Rajesh P. N. Rao

Sheng Zhang, Yanbo Xu, Naoto Usuyama et al. · cambridge, microsoft-research

Biomedical data is inherently multimodal, comprising physical measurements and natural language narratives. A generalist biomedical AI model needs to simultaneously process different modalities of data, including text and images. Therefore, training an effective generalist biomedical model requires high-quality multimodal data, such as parallel image-text pairs. Here, we present PMC-15M, a novel dataset that is two orders of magnitude larger than existing biomedical multimodal datasets such as MIMIC-CXR, and spans a diverse range of biomedical image types. PMC-15M contains 15 million biomedical image-text pairs collected from 4.4 million scientific articles. Based on PMC-15M, we have pretrained BiomedCLIP, a multimodal foundation model, with domain-specific adaptations tailored to biomedical vision-language processing. We conducted extensive experiments and ablation studies on standard biomedical imaging tasks from retrieval to classification to visual question-answering (VQA). BiomedCLIP achieved new state-of-the-art results in a wide range of standard datasets, substantially outperforming prior approaches. Intriguingly, by large-scale pretraining on diverse biomedical image types, BiomedCLIP even outperforms state-of-the-art radiology-specific models such as BioViL in radiology-specific tasks such as RSNA pneumonia detection. In summary, BiomedCLIP is a fully open-access foundation model that achieves state-of-the-art performance on various biomedical tasks, paving the way for transformative multimodal biomedical discovery and applications. We release our models at https://aka.ms/biomedclip to facilitate future research in multimodal biomedical AI.

Tommaso Salvatori, Ankur Mali, Christopher L. Buckley et al. · uw

Artificial intelligence (AI) is rapidly becoming one of the key technologies of this century. The majority of results in AI thus far have been achieved using deep neural networks trained with a learning algorithm called error backpropagation, always considered biologically implausible. To this end, recent works have studied learning algorithms for deep neural networks inspired by the neurosciences. One such theory, called predictive coding (PC), has shown promising properties that make it potentially valuable for the machine learning community: it can model information processing in different areas of the brain, can be used in control and robotics, has a solid mathematical foundation in variational inference, and performs its computations asynchronously. Inspired by such properties, works that propose novel PC-like algorithms are starting to be present in multiple sub-fields of machine learning and AI at large. Here, we survey such efforts by first providing a broad overview of the history of PC to provide common ground for the understanding of the recent developments, then by describing current efforts and results, and concluding with a large discussion of possible implications and ways forward.

Sam Preston, Mu Wei, Rajesh Rao et al. · cambridge, microsoft-research

Objective: The majority of detailed patient information in real-world data (RWD) is only consistently available in free-text clinical documents. Manual curation is expensive and time-consuming. Developing natural language processing (NLP) methods for structuring RWD is thus essential for scaling real-world evidence generation. Materials and Methods: Traditional rule-based systems are vulnerable to the prevalent linguistic variations and ambiguities in clinical text, and prior applications of machine-learning methods typically require sentence-level or report-level labeled examples that are hard to produce at scale. We propose leveraging patient-level supervision from medical registries, which are often readily available and capture key patient information, for general RWD applications. To combat the lack of sentence-level or report-level annotations, we explore advanced deep-learning methods by combining domain-specific pretraining, recurrent neural networks, and hierarchical attention. Results: We conduct an extensive study on 135,107 patients from the cancer registry of a large integrated delivery network (IDN) comprising healthcare systems in five western US states. Our deep learning methods attain test AUROC of 94-99% for key tumor attributes and comparable performance on held-out data from separate health systems and states. Discussion and Conclusion: Ablation results demonstrate clear superiority of these advanced deep-learning methods over prior approaches. Error analysis shows that our NLP system sometimes even corrects errors in registrar labels. We also conduct a preliminary investigation in accelerating registry curation and general RWD structuring via assisted curation for over 1.2 million cancer patients in this healthcare network.

Shirui Chen, Linxing Preston Jiang, Rajesh P. N. Rao et al. · uw

In sampling-based Bayesian models of brain function, neural activities are assumed to be samples from probability distributions that the brain uses for probabilistic computation. However, a comprehensive understanding of how mechanistic models of neural dynamics can sample from arbitrary distributions is still lacking. We use tools from functional analysis and stochastic differential equations to explore the minimum architectural requirements for $\textit{recurrent}$ neural circuits to sample from complex distributions. We first consider the traditional sampling model consisting of a network of neurons whose outputs directly represent the samples (sampler-only network). We argue that synaptic current and firing-rate dynamics in the traditional model have limited capacity to sample from a complex probability distribution. We show that the firing rate dynamics of a recurrent neural circuit with a separate set of output units can sample from an arbitrary probability distribution. We call such circuits reservoir-sampler networks (RSNs). We propose an efficient training procedure based on denoising score matching that finds recurrent and output weights such that the RSN implements Langevin sampling. We empirically demonstrate our model's ability to sample from several complex data distributions using the proposed neural dynamics and discuss its applicability to developing the next generation of sampling-based brain models.

Dimitrios C. Gklezakos, Rishi Jha, Rajesh P. N. Rao · uw

Inspired by Gibson's notion of object affordances in human vision, we ask the question: how can an agent learn to predict an entire action policy for a novel object or environment given only a single glimpse? To tackle this problem, we introduce the concept of Universal Policy Functions (UPFs) which are state-to-action mappings that generalize not only to new goals but most importantly to novel, unseen environments. Specifically, we consider the problem of efficiently learning such policies for agents with limited computational and communication capacity, constraints that are frequently encountered in edge devices. We propose the Hyper-Universal Policy Approximator (HUPA), a hypernetwork-based model to generate small task- and environment-conditional policy networks from a single image, with good generalization properties. Our results show that HUPAs significantly outperform an embedding-based alternative for generated policies that are size-constrained. Although this work is restricted to a simple map-based navigation task, future work includes applying the principles behind HUPAs to learning more general affordances for objects and environments.

Cailin Winston, Caleb Winston, Chloe N Winston et al. · uw

Brain-computer interfaces (BCIs) decode recorded neural signals from the brain and/or stimulate the brain with encoded neural signals. BCIs span both hardware and software and have a wide range of applications in restorative medicine, from restoring movement through prostheses and robotic limbs to restoring sensation and communication through spellers. BCIs also have applications in diagnostic medicine, e.g., providing clinicians with data for detecting seizures, sleep patterns, or emotions. Despite their promise, BCIs have not yet been adopted for long-term, day-to-day use because of challenges related to reliability and robustness, which are needed for safe operation in all scenarios. Ensuring safe operation currently requires hours of manual data collection and recalibration, involving both patients and clinicians. However, data collection is not targeted at eliminating specific faults in a BCI. This paper presents a new methodology for characterizing, detecting, and localizing faults in BCIs. Specifically, it proposes partial test oracles as a method for detecting faults and slice functions as a method for localizing faults to characteristic patterns in the input data or relevant tasks performed by the user. Through targeted data acquisition and retraining, the proposed methodology improves the correctness of BCIs. We evaluated the proposed methodology on five BCI applications. The results show that the proposed methodology (1) precisely localizes faults and (2) can significantly reduce the frequency of faults through retraining based on targeted, fault-based data acquisition. These results suggest that the proposed methodology is a promising step towards repairing faulty BCIs.

Rajesh P. N. Rao, Dimitrios C. Gklezakos, Vishwas Sathish

Predictive coding has emerged as a prominent model of how the brain learns through predictions, anticipating the importance accorded to predictive learning in recent AI architectures such as transformers. Here we propose a new framework for predictive coding called active predictive coding which can learn hierarchical world models and solve two radically different open problems in AI: (1) how do we learn compositional representations, e.g., part-whole hierarchies, for equivariant vision? and (2) how do we solve large-scale planning problems, which are hard for traditional reinforcement learning, by composing complex action sequences from primitive policies? Our approach exploits hypernetworks, self-supervised learning and reinforcement learning to learn hierarchical world models that combine task-invariant state transition networks and task-dependent policy networks at multiple abstraction levels. We demonstrate the viability of our approach on a variety of vision datasets (MNIST, FashionMNIST, Omniglot) as well as on a scalable hierarchical planning problem. Our results represent, to our knowledge, the first demonstration of a unified solution to the part-whole learning problem posed by Hinton, the nested reference frames problem posed by Hawkins, and the integrated state-action hierarchy learning problem in reinforcement learning.

Ares Fisher, Rajesh P. N. Rao

Human vision involves parsing and representing objects and scenes using structured representations based on part-whole hierarchies. Computer vision and machine learning researchers have recently sought to emulate this capability using capsule networks, reference frames and active predictive coding, but a generative model formulation has been lacking. We introduce Recursive Neural Programs (RNPs), which, to our knowledge, is the first neural generative model to address the part-whole hierarchy learning problem. RNPs model images as hierarchical trees of probabilistic sensory-motor programs that recursively reuse learned sensory-motor primitives to model an image within different reference frames, forming recursive image grammars. We express RNPs as structured variational autoencoders (sVAEs) for inference and sampling, and demonstrate parts-based parsing, sampling and one-shot transfer learning for MNIST, Omniglot and Fashion-MNIST datasets, demonstrating the model's expressive power. Our results show that RNPs provide an intuitive and explainable way of composing objects and scenes, allowing rich compositionality and intuitive interpretations of objects in terms of part-whole hierarchies.

Rajesh P. N. Rao, Vishwas Sathish, Linxing Preston Jiang et al.

The phenomenal advances in large language models (LLMs) and other foundation models over the past few years have been based on optimizing large-scale transformer models on the surprisingly simple objective of minimizing next-token prediction loss, a form of predictive coding that is also the backbone of an increasingly popular model of brain function in neuroscience and cognitive science. However, current foundation models ignore three other important components of state-of-the-art predictive coding models: tight integration of actions with generative models, hierarchical compositional structure, and episodic memory. We propose that to achieve safe, interpretable, energy-efficient, and human-like AI, foundation models should integrate actions, at multiple scales of abstraction, with a compositional generative architecture and episodic memory. We present recent evidence from neuroscience and cognitive science on the importance of each of these components. We describe how the addition of these missing components to foundation models could help address some of their current deficiencies: hallucinations and superficial understanding of concepts due to lack of grounding, a missing sense of agency/responsibility due to lack of control, threats to safety and trustworthiness due to lack of interpretability, and energy inefficiency. We compare our proposal to current trends, such as adding chain-of-thought (CoT) reasoning and retrieval-augmented generation (RAG) to foundation models, and discuss new ways of augmenting these models with brain-inspired components. We conclude by arguing that a rekindling of the historically fruitful exchange of ideas between brain science and AI will help pave the way towards safe and interpretable human-centered AI.

Nigini Oliveira, Jasmine Li, Koosha Khalvati et al.

Constructing a universal moral code for artificial intelligence (AI) is difficult or even impossible, given that different human cultures have different definitions of morality and different societal norms. We therefore argue that the value system of an AI should be culturally attuned: just as a child raised in a particular culture learns the specific values and norms of that culture, we propose that an AI agent operating in a particular human community should acquire that community's moral, ethical, and cultural codes. How AI systems might acquire such codes from human observation and interaction has remained an open question. Here, we propose using inverse reinforcement learning (IRL) as a method for AI agents to acquire a culturally-attuned value system implicitly. We test our approach using an experimental paradigm in which AI agents use IRL to learn different reward functions, which govern the agents' moral values, by observing the behavior of different cultural groups in an online virtual world requiring real-time decision making. We show that an AI agent learning from the average behavior of a particular cultural group can acquire altruistic characteristics reflective of that group's behavior, and this learned value system can generalize to new scenarios requiring altruistic judgments. Our results provide, to our knowledge, the first demonstration that AI agents could potentially be endowed with the ability to continually learn their values and norms from observing and interacting with humans, thereby becoming attuned to the culture they are operating in.

Dimitrios C. Gklezakos, Rajesh P. N. Rao

We introduce Active Predictive Coding Networks (APCNs), a new class of neural networks that solve a major problem posed by Hinton and others in the fields of artificial intelligence and brain modeling: how can neural networks learn intrinsic reference frames for objects and parse visual scenes into part-whole hierarchies by dynamically allocating nodes in a parse tree? APCNs address this problem by using a novel combination of ideas: (1) hypernetworks are used for dynamically generating recurrent neural networks that predict parts and their locations within intrinsic reference frames conditioned on higher object-level embedding vectors, and (2) reinforcement learning is used in conjunction with backpropagation for end-to-end learning of model parameters. The APCN architecture lends itself naturally to multi-level hierarchical learning and is closely related to predictive coding models of cortical function. Using the MNIST, Fashion-MNIST and Omniglot datasets, we demonstrate that APCNs can (a) learn to parse images into part-whole hierarchies, (b) learn compositional representations, and (c) transfer their knowledge to unseen classes of objects. With their ability to dynamically generate parse trees with part locations for objects, APCNs offer a new framework for explainable AI that leverages advances in deep learning while retaining interpretability and compositionality.

Satpreet Harcharan Singh, Floris van Breugel, Rajesh P. N. Rao et al.

Tracking a turbulent plume to locate its source is a complex control problem because it requires multi-sensory integration and must be robust to intermittent odors, changing wind direction, and variable plume statistics. This task is routinely performed by flying insects, often over long distances, in pursuit of food or mates. Several aspects of this remarkable behavior have been studied in detail in many experimental studies. Here, we take a complementary in silico approach, using artificial agents trained with reinforcement learning to develop an integrated understanding of the behaviors and neural computations that support plume tracking. Specifically, we use deep reinforcement learning (DRL) to train recurrent neural network (RNN) agents to locate the source of simulated turbulent plumes. Interestingly, the agents' emergent behaviors resemble those of flying insects, and the RNNs learn to represent task-relevant variables, such as head direction and time since last odor encounter. Our analyses suggest an intriguing experimentally testable hypothesis for tracking plumes in changing wind direction -- that agents follow local plume shape rather than the current wind direction. While reflexive short-memory behaviors are sufficient for tracking plumes in constant wind, longer timescales of memory are essential for tracking plumes that switch direction. At the level of neural dynamics, the RNNs' population activity is low-dimensional and organized into distinct dynamical structures, with some correspondence to behavioral modules. Our in silico approach provides key intuitions for turbulent plume tracking strategies and motivates future targeted experimental and theoretical developments.

Rajesh P. N. Rao

Brain-computer interfaces (BCIs) use decoding algorithms to control prosthetic devices based on brain signals for restoration of lost function. Computer-brain interfaces (CBIs), on the other hand, use encoding algorithms to transform external sensory signals into neural stimulation patterns for restoring sensation or providing sensory feedback for closed-loop prosthetic control. In this article, we introduce brain co-processors, devices that combine decoding and encoding in a unified framework using artificial intelligence (AI) to supplement or augment brain function. Brain co-processors can be used for a range of applications, from inducing Hebbian plasticity for rehabilitation after brain injury to reanimating paralyzed limbs and enhancing memory. A key challenge is simultaneous multi-channel neural decoding and encoding for optimization of external behavioral or task-related goals. We describe a new framework for developing brain co-processors based on artificial neural networks, deep learning and reinforcement learning. These "neural co-processors" allow joint optimization of cost functions with the nervous system to achieve desired behaviors. By coupling artificial neural networks with their biological counterparts, neural co-processors offer a new way of restoring and augmenting the brain, as well as a new scientific tool for brain research. We conclude by discussing the potential applications and ethical implications of brain co-processors.

Satpreet H. Singh, Steven M. Peterson, Rajesh P. N. Rao et al.

Recent technological advances in brain recording and artificial intelligence are propelling a new paradigm in neuroscience beyond the traditional controlled experiment. Rather than focusing on cued, repeated trials, naturalistic neuroscience studies neural processes underlying spontaneous behaviors performed in unconstrained settings. However, analyzing such unstructured data lacking a priori experimental design remains a significant challenge, especially when the data is multi-modal and long-term. Here we describe an automated approach for analyzing simultaneously recorded long-term, naturalistic electrocorticography (ECoG) and naturalistic behavior video data. We take a behavior-first approach to analyzing the long-term recordings. Using a combination of computer vision, discrete latent-variable modeling, and string pattern-matching on the behavioral video data, we find and annotate spontaneous human upper-limb movement events. We show results from our approach applied to data collected for 12 human subjects over 7--9 days for each subject. Our pipeline discovers and annotates over 40,000 instances of naturalistic human upper-limb movement events in the behavioral videos. Analysis of the simultaneously recorded brain data reveals neural signatures of movement that corroborate prior findings from traditional controlled experiments. We also prototype a decoder for a movement initiation detection task to demonstrate the efficacy of our pipeline as a source of training data for brain-computer interfacing applications. Our work addresses the unique data analysis challenges in studying naturalistic human behaviors, and contributes methods that may generalize to other neural recording modalities beyond ECoG. We publicly release our curated dataset, providing a resource to study naturalistic neural and behavioral variability at a scale not previously available.

Kameron Decker Harris, Aleksandr Aravkin, Rajesh Rao et al.

We present a windowed technique to learn parsimonious time-varying autoregressive models from multivariate timeseries. This unsupervised method uncovers interpretable spatiotemporal structure in data via non-smooth and non-convex optimization. In each time window, we assume the data follow a linear model parameterized by a system matrix, and we model this stack of potentially different system matrices as a low rank tensor. Because of its structure, the model is scalable to high-dimensional data and can easily incorporate priors such as smoothness over time. We find the components of the tensor using alternating minimization and prove that any stationary point of this algorithm is a local minimum. We demonstrate on a synthetic example that our method identifies the true rank of a switching linear system in the presence of noise. We illustrate our model's utility and superior scalability over extant methods when applied to several synthetic and real-world example: two types of time-varying linear systems, worm behavior, sea surface temperature, and monkey brain datasets.

Rajesh P. N. Rao

The Indus script remains one of the last major undeciphered scripts of the ancient world. We focus here on Indus inscriptions on a group of miniature tablets discovered by Meadow and Kenoyer in Harappa in 1997. By drawing parallels with proto-Elamite and proto-Cuneiform inscriptions, we explore how these miniature tablets may have been used to record rations allocated to porters or laborers. We then show that similar inscriptions are found on stamp seals, leading to the potentially provocative conclusion that rather than simply indicating ownership of property, Indus seals may have been used for generating tokens, tablets and sealings for repetitive economic transactions such as rations and exchange of canonical amounts of goods, grains, animals, and labor in a barter-based economy.

Rajesh P. N. Rao

The field of brain-computer interfaces is poised to advance from the traditional goal of controlling prosthetic devices using brain signals to combining neural decoding and encoding within a single neuroprosthetic device. Such a device acts as a "co-processor" for the brain, with applications ranging from inducing Hebbian plasticity for rehabilitation after brain injury to reanimating paralyzed limbs and enhancing memory. We review recent progress in simultaneous decoding and encoding for closed-loop control and plasticity induction. To address the challenge of multi-channel decoding and encoding, we introduce a unifying framework for developing brain co-processors based on artificial neural networks and deep learning. These "neural co-processors" can be used to jointly optimize cost functions with the nervous system to achieve desired behaviors ranging from targeted neuro-rehabilitation to augmentation of brain function.

Dimitrios C. Gklezakos, Rajesh P. N. Rao

A fundamental problem faced by object recognition systems is that objects and their features can appear in different locations, scales and orientations. Current deep learning methods attempt to achieve invariance to local translations via pooling, discarding the locations of features in the process. Other approaches explicitly learn transformed versions of the same feature, leading to representations that quickly explode in size. Instead of discarding the rich and useful information about feature transformations to achieve invariance, we argue that models should learn object features conjointly with their transformations to achieve equivariance. We propose a new model of unsupervised learning based on sparse coding that can learn object features jointly with their affine transformations directly from images. Results based on learning from natural images indicate that our approach matches the reconstruction quality of traditional sparse coding but with significantly fewer degrees of freedom while simultaneously learning transformations from data. These results open the door to scaling up unsupervised learning to allow deep feature+transformation learning in a manner consistent with the ventral+dorsal stream architecture of the primate visual cortex.

Kaiyu Zheng, Andrzej Pronobis, Rajesh P. N. Rao

We introduce Graph-Structured Sum-Product Networks (GraphSPNs), a probabilistic approach to structured prediction for problems where dependencies between latent variables are expressed in terms of arbitrary, dynamic graphs. While many approaches to structured prediction place strict constraints on the interactions between inferred variables, many real-world problems can be only characterized using complex graph structures of varying size, often contaminated with noise when obtained from real data. Here, we focus on one such problem in the domain of robotics. We demonstrate how GraphSPNs can be used to bolster inference about semantic, conceptual place descriptions using noisy topological relations discovered by a robot exploring large-scale office spaces. Through experiments, we show that GraphSPNs consistently outperform the traditional approach based on undirected graphical models, successfully disambiguating information in global semantic maps built from uncertain, noisy local evidence. We further exploit the probabilistic nature of the model to infer marginal distributions over semantic descriptions of as yet unexplored places and detect spatial environment configurations that are novel and incongruent with the known evidence.

Nancy Xin Ru Wang, Ali Farhadi, Rajesh Rao et al.

Developing useful interfaces between brains and machines is a grand challenge of neuroengineering. An effective interface has the capacity to not only interpret neural signals, but predict the intentions of the human to perform an action in the near future; prediction is made even more challenging outside well-controlled laboratory experiments. This paper describes our approach to detect and to predict natural human arm movements in the future, a key challenge in brain computer interfacing that has never before been attempted. We introduce the novel Annotated Joints in Long-term ECoG (AJILE) dataset; AJILE includes automatically annotated poses of 7 upper body joints for four human subjects over 670 total hours (more than 72 million frames), along with the corresponding simultaneously acquired intracranial neural recordings. The size and scope of AJILE greatly exceeds all previous datasets with movements and electrocorticography (ECoG), making it possible to take a deep learning approach to movement prediction. We propose a multimodal model that combines deep convolutional neural networks (CNN) with long short-term memory (LSTM) blocks, leveraging both ECoG and video modalities. We demonstrate that our models are able to detect movements and predict future movements up to 800 msec before movement initiation. Further, our multimodal movement prediction models exhibit resilience to simulated ablation of input neural signals. We believe a multimodal approach to natural neural decoding that takes context into account is critical in advancing bioelectronic technologies and human neuroscience.

Jing Wu, Kaitlyn Casimo, David J. Caldwell et al.

Identification of intended movement type and movement phase of hand grasp shaping are critical features for the control of volitional neuroprosthetics. We demonstrate that neural dynamics during visually-guided imagined grasp shaping can encode intended movement. We apply Procrustes analysis and LASSO regression to achieve 72% accuracy (chance = 25%) in distinguishing between visually-guided imagined grasp trajectories. Further, we can predict the stage of grasp shaping in the form of elapsed time from start of trial (R2=0.4). Our approach contributes to more accurate single-trial decoding of higher-level movement goals and the phase of grasping movements in individuals not trained with brain-computer interfaces. We also find that the overall time-varying trajectory structure of imagined movements tend to be consistent within individuals, and that transient trajectory deviations within trials return to the task-dependent trajectory mean. These overall findings may contribute to the further understanding of the cortical dynamics of human motor imagery.

Jing Wu, Benjamin R. Shuman, Bingni W. Brunton et al.

Neural correlates of movement planning onset and direction may be present in human electrocorticography in the signal dynamics of both motor and non-motor cortical regions. We use a three-stage model of jPCA reduced-rank hidden Markov model (jPCA-RR-HMM), regularized shrunken-centroid discriminant analysis (RDA), and LASSO regression to extract direction-sensitive planning information and movement onset in an upper-limb 3D isometric force task in a human subject. This mode achieves a relatively high true positive force-onset prediction rate of 60% within 250ms, and an above-chance 36% accuracy (17% chance) in predicting one of six planned 3D directions of isometric force using pre-movement signals. We also find direction-distinguishing information up to 400ms before force onset in the pre-movement signals, captured by electrodes placed over the limb-ipsilateral dorsal premotor regions. This approach can contribute to more accurate decoding of higher-level movement goals, at earlier timescales, and inform sensor placement. Our results also contribute to further understanding of the spatiotemporal features of human motor planning.