, Alexander Lerchner et al.

We introduce SIMA 2, a generalist embodied agent that understands and acts in a wide variety of 3D virtual worlds. Built upon a Gemini foundation model, SIMA 2 represents a significant step toward active, goal-directed interaction within an embodied environment. Unlike prior work (e.g., SIMA 1) limited to simple language commands, SIMA 2 acts as an interactive partner, capable of reasoning about high-level goals, conversing with the user, and handling complex instructions given through language and images. Across a diverse portfolio of games, SIMA 2 substantially closes the gap with human performance and demonstrates robust generalization to previously unseen environments, all while retaining the base model's core reasoning capabilities. Furthermore, we demonstrate a capacity for open-ended self-improvement: by leveraging Gemini to generate tasks and provide rewards, SIMA 2 can autonomously learn new skills from scratch in a new environment. This work validates a path toward creating versatile and continuously learning agents for both virtual and, eventually, physical worlds.

Talat Zehra, Joseph Marino, Wendy Wang et al.

Histology slide digitization is becoming essential for telepathology (remote consultation), knowledge sharing (education), and using the state-of-the-art artificial intelligence algorithms (augmented/automated end-to-end clinical workflows). However, the cumulative costs of digital multi-slide high-speed brightfield scanners, cloud/on-premises storage, and personnel (IT and technicians) make the current slide digitization workflows out-of-reach for limited-resource settings, further widening the health equity gap; even single-slide manual scanning commercial solutions are costly due to hardware requirements (high-resolution cameras, high-spec PC/workstation, and support for only high-end microscopes). In this work, we present a new cloud slide digitization workflow for creating scanner-quality whole-slide images (WSIs) from uploaded low-quality videos, acquired from cheap and inexpensive microscopes with built-in cameras. Specifically, we present a pipeline to create stitched WSIs while automatically deblurring out-of-focus regions, upsampling input 10X images to 40X resolution, and reducing brightness/contrast and light-source illumination variations. We demonstrate the WSI creation efficacy from our workflow on World Health Organization-declared neglected tropical disease, Cutaneous Leishmaniasis (prevalent only in the poorest regions of the world and only diagnosed by sub-specialist dermatopathologists, rare in poor countries), as well as other common pathologies on core biopsies of breast, liver, duodenum, stomach and lymph node. The code and pretrained models will be accessible via our GitHub (https://github.com/nadeemlab/DeepLIIF), and the cloud platform will be available at https://deepliif.org for uploading microscope videos and downloading/viewing WSIs with shareable links (no sign-in required) for telepathology and knowledge sharing.

Parmida Ghahremani, Joseph Marino, Juan Hernandez-Prera et al.

We introduce a new AI-ready computational pathology dataset containing restained and co-registered digitized images from eight head-and-neck squamous cell carcinoma patients. Specifically, the same tumor sections were stained with the expensive multiplex immunofluorescence (mIF) assay first and then restained with cheaper multiplex immunohistochemistry (mIHC). This is a first public dataset that demonstrates the equivalence of these two staining methods which in turn allows several use cases; due to the equivalence, our cheaper mIHC staining protocol can offset the need for expensive mIF staining/scanning which requires highly-skilled lab technicians. As opposed to subjective and error-prone immune cell annotations from individual pathologists (disagreement > 50%) to drive SOTA deep learning approaches, this dataset provides objective immune and tumor cell annotations via mIF/mIHC restaining for more reproducible and accurate characterization of tumor immune microenvironment (e.g. for immunotherapy). We demonstrate the effectiveness of this dataset in three use cases: (1) IHC quantification of CD3/CD8 tumor-infiltrating lymphocytes via style transfer, (2) virtual translation of cheap mIHC stains to more expensive mIF stains, and (3) virtual tumor/immune cellular phenotyping on standard hematoxylin images. The dataset is available at \url{https://github.com/nadeemlab/DeepLIIF}.

SIMA Team, Maria Abi Raad, Arun Ahuja et al. · deepmind, stanford

Building embodied AI systems that can follow arbitrary language instructions in any 3D environment is a key challenge for creating general AI. Accomplishing this goal requires learning to ground language in perception and embodied actions, in order to accomplish complex tasks. The Scalable, Instructable, Multiworld Agent (SIMA) project tackles this by training agents to follow free-form instructions across a diverse range of virtual 3D environments, including curated research environments as well as open-ended, commercial video games. Our goal is to develop an instructable agent that can accomplish anything a human can do in any simulated 3D environment. Our approach focuses on language-driven generality while imposing minimal assumptions. Our agents interact with environments in real-time using a generic, human-like interface: the inputs are image observations and language instructions and the outputs are keyboard-and-mouse actions. This general approach is challenging, but it allows agents to ground language across many visually complex and semantically rich environments while also allowing us to readily run agents in new environments. In this paper we describe our motivation and goal, the initial progress we have made, and promising preliminary results on several diverse research environments and a variety of commercial video games.

Ping Hu, Saeed Boorboor, Joseph Marino et al.

The growing complexity of spatial and structural information in 3D data makes data inspection and visualization a challenging task. We describe a method to create a planar embedding of 3D treelike structures using their skeleton representations. Our method maintains the original geometry, without overlaps, to the best extent possible, allowing exploration of the topology within a single view. We present a novel camera view generation method which maximizes the visible geometric attributes (segment shape and relative placement between segments). Camera views are created for individual segments and are used to determine local bending angles at each node by projecting them to 2D. The final embedding is generated by minimizing an energy function (the weights of which are user adjustable) based on branch length and the 2D angles, while avoiding intersections. The user can also interactively modify segment placement within the 2D embedding, and the overall embedding will update accordingly. A global to local interactive exploration is provided using hierarchical camera views that are created for subtrees within the structure. We evaluate our method both qualitatively and quantitatively and demonstrate our results by constructing planar visualizations of line data (traced neurons) and volume data (CT vascular and bronchial data

Ruihan Yang, Yibo Yang, Joseph Marino et al.

While recent machine learning research has revealed connections between deep generative models such as VAEs and rate-distortion losses used in learned compression, most of this work has focused on images. In a similar spirit, we view recently proposed neural video coding algorithms through the lens of deep autoregressive and latent variable modeling. We present these codecs as instances of a generalized stochastic temporal autoregressive transform, and propose new avenues for further improvements inspired by normalizing flows and structured priors. We propose several architectures that yield state-of-the-art video compression performance on high-resolution video and discuss their tradeoffs and ablations. In particular, we propose (i) improved temporal autoregressive transforms, (ii) improved entropy models with structured and temporal dependencies, and (iii) variable bitrate versions of our algorithms. Since our improvements are compatible with a large class of existing models, we provide further evidence that the generative modeling viewpoint can advance the neural video coding field.

Alexandre Piché, Valentin Thomas, Rafael Pardinas et al.

Bootstrapping is behind much of the successes of deep Reinforcement Learning. However, learning the value function via bootstrapping often leads to unstable training due to fast-changing target values. Target Networks are employed to stabilize training by using an additional set of lagging parameters to estimate the target values. Despite the popularity of Target Networks, their effect on the optimization is still misunderstood. In this work, we show that they act as an implicit regularizer which can be beneficial in some cases, but also have disadvantages such as being inflexible and can result in instabilities, even when vanilla TD(0) converges. To overcome these issues, we propose an explicit Functional Regularization alternative that is flexible and a convex regularizer in function space and we theoretically study its convergence. We conduct an experimental study across a range of environments, discount factors, and off-policiness data collections to investigate the effectiveness of the regularization induced by Target Networks and Functional Regularization in terms of performance, accuracy, and stability. Our findings emphasize that Functional Regularization can be used as a drop-in replacement for Target Networks and result in performance improvement. Furthermore, adjusting both the regularization weight and the network update period in Functional Regularization can result in further performance improvements compared to solely adjusting the network update period as typically done with Target Networks. Our approach also enhances the ability to networks to recover accurate $Q$-values.

Joseph Marino

This paper reviews predictive coding, from theoretical neuroscience, and variational autoencoders, from machine learning, identifying the common origin and mathematical framework underlying both areas. As each area is prominent within its respective field, more firmly connecting these areas could prove useful in the dialogue between neuroscience and machine learning. After reviewing each area, we discuss two possible correspondences implied by this perspective: cortical pyramidal dendrites as analogous to (non-linear) deep networks and lateral inhibition as analogous to normalizing flows. These connections may provide new directions for further investigations in each field.

Joseph Marino, Alexandre Piché, Alessandro Davide Ialongo et al.

Policy networks are a central feature of deep reinforcement learning (RL) algorithms for continuous control, enabling the estimation and sampling of high-value actions. From the variational inference perspective on RL, policy networks, when used with entropy or KL regularization, are a form of \textit{amortized optimization}, optimizing network parameters rather than the policy distributions directly. However, \textit{direct} amortized mappings can yield suboptimal policy estimates and restricted distributions, limiting performance and exploration. Given this perspective, we consider the more flexible class of \textit{iterative} amortized optimizers. We demonstrate that the resulting technique, iterative amortized policy optimization, yields performance improvements over direct amortization on benchmark continuous control tasks.

Ruihan Yang, Yibo Yang, Joseph Marino et al.

Recent work by Marino et al. (2020) showed improved performance in sequential density estimation by combining masked autoregressive flows with hierarchical latent variable models. We draw a connection between such autoregressive generative models and the task of lossy video compression. Specifically, we view recent neural video compression methods (Lu et al., 2019; Yang et al., 2020b; Agustssonet al., 2020) as instances of a generalized stochastic temporal autoregressive transform, and propose avenues for enhancement based on this insight. Comprehensive evaluations on large-scale video data show improved rate-distortion performance over both state-of-the-art neural and conventional video compression methods.

Joseph Marino, Lei Chen, Jiawei He et al.

We propose an approach for improving sequence modeling based on autoregressive normalizing flows. Each autoregressive transform, acting across time, serves as a moving frame of reference, removing temporal correlations, and simplifying the modeling of higher-level dynamics. This technique provides a simple, general-purpose method for improving sequence modeling, with connections to existing and classical techniques. We demonstrate the proposed approach both with standalone flow-based models and as a component within sequential latent variable models. Results are presented on three benchmark video datasets, where autoregressive flow-based dynamics improve log-likelihood performance over baseline models. Finally, we illustrate the decorrelation and improved generalization properties of using flow-based dynamics.

Joseph Marino, Milan Cvitkovic, Yisong Yue

We introduce the variational filtering EM algorithm, a simple, general-purpose method for performing variational inference in dynamical latent variable models using information from only past and present variables, i.e. filtering. The algorithm is derived from the variational objective in the filtering setting and consists of an optimization procedure at each time step. By performing each inference optimization procedure with an iterative amortized inference model, we obtain a computationally efficient implementation of the algorithm, which we call amortized variational filtering. We present experiments demonstrating that this general-purpose method improves performance across several deep dynamical latent variable models.

Saad Nadeem, Joseph Marino, Xianfeng Gu et al.

We present a method for registration and visualization of corresponding supine and prone virtual colonoscopy scans based on eigenfunction analysis and fold modeling. In virtual colonoscopy, CT scans are acquired with the patient in two positions, and their registration is desirable so that physicians can corroborate findings between scans. Our algorithm performs this registration efficiently through the use of Fiedler vector representation (the second eigenfunction of the Laplace-Beltrami operator). This representation is employed to first perform global registration of the two colon positions. The registration is then locally refined using the haustral folds, which are automatically segmented using the 3D level sets of the Fiedler vector. The use of Fiedler vectors and the segmented folds presents a precise way of visualizing corresponding regions across datasets and visual modalities. We present multiple methods of visualizing the results, including 2D flattened rendering and the corresponding 3D endoluminal views. The precise fold modeling is used to automatically find a suitable cut for the 2D flattening, which provides a less distorted visualization. Our approach is robust, and we demonstrate its efficiency and efficacy by showing matched views on both the 2D flattened colons and in the 3D endoluminal view. We analytically evaluate the results by measuring the distance between features on the registered colons, and we also assess our fold segmentation against 20 manually labeled datasets. We have compared our results analytically to previous methods, and have found our method to achieve superior results. We also prove the hot spots conjecture for modeling cylindrical topology using Fiedler vector representation, which allows our approach to be used for general cylindrical geometry modeling and feature extraction.

Ji Hwan Park, Saad Nadeem, Joseph Marino et al.

Virtual colonoscopy (VC) allows a radiologist to navigate through a 3D colon model reconstructed from a computed tomography scan of the abdomen, looking for polyps, the precursors of colon cancer. Polyps are seen as protrusions on the colon wall and haustral folds, visible in the VC fly-through videos. A complete review of the colon surface requires full navigation from the rectum to the cecum in antegrade and retrograde directions, which is a tedious task that takes an average of 30 minutes. Crowdsourcing is a technique for non-expert users to perform certain tasks, such as image or video annotation. In this work, we use crowdsourcing for the examination of complete VC fly-through videos for polyp annotation by non-experts. The motivation for this is to potentially help the radiologist reach a diagnosis in a shorter period of time, and provide a stronger confirmation of the eventual diagnosis. The crowdsourcing interface includes an interactive tool for the crowd to annotate suspected polyps in the video with an enclosing box. Using our workflow, we achieve an overall polyps-per-patient sensitivity of 87.88% (95.65% for polyps $\geq$5mm and 70% for polyps $<$5mm). We also demonstrate the efficacy and effectiveness of a non-expert user in detecting and annotating polyps and discuss their possibility in aiding radiologists in VC examinations.

Joseph Marino, Yisong Yue, Stephan Mandt

Inference models are a key component in scaling variational inference to deep latent variable models, most notably as encoder networks in variational auto-encoders (VAEs). By replacing conventional optimization-based inference with a learned model, inference is amortized over data examples and therefore more computationally efficient. However, standard inference models are restricted to direct mappings from data to approximate posterior estimates. The failure of these models to reach fully optimized approximate posterior estimates results in an amortization gap. We aim toward closing this gap by proposing iterative inference models, which learn to perform inference optimization through repeatedly encoding gradients. Our approach generalizes standard inference models in VAEs and provides insight into several empirical findings, including top-down inference techniques. We demonstrate the inference optimization capabilities of iterative inference models and show that they outperform standard inference models on several benchmark data sets of images and text.

Jiawei He, Andreas Lehrmann, Joseph Marino et al.

Videos express highly structured spatio-temporal patterns of visual data. A video can be thought of as being governed by two factors: (i) temporally invariant (e.g., person identity), or slowly varying (e.g., activity), attribute-induced appearance, encoding the persistent content of each frame, and (ii) an inter-frame motion or scene dynamics (e.g., encoding evolution of the person ex-ecuting the action). Based on this intuition, we propose a generative framework for video generation and future prediction. The proposed framework generates a video (short clip) by decoding samples sequentially drawn from a latent space distribution into full video frames. Variational Autoencoders (VAEs) are used as a means of encoding/decoding frames into/from the latent space and RNN as a wayto model the dynamics in the latent space. We improve the video generation consistency through temporally-conditional sampling and quality by structuring the latent space with attribute controls; ensuring that attributes can be both inferred and conditioned on during learning/generation. As a result, given attributes and/orthe first frame, our model is able to generate diverse but highly consistent sets ofvideo sequences, accounting for the inherent uncertainty in the prediction task. Experimental results on Chair CAD, Weizmann Human Action, and MIT-Flickr datasets, along with detailed comparison to the state-of-the-art, verify effectiveness of the framework.

Ji Hwan Park, Seyedkoosha Mirhosseini, Saad Nadeem et al.

Virtual colonoscopy (VC) allows a physician to virtually navigate within a reconstructed 3D colon model searching for colorectal polyps. Though VC is widely recognized as a highly sensitive and specific test for identifying polyps, one limitation is the reading time, which can take over 30 minutes per patient. Large amounts of the colon are often devoid of polyps, and a way of identifying these polyp-free segments could be of valuable use in reducing the required reading time for the interrogating radiologist. To this end, we have tested the ability of the collective crowd intelligence of non-expert workers to identify polyp candidates and polyp-free regions. We presented twenty short videos flying through a segment of a virtual colon to each worker, and the crowd was asked to determine whether or not a possible polyp was observed within that video segment. We evaluated our framework on Amazon Mechanical Turk and found that the crowd was able to achieve a sensitivity of 80.0% and specificity of 86.5% in identifying video segments which contained a clinically proven polyp. Since each polyp appeared in multiple consecutive segments, all polyps were in fact identified. Using the crowd results as a first pass, 80% of the video segments could in theory be skipped by the radiologist, equating to a significant time savings and enabling more VC examinations to be performed.