Bruno A. Olshausen

Anthony Zador, Sean Escola, Blake Richards et al. · stanford

Neuroscience has long been an essential driver of progress in artificial intelligence (AI). We propose that to accelerate progress in AI, we must invest in fundamental research in NeuroAI. A core component of this is the embodied Turing test, which challenges AI animal models to interact with the sensorimotor world at skill levels akin to their living counterparts. The embodied Turing test shifts the focus from those capabilities like game playing and language that are especially well-developed or uniquely human to those capabilities, inherited from over 500 million years of evolution, that are shared with all animals. Building models that can pass the embodied Turing test will provide a roadmap for the next generation of AI.

Alpha Renner, Lazar Supic, Andreea Danielescu et al. · eth-zurich

Analyzing a visual scene by inferring the configuration of a generative model is widely considered the most flexible and generalizable approach to scene understanding. Yet, one major problem is the computational challenge of the inference procedure, involving a combinatorial search across object identities and poses. Here we propose a neuromorphic solution exploiting three key concepts: (1) a computational framework based on Vector Symbolic Architectures (VSA) with complex-valued vectors; (2) the design of Hierarchical Resonator Networks (HRN) to factorize the non-commutative transforms translation and rotation in visual scenes; (3) the design of a multi-compartment spiking phasor neuron model for implementing complex-valued resonator networks on neuromorphic hardware. The VSA framework uses vector binding operations to form a generative image model in which binding acts as the equivariant operation for geometric transformations. A scene can, therefore, be described as a sum of vector products, which can then be efficiently factorized by a resonator network to infer objects and their poses. The HRN features a partitioned architecture in which vector binding is equivariant for horizontal and vertical translation within one partition and for rotation and scaling within the other partition. The spiking neuron model allows mapping the resonator network onto efficient and low-power neuromorphic hardware. Our approach is demonstrated on synthetic scenes composed of simple 2D shapes undergoing rigid geometric transformations and color changes. A companion paper demonstrates the same approach in real-world application scenarios for machine vision and robotics.

Yubei Chen, Zeyu Yun, Yi Ma et al.

We describe a minimalistic and interpretable method for unsupervised learning, without resorting to data augmentation, hyperparameter tuning, or other engineering designs, that achieves performance close to the SOTA SSL methods. Our approach leverages the sparse manifold transform, which unifies sparse coding, manifold learning, and slow feature analysis. With a one-layer deterministic sparse manifold transform, one can achieve 99.3% KNN top-1 accuracy on MNIST, 81.1% KNN top-1 accuracy on CIFAR-10 and 53.2% on CIFAR-100. With a simple gray-scale augmentation, the model gets 83.2% KNN top-1 accuracy on CIFAR-10 and 57% on CIFAR-100. These results significantly close the gap between simplistic "white-box" methods and the SOTA methods. Additionally, we provide visualization to explain how an unsupervised representation transform is formed. The proposed method is closely connected to latent-embedding self-supervised methods and can be treated as the simplest form of VICReg. Though there remains a small performance gap between our simple constructive model and SOTA methods, the evidence points to this as a promising direction for achieving a principled and white-box approach to unsupervised learning.

Sophia Sanborn, Christian Shewmake, Bruno Olshausen et al.

We present a neural network architecture, Bispectral Neural Networks (BNNs) for learning representations that are invariant to the actions of compact commutative groups on the space over which a signal is defined. The model incorporates the ansatz of the bispectrum, an analytically defined group invariant that is complete -- that is, it preserves all signal structure while removing only the variation due to group actions. Here, we demonstrate that BNNs are able to simultaneously learn groups, their irreducible representations, and corresponding equivariant and complete-invariant maps purely from the symmetries implicit in data. Further, we demonstrate that the completeness property endows these networks with strong invariance-based adversarial robustness. This work establishes Bispectral Neural Networks as a powerful computational primitive for robust invariant representation learning

Ping-Chen Huang, Denis Kleyko, Jan M. Rabaey et al.

We introduce a method to identify speakers by computing with high-dimensional random vectors. Its strengths are simplicity and speed. With only 1.02k active parameters and a 128-minute pass through the training data we achieve Top-1 and Top-5 scores of 31% and 52% on the VoxCeleb1 dataset of 1,251 speakers. This is in contrast to CNN models requiring several million parameters and orders of magnitude higher computational complexity for only a 2$\times$ gain in discriminative power as measured in mutual information. An additional 92 seconds of training with Generalized Learning Vector Quantization (GLVQ) raises the scores to 48% and 67%. A trained classifier classifies 1 second of speech in 5.7 ms. All processing was done on standard CPU-based machines.

E. Paxon Frady, Spencer Kent, Quinn Tran et al.

Complex visual scenes that are composed of multiple objects, each with attributes, such as object name, location, pose, color, etc., are challenging to describe in order to train neural networks. Usually,deep learning networks are trained supervised by categorical scene descriptions. The common categorical description of a scene contains the names of individual objects but lacks information about other attributes. Here, we use distributed representations of object attributes and vector operations in a vector symbolic architecture to create a full compositional description of a scene in a high-dimensional vector. To control the scene composition, we use artificial images composed of multiple, translated and colored MNIST digits. In contrast to learning category labels, here we train deep neural networks to output the full compositional vector description of an input image. The output of the deep network can then be interpreted by a VSA resonator network, to extract object identity or other properties of indiviual objects. We evaluate the performance and generalization properties of the system on randomly generated scenes. Specifically, we show that the network is able to learn the task and generalize to unseen seen digit shapes and scene configurations. Further, the generalisation ability of the trained model is limited. For example, with a gap in the training data, like an object not shown in a particular image location during training, the learning does not automatically fill this gap.

Michael Y. -S. Fang, Mayur Mudigonda, Ryan Zarcone et al.

We describe a stochastic, dynamical system capable of inference and learning in a probabilistic latent variable model. The most challenging problem in such models - sampling the posterior distribution over latent variables - is proposed to be solved by harnessing natural sources of stochasticity inherent in electronic and neural systems. We demonstrate this idea for a sparse coding model by deriving a continuous-time equation for inferring its latent variables via Langevin dynamics. The model parameters are learned by simultaneously evolving according to another continuous-time equation, thus bypassing the need for digital accumulators or a global clock. Moreover we show that Langevin dynamics lead to an efficient procedure for sampling from the posterior distribution in the 'L0 sparse' regime, where latent variables are encouraged to be set to zero as opposed to having a small L1 norm. This allows the model to properly incorporate the notion of sparsity rather than having to resort to a relaxed version of sparsity to make optimization tractable. Simulations of the proposed dynamical system on both synthetic and natural image datasets demonstrate that the model is capable of probabilistically correct inference, enabling learning of the dictionary as well as parameters of the prior.

Marissa Connor, Bruno Olshausen, Christopher Rozell

When interacting in a three dimensional world, humans must estimate 3D structure from visual inputs projected down to two dimensional retinal images. It has been shown that humans use the persistence of object shape over motion-induced transformations as a cue to resolve depth ambiguity when solving this underconstrained problem. With the aim of understanding how biological vision systems may internally represent 3D transformations, we propose a computational model, based on a generative manifold model, which can be used to infer 3D structure from the motion of 2D points. Our model can also learn representations of the transformations with minimal supervision, providing a proof of concept for how humans may develop internal representations on a developmental or evolutionary time scale. Focused on rotational motion, we show how our model infers depth from moving 2D projected points, learns 3D rotational transformations from 2D training stimuli, and compares to human performance on psychophysical structure-from-motion experiments.

Christopher J. Kymn, Denis Kleyko, E. Paxon Frady et al.

We introduce Residue Hyperdimensional Computing, a computing framework that unifies residue number systems with an algebra defined over random, high-dimensional vectors. We show how residue numbers can be represented as high-dimensional vectors in a manner that allows algebraic operations to be performed with component-wise, parallelizable operations on the vector elements. The resulting framework, when combined with an efficient method for factorizing high-dimensional vectors, can represent and operate on numerical values over a large dynamic range using vastly fewer resources than previous methods, and it exhibits impressive robustness to noise. We demonstrate the potential for this framework to solve computationally difficult problems in visual perception and combinatorial optimization, showing improvement over baseline methods. More broadly, the framework provides a possible account for the computational operations of grid cells in the brain, and it suggests new machine learning architectures for representing and manipulating numerical data.

Zeyu Yun, Yubei Chen, Bruno A Olshausen et al.

Transformer networks have revolutionized NLP representation learning since they were introduced. Though a great effort has been made to explain the representation in transformers, it is widely recognized that our understanding is not sufficient. One important reason is that there lack enough visualization tools for detailed analysis. In this paper, we propose to use dictionary learning to open up these "black boxes" as linear superpositions of transformer factors. Through visualization, we demonstrate the hierarchical semantic structures captured by the transformer factors, e.g., word-level polysemy disambiguation, sentence-level pattern formation, and long-range dependency. While some of these patterns confirm the conventional prior linguistic knowledge, the rest are relatively unexpected, which may provide new insights. We hope this visualization tool can bring further knowledge and a better understanding of how transformer networks work. The code is available at https://github.com/zeyuyun1/TransformerVis

Shariq Mobin, Brian Cheung, Bruno Olshausen

Recent work has shown that recurrent neural networks can be trained to separate individual speakers in a sound mixture with high fidelity. Here we explore convolutional neural network models as an alternative and show that they achieve state-of-the-art results with an order of magnitude fewer parameters. We also characterize and compare the robustness and ability of these different approaches to generalize under three different test conditions: longer time sequences, the addition of intermittent noise, and different datasets not seen during training. For the last condition, we create a new dataset, RealTalkLibri, to test source separation in real-world environments. We show that the acoustics of the environment have significant impact on the structure of the waveform and the overall performance of neural network models, with the convolutional model showing superior ability to generalize to new environments. The code for our study is available at https://github.com/ShariqM/source_separation.

Christopher J. Kymn, Sonia Mazelet, Annabel Ng et al.

We propose a system for visual scene analysis and recognition based on encoding the sparse, latent feature-representation of an image into a high-dimensional vector that is subsequently factorized to parse scene content. The sparse feature representation is learned from image statistics via convolutional sparse coding, while scene parsing is performed by a resonator network. The integration of sparse coding with the resonator network increases the capacity of distributed representations and reduces collisions in the combinatorial search space during factorization. We find that for this problem the resonator network is capable of fast and accurate vector factorization, and we develop a confidence-based metric that assists in tracking the convergence of the resonator network.

E. Paxon Frady, Denis Kleyko, Christopher J. Kymn et al.

Vector space models for symbolic processing that encode symbols by random vectors have been proposed in cognitive science and connectionist communities under the names Vector Symbolic Architecture (VSA), and, synonymously, Hyperdimensional (HD) computing. In this paper, we generalize VSAs to function spaces by mapping continuous-valued data into a vector space such that the inner product between the representations of any two data points represents a similarity kernel. By analogy to VSA, we call this new function encoding and computing framework Vector Function Architecture (VFA). In VFAs, vectors can represent individual data points as well as elements of a function space (a reproducing kernel Hilbert space). The algebraic vector operations, inherited from VSA, correspond to well-defined operations in function space. Furthermore, we study a previously proposed method for encoding continuous data, fractional power encoding (FPE), which uses exponentiation of a random base vector to produce randomized representations of data points and fulfills the kernel properties for inducing a VFA. We show that the distribution from which elements of the base vector are sampled determines the shape of the FPE kernel, which in turn induces a VFA for computing with band-limited functions. In particular, VFAs provide an algebraic framework for implementing large-scale kernel machines with random features, extending Rahimi and Recht, 2007. Finally, we demonstrate several applications of VFA models to problems in image recognition, density estimation and nonlinear regression. Our analyses and results suggest that VFAs constitute a powerful new framework for representing and manipulating functions in distributed neural systems, with myriad applications in artificial intelligence.

Cameron Diao, Denis Kleyko, Jan M. Rabaey et al.

Machine learning algorithms deployed on edge devices must meet certain resource constraints and efficiency requirements. Random Vector Functional Link (RVFL) networks are favored for such applications due to their simple design and training efficiency. We propose a modified RVFL network that avoids computationally expensive matrix operations during training, thus expanding the network's range of potential applications. Our modification replaces the least-squares classifier with the Generalized Learning Vector Quantization (GLVQ) classifier, which only employs simple vector and distance calculations. The GLVQ classifier can also be considered an improvement upon certain classification algorithms popularly used in the area of Hyperdimensional Computing. The proposed approach achieved state-of-the-art accuracy on a collection of datasets from the UCI Machine Learning Repository - higher than previously proposed RVFL networks. We further demonstrate that our approach still achieves high accuracy while severely limited in training iterations (using on average only 21% of the least-squares classifier computational costs).

Denis Kleyko, Mike Davies, E. Paxon Frady et al.

This article reviews recent progress in the development of the computing framework vector symbolic architectures (VSA) (also known as hyperdimensional computing). This framework is well suited for implementation in stochastic, emerging hardware, and it naturally expresses the types of cognitive operations required for artificial intelligence (AI). We demonstrate in this article that the field-like algebraic structure of VSA offers simple but powerful operations on high-dimensional vectors that can support all data structures and manipulations relevant to modern computing. In addition, we illustrate the distinguishing feature of VSA, "computing in superposition," which sets it apart from conventional computing. It also opens the door to efficient solutions to the difficult combinatorial search problems inherent in AI applications. We sketch ways of demonstrating that VSA are computationally universal. We see them acting as a framework for computing with distributed representations that can play a role of an abstraction layer for emerging computing hardware. This article serves as a reference for computer architects by illustrating the philosophy behind VSA, techniques of distributed computing with them, and their relevance to emerging computing hardware, such as neuromorphic computing.

Ho Yin Chau, Frank Qiu, Yubei Chen et al.

Discrete spatial patterns and their continuous transformations are two important regularities contained in natural signals. Lie groups and representation theory are mathematical tools that have been used in previous works to model continuous image transformations. On the other hand, sparse coding is an important tool for learning dictionaries of patterns in natural signals. In this paper, we combine these ideas in a Bayesian generative model that learns to disentangle spatial patterns and their continuous transformations in a completely unsupervised manner. Images are modeled as a sparse superposition of shape components followed by a transformation that is parameterized by n continuous variables. The shape components and transformations are not predefined, but are instead adapted to learn the symmetries in the data, with the constraint that the transformations form a representation of an n-dimensional torus. Training the model on a dataset consisting of controlled geometric transformations of specific MNIST digits shows that it can recover these transformations along with the digits. Training on the full MNIST dataset shows that it can learn both the basic digit shapes and the natural transformations such as shearing and stretching that are contained in this data.

Hong-Ye Hu, Dian Wu, Yi-Zhuang You et al.

Flow-based generative models have become an important class of unsupervised learning approaches. In this work, we incorporate the key ideas of renormalization group (RG) and sparse prior distribution to design a hierarchical flow-based generative model, RG-Flow, which can separate information at different scales of images and extract disentangled representations at each scale. We demonstrate our method on synthetic multi-scale image datasets and the CelebA dataset, showing that the disentangled representations enable semantic manipulation and style mixing of the images at different scales. To visualize the latent representations, we introduce receptive fields for flow-based models and show that the receptive fields of RG-Flow are similar to those of convolutional neural networks. In addition, we replace the widely adopted isotropic Gaussian prior distribution by the sparse Laplacian distribution to further enhance the disentanglement of representations. From a theoretical perspective, our proposed method has $O(\log L)$ complexity for inpainting of an image with edge length $L$, compared to previous generative models with $O(L^2)$ complexity.

E. Paxon Frady, Spencer Kent, Bruno A. Olshausen et al.

The ability to encode and manipulate data structures with distributed neural representations could qualitatively enhance the capabilities of traditional neural networks by supporting rule-based symbolic reasoning, a central property of cognition. Here we show how this may be accomplished within the framework of Vector Symbolic Architectures (VSA) (Plate, 1991; Gayler, 1998; Kanerva, 1996), whereby data structures are encoded by combining high-dimensional vectors with operations that together form an algebra on the space of distributed representations. In particular, we propose an efficient solution to a hard combinatorial search problem that arises when decoding elements of a VSA data structure: the factorization of products of multiple code vectors. Our proposed algorithm, called a resonator network, is a new type of recurrent neural network that interleaves VSA multiplication operations and pattern completion. We show in two examples -- parsing of a tree-like data structure and parsing of a visual scene -- how the factorization problem arises and how the resonator network can solve it. More broadly, resonator networks open the possibility to apply VSAs to myriad artificial intelligence problems in real-world domains. A companion paper (Kent et al., 2020) presents a rigorous analysis and evaluation of the performance of resonator networks, showing it out-performs alternative approaches.

Dequan Wang, Evan Shelhamer, Shaoteng Liu et al.

A model must adapt itself to generalize to new and different data during testing. In this setting of fully test-time adaptation the model has only the test data and its own parameters. We propose to adapt by test entropy minimization (tent): we optimize the model for confidence as measured by the entropy of its predictions. Our method estimates normalization statistics and optimizes channel-wise affine transformations to update online on each batch. Tent reduces generalization error for image classification on corrupted ImageNet and CIFAR-10/100 and reaches a new state-of-the-art error on ImageNet-C. Tent handles source-free domain adaptation on digit recognition from SVHN to MNIST/MNIST-M/USPS, on semantic segmentation from GTA to Cityscapes, and on the VisDA-C benchmark. These results are achieved in one epoch of test-time optimization without altering training.

Juexiao Zhang, Yubei Chen, Brian Cheung et al.

Co-occurrence statistics based word embedding techniques have proved to be very useful in extracting the semantic and syntactic representation of words as low dimensional continuous vectors. In this work, we discovered that dictionary learning can open up these word vectors as a linear combination of more elementary word factors. We demonstrate many of the learned factors have surprisingly strong semantic or syntactic meaning corresponding to the factors previously identified manually by human inspection. Thus dictionary learning provides a powerful visualization tool for understanding word embedding representations. Furthermore, we show that the word factors can help in identifying key semantic and syntactic differences in word analogy tasks and improve upon the state-of-the-art word embedding techniques in these tasks by a large margin.

Dequan Wang, Evan Shelhamer, Bruno Olshausen et al.

Given the variety of the visual world there is not one true scale for recognition: objects may appear at drastically different sizes across the visual field. Rather than enumerate variations across filter channels or pyramid levels, dynamic models locally predict scale and adapt receptive fields accordingly. The degree of variation and diversity of inputs makes this a difficult task. Existing methods either learn a feedforward predictor, which is not itself totally immune to the scale variation it is meant to counter, or select scales by a fixed algorithm, which cannot learn from the given task and data. We extend dynamic scale inference from feedforward prediction to iterative optimization for further adaptivity. We propose a novel entropy minimization objective for inference and optimize over task and structure parameters to tune the model to each input. Optimization during inference improves semantic segmentation accuracy and generalizes better to extreme scale variations that cause feedforward dynamic inference to falter.

Spencer J. Kent, E. Paxon Frady, Friedrich T. Sommer et al.

We develop theoretical foundations of Resonator Networks, a new type of recurrent neural network introduced in Frady et al. (2020) to solve a high-dimensional vector factorization problem arising in Vector Symbolic Architectures. Given a composite vector formed by the Hadamard product between a discrete set of high-dimensional vectors, a Resonator Network can efficiently decompose the composite into these factors. We compare the performance of Resonator Networks against optimization-based methods, including Alternating Least Squares and several gradient-based algorithms, showing that Resonator Networks are superior in several important ways. This advantage is achieved by leveraging a combination of nonlinear dynamics and "searching in superposition," by which estimates of the correct solution are formed from a weighted superposition of all possible solutions. While the alternative methods also search in superposition, the dynamics of Resonator Networks allow them to strike a more effective balance between exploring the solution space and exploiting local information to drive the network toward probable solutions. Resonator Networks are not guaranteed to converge, but within a particular regime they almost always do. In exchange for relaxing this guarantee of global convergence, Resonator Networks are dramatically more effective at finding factorizations than all alternative approaches considered.

Shariq Mobin, Bruno Olshausen

We present a model for separating a set of voices out of a sound mixture containing an unknown number of sources. Our Attentional Gating Network (AGN) uses a variable attentional context to specify which speakers in the mixture are of interest. The attentional context is specified by an embedding vector which modifies the processing of a neural network through an additive bias. Individual speaker embeddings are learned to separate a single speaker while superpositions of the individual speaker embeddings are used to separate sets of speakers. We first evaluate AGN on a traditional single speaker separation task and show an improvement of 9% with respect to comparable models. Then, we introduce a new task to separate an arbitrary subset of voices from a mixture of an unknown-sized set of voices, inspired by the human ability to separate a conversation of interest from background chatter at a cafeteria. We show that AGN is the only model capable of solving this task, performing only 7% worse than on the single speaker separation task.

Brian Cheung, Alex Terekhov, Yubei Chen et al.

We present a method for storing multiple models within a single set of parameters. Models can coexist in superposition and still be retrieved individually. In experiments with neural networks, we show that a surprisingly large number of models can be effectively stored within a single parameter instance. Furthermore, each of these models can undergo thousands of training steps without significantly interfering with other models within the superposition. This approach may be viewed as the online complement of compression: rather than reducing the size of a network after training, we make use of the unrealized capacity of a network during training.

Yubei Chen, Dylan M. Paiton, Bruno A. Olshausen

We present a signal representation framework called the sparse manifold transform that combines key ideas from sparse coding, manifold learning, and slow feature analysis. It turns non-linear transformations in the primary sensory signal space into linear interpolations in a representational embedding space while maintaining approximate invertibility. The sparse manifold transform is an unsupervised and generative framework that explicitly and simultaneously models the sparse discreteness and low-dimensional manifold structure found in natural scenes. When stacked, it also models hierarchical composition. We provide a theoretical description of the transform and demonstrate properties of the learned representation on both synthetic data and natural videos.

Brian Cheung, Eric Weiss, Bruno Olshausen

We describe a neural attention model with a learnable retinal sampling lattice. The model is trained on a visual search task requiring the classification of an object embedded in a visual scene amidst background distractors using the smallest number of fixations. We explore the tiling properties that emerge in the model's retinal sampling lattice after training. Specifically, we show that this lattice resembles the eccentricity dependent sampling lattice of the primate retina, with a high resolution region in the fovea surrounded by a low resolution periphery. Furthermore, we find conditions where these emergent properties are amplified or eliminated providing clues to their function.

Alexander G. Anderson, Cory P. Berg, Daniel P. Mossing et al.

A recent paper by Gatys et al. describes a method for rendering an image in the style of another image. First, they use convolutional neural network features to build a statistical model for the style of an image. Then they create a new image with the content of one image but the style statistics of another image. Here, we extend this method to render a movie in a given artistic style. The naive solution that independently renders each frame produces poor results because the features of the style move substantially from one frame to the next. The other naive method that initializes the optimization for the next frame using the rendered version of the previous frame also produces poor results because the features of the texture stay fixed relative to the frame of the movie instead of moving with objects in the scene. The main contribution of this paper is to use optical flow to initialize the style transfer optimization so that the texture features move with the objects in the video. Finally, we suggest a method to incorporate optical flow explicitly into the cost function.

Brian Cheung, Jesse A. Livezey, Arjun K. Bansal et al.

Deep learning has enjoyed a great deal of success because of its ability to learn useful features for tasks such as classification. But there has been less exploration in learning the factors of variation apart from the classification signal. By augmenting autoencoders with simple regularization terms during training, we demonstrate that standard deep architectures can discover and explicitly represent factors of variation beyond those relevant for categorization. We introduce a cross-covariance penalty (XCov) as a method to disentangle factors like handwriting style for digits and subject identity in faces. We demonstrate this on the MNIST handwritten digit database, the Toronto Faces Database (TFD) and the Multi-PIE dataset by generating manipulated instances of the data. Furthermore, we demonstrate these deep networks can extrapolate `hidden' variation in the supervised signal.