Skip to main content
eScholarship
Open Access Publications from the University of California

Scholarly Works (7 results)

  • Thesis
  • Peer Reviewed
UC Berkeley Electronic Theses and Dissertations (2018)

In the past decade, research in machine learning has been principally focused on the development of algorithms and models with high predictive capabilities. Models such as convolutional neural networks (CNNs) achieve state-of-the-art predictive performance for many tasks in computer vision, autonomous driving, and transfer learning. However, interpreting these models remains a challenge, primarily because of the large number of parameters involved.

In this thesis, we investigate two regimes based on (1) compression and (2) stability to build more interpretable machine learning models. These regimes will be demonstrated in a computational neuroscience study. In the first part of the thesis, we introduce a greedy structural compression scheme that prunes filters in a trained CNN. To do this, we define a filter importance index equal to the classification accuracy reduction (CAR) of the network after pruning that filter (similarly defined as RAR for regression). CAR achieves state-of-the-art classification accuracy compared to other filter pruning schemes. Furthermore, we demonstrate the interpretability of CAR-compressed CNNs by showing that our algorithm prunes filters with visually redundant functionalities such as color filters.

In the second part of this thesis, we introduce DeepTune, a stability-driven visu- alization and interpretation framework for CNN-based models. DeepTune is used to characterize biological neurons in the V4 area of the primate visual cortex. V4 is a large retinotopically-organized area of the visual cortex located between the primary visual cor- tex and high-level areas in the inferior temporal lobe. V4 neurons have highly nonlinear response properties and it is notoriously di cult to construct quantitative models that accurately describe how visual information is represented in V4. To better understand the filtering properties of these neurons, we study recordings from 71 well isolated cells stimulated with 4000-12000 static grayscale natural images collected by the Gallant Lab at UC Berkeley. Our CNN-based models of V4 neurons achieve state-of-the-art accuracy in predicting neural spike rates in a hold-out validation set (average predictive correlation of 0.53 for 71 neurons). Then, we employ our DeepTune stability-driven interpretation framework and discover that the V4 neurons are tuned to a remarkable diversity of tex- tures (40% of the neurons), contour shapes (30% of the neurons), and complex patterns (30% of the neurons). Most importantly, these smooth DeepTune images provide testable naturalistic stimuli for future experiments on V4 neurons.

In the final part of this thesis, we study the application of CAR and RAR compressed CNNs in modeling V4 neurons. Both CAR and RAR compression give rise to a new set of simpler models for V4 neurons with similar accuracy to existing state-of-the-art models. For each of the accurate CAR and RAR compressed models of V4 neurons (up to 90% compression rate), the smooth DeepTune images are stable and exhibit similar patterns to the uncompressed model’s consensus DeepTune image. Our results suggest, to some extent, that these CNNs resemble the structure of the primate brain.

    Cover page: Interpretable Machine Learning with Applications in Neuroscience
    • Article
    • Peer Reviewed
    UC Berkeley Previously Published Works (2021)
    Deep convolutional neural networks (CNNs) have been successful in many tasks in machine vision, however, millions of weights in the form of thousands of convolutional filters in CNNs make them difficult for human interpretation or understanding in science. In this article, we introduce a greedy structural compression scheme to obtain smaller and more interpretable CNNs, while achieving close to original accuracy. The compression is based on pruning filters with the least contribution to the classification accuracy or the lowest Classification Accuracy Reduction (CAR) importance index. We demonstrate the interpretability of CAR-compressed CNNs by showing that our algorithm prunes filters with visually redundant functionalities such as color filters. These compressed networks are easier to interpret because they retain the filter diversity of uncompressed networks with an order of magnitude fewer filters. Finally, a variant of CAR is introduced to quantify the importance of each image category to each CNN filter. Specifically, the most and the least important class labels are shown to be meaningful interpretations of each filter.
      Cover page: Structural Compression of Convolutional Neural Networks with Applications in Interpretability
      • Thesis
      • Peer Reviewed
      UC San Francisco Electronic Theses and Dissertations (2024)

      In Parkinson’s disease, imbalances between “antikinetic” and “prokinetic” patterns of neuronal oscillatory activity are related to motor dysfunction. Invasive brain recordings from the motor network have suggested that medical or surgical therapy can promote a prokinetic state by inducing narrowband gamma rhythms (60-90 Hz), which are associated with dyskinesia. Here, I investigate the behavioral and statistical properties of levodopa-induced and deep brain stimulation-entrained narrowband gamma rhythms. Collaborating with Oxford researchers, we also model the deep brain stimulation induced entrainment of gamma oscillations using interacting neuronal populations and patient-specific features of the levodopa-induced gamma oscillations. Using a sensing-enabled deep brain stimulation system, attached to both motor cortex and subthalamic or pallidal leads, the Starr Lab recorded over 900 hours of multisite field potentials prior to initiating deep brain stimulation, and over 600 hours with deep brain stimulation. I find that levodopa-induced gamma oscillations are more strongly associated with dyskinesia than deep brain stimulation-entrained gamma oscillations. Statistical comparisons revealed that levodopa-induced gamma oscillations exhibit increased variance in peak frequency, decreased spectral power, and higher variance in spectral power. Furthermore, I also work with collaborators at Oxford to show that entrainment can be predicted using neural circuit models fitted to patient data at various stimulation parameters. Put together, this work suggests that gamma oscillations as programming biomarkers should be leveraged distinctly across medical and surgical interventions to mitigate dyskinesia – while excessive levodopa-induced gamma oscillations are a marker for dyskinesia, increased entrained gamma oscillations may be a marker of a non-pathological prokinetic movement state. Our modeling work can subsequently be leveraged to predict the amount of entrained gamma across stimulation parameters.

        Cover page: Levodopa-induced gamma oscillations in patients with Parkinson’s disease using chronic invasive brain recording: detection, entrainment, and functional relevance
        • Article
        • Peer Reviewed
        UC Berkeley Previously Published Works (2016)
          • Article
          • Peer Reviewed
          UC Berkeley Previously Published Works (2019)
          Machine-learning models have demonstrated great success in learning complex patterns that enable them to make predictions about unobserved data. In addition to using models for prediction, the ability to interpret what a model has learned is receiving an increasing amount of attention. However, this increased focus has led to considerable confusion about the notion of interpretability. In particular, it is unclear how the wide array of proposed interpretation methods are related and what common concepts can be used to evaluate them. We aim to address these concerns by defining interpretability in the context of machine learning and introducing the predictive, descriptive, relevant (PDR) framework for discussing interpretations. The PDR framework provides 3 overarching desiderata for evaluation: predictive accuracy, descriptive accuracy, and relevancy, with relevancy judged relative to a human audience. Moreover, to help manage the deluge of interpretation methods, we introduce a categorization of existing techniques into model-based and post hoc categories, with subgroups including sparsity, modularity, and simulatability. To demonstrate how practitioners can use the PDR framework to evaluate and understand interpretations, we provide numerous real-world examples. These examples highlight the often underappreciated role played by human audiences in discussions of interpretability. Finally, based on our framework, we discuss limitations of existing methods and directions for future work. We hope that this work will provide a common vocabulary that will make it easier for both practitioners and researchers to discuss and choose from the full range of interpretation methods.
            Cover page: Definitions, methods, and applications in interpretable machine learning
            • Article
            • Peer Reviewed
            UC San Francisco Previously Published Works (2024)
            Quantification of motor symptom progression in Parkinsons disease (PD) patients is crucial for assessing disease progression and for optimizing therapeutic interventions, such as dopaminergic medications and deep brain stimulation. Cumulative and heuristic clinical experience has identified various clinical signs associated with PD severity, but these are neither objectively quantifiable nor robustly validated. Video-based objective symptom quantification enabled by machine learning (ML) introduces a potential solution. However, video-based diagnostic tools often have implementation challenges due to expensive and inaccessible technology, and typical black-box ML implementations are not tailored to be clinically interpretable. Here, we address these needs by releasing a comprehensive kinematic dataset and developing an interpretable video-based framework that predicts high versus low PD motor symptom severity according to MDS-UPDRS Part III metrics. This data driven approach validated and robustly quantified canonical movement features and identified new clinical insights, not previously appreciated as related to clinical severity, including pinkie finger movements and lower limb and axial features of gait. Our framework is enabled by retrospective, single-view, seconds-long videos recorded on consumer-grade devices such as smartphones, tablets, and digital cameras, thereby eliminating the requirement for specialized equipment. Following interpretable ML principles, our framework enforces robustness and interpretability by integrating (1) automatic, data-driven kinematic metric evaluation guided by pre-defined digital features of movement, (2) combination of bi-domain (body and hand) kinematic features, and (3) sparsity-inducing and stability-driven ML analysis with simple-to-interpret models. These elements ensure that the proposed framework quantifies clinically meaningful motor features useful for both ML predictions and clinical analysis.
              Cover page: Interpretable video-based tracking and quantification of parkinsonism clinical motor states.
              • Article
              • Peer Reviewed
              UC San Francisco Previously Published Works (2024)
              The rapid growth of large-scale spatial gene expression data demands efficient and reliable computational tools to extract major trends of gene expression in their native spatial context. Here, we used stability-driven unsupervised learning (i.e., staNMF) to identify principal patterns (PPs) of 3D gene expression profiles and understand spatial gene distribution and anatomical localization at the whole mouse brain level. Our subsequent spatial correlation analysis systematically compared the PPs to known anatomical regions and ontology from the Allen Mouse Brain Atlas using spatial neighborhoods. We demonstrate that our stable and spatially coherent PPs, whose linear combinations accurately approximate the spatial gene data, are highly correlated with combinations of expert-annotated brain regions. These PPs yield a brain ontology based purely on spatial gene expression. Our PP identification approach outperforms principal component analysis and typical clustering algorithms on the same task. Moreover, we show that the stable PPs reveal marked regional imbalance of brainwide genetic architecture, leading to region-specific marker genes and gene coexpression networks. Our findings highlight the advantages of stability-driven machine learning for plausible biological discovery from dense spatial gene expression data, streamlining tasks that are infeasible by conventional manual approaches.
                Cover page: Unsupervised pattern identification in spatial gene expression atlas reveals mouse brain regions beyond established ontology.
                Top