Daniel Yamins

Performance-optimized hierarchical models predict neural responses in higher visual cortex

Significance Humans and monkeys easily recognize objects in scenes. This ability is known to be supported by a network of hierarchically interconnected brain areas. However, understanding neurons in higher levels of this hierarchy has long remained a major challenge in visual systems neuroscience. We use computational techniques to identify a neural network model that matches human performance on challenging object categorization tasks. Although not explicitly constrained to match neural data, this model turns out to be highly predictive of neural responses in both the V4 and inferior temporal cortex, the top two layers of the ventral visual hierarchy. In addition to yielding greatly improved models of visual cortex, these results suggest that a process of biological performance optimization directly shaped neural mechanisms. The ventral visual stream underlies key human visual object recognition abilities. However, neural encoding in the higher areas of the ventral stream remains poorly understood. Here, we describe a modeling approach that yields a quantitatively accurate model of inferior temporal (IT) cortex, the highest ventral cortical area. Using high-throughput computational techniques, we discovered that, within a class of biologically plausible hierarchical neural network models, there is a strong correlation between a model’s categorization performance and its ability to predict individual IT neural unit response data. To pursue this idea, we then identified a high-performing neural network that matches human performance on a range of recognition tasks. Critically, even though we did not constrain this model to match neural data, its top output layer turns out to be highly predictive of IT spiking responses to complex naturalistic images at both the single site and population levels. Moreover, the model’s intermediate layers are highly predictive of neural responses in the V4 cortex, a midlevel visual area that provides the dominant cortical input to IT. These results show that performance optimization—applied in a biologically appropriate model class—can be used to build quantitative predictive models of neural processing.

Deep Neural Networks Rival the Representation of Primate IT Cortex for Core Visual Object Recognition

The primate visual system achieves remarkable visual object recognition performance even in brief presentations, and under changes to object exemplar, geometric transformations, and background variation (a.k.a. core visual object recognition). This remarkable performance is mediated by the representation formed in inferior temporal (IT) cortex. In parallel, recent advances in machine learning have led to ever higher performing models of object recognition using artificial deep neural networks (DNNs). It remains unclear, however, whether the representational performance of DNNs rivals that of the brain. To accurately produce such a comparison, a major difficulty has been a unifying metric that accounts for experimental limitations, such as the amount of noise, the number of neural recording sites, and the number of trials, and computational limitations, such as the complexity of the decoding classifier and the number of classifier training examples. In this work, we perform a direct comparison that corrects for these experimental limitations and computational considerations. As part of our methodology, we propose an extension of “kernel analysis” that measures the generalization accuracy as a function of representational complexity. Our evaluations show that, unlike previous bio-inspired models, the latest DNNs rival the representational performance of IT cortex on this visual object recognition task. Furthermore, we show that models that perform well on measures of representational performance also perform well on measures of representational similarity to IT, and on measures of predicting individual IT multi-unit responses. Whether these DNNs rely on computational mechanisms similar to the primate visual system is yet to be determined, but, unlike all previous bio-inspired models, that possibility cannot be ruled out merely on representational performance grounds.

Using goal-driven deep learning models to understand sensory cortex.

Fueled by innovation in the computer vision and artificial intelligence communities, recent developments in computational neuroscience have used goal-driven hierarchical convolutional neural networks (HCNNs) to make strides in modeling neural single-unit and population responses in higher visual cortical areas. In this Perspective, we review the recent progress in a broader modeling context and describe some of the key technical innovations that have supported it. We then outline how the goal-driven HCNN approach can be used to delve even more deeply into understanding the development and organization of sensory cortical processing.

Identification and functional validation of the novel antimalarial resistance locus PF10_0355 in Plasmodium falciparum.

The Plasmodium falciparum parasites ability to adapt to environmental pressures, such as the human immune system and antimalarial drugs, makes malaria an enduring burden to public health. Understanding the genetic basis of these adaptations is critical to intervening successfully against malaria. To that end, we created a high-density genotyping array that assays over 17,000 single nucleotide polymorphisms (∼1 SNP/kb), and applied it to 57 culture-adapted parasites from three continents. We characterized genome-wide genetic diversity within and between populations and identified numerous loci with signals of natural selection, suggesting their role in recent adaptation. In addition, we performed a genome-wide association study (GWAS), searching for loci correlated with resistance to thirteen antimalarials; we detected both known and novel resistance loci, including a new halofantrine resistance locus, PF10_0355. Through functional testing we demonstrated that PF10_0355 overexpression decreases sensitivity to halofantrine, mefloquine, and lumefantrine, but not to structurally unrelated antimalarials, and that increased gene copy number mediates resistance. Our GWAS and follow-on functional validation demonstrate the potential of genome-wide studies to elucidate functionally important loci in the malaria parasite genome.

Dynamic Task Assignment in Robot Swarms.

A large group of robots will often be partitioned into subgroups, each subgroup performing a different task. This paper presents four distributed algorithms for assigning swarms of homogenous robots to subgroups to meet a specified global task distribution. Algorithm Random-Choice selects tasks randomly, but runs in constant time. Algorithm Extreme-Comm compiles a complete inventory of all the robots on every robot, runs quickly, but uses a great deal of communication. The CardDealer’s algorithm assigns tasks to individual robots sequentially, using minimal communications but a great deal of time. The TreeRecolor algorithm is a compromise between Extreme-Comm and Card-Dealer’s, balancing communications use and running time. The three deterministic algorithms drive the system towards the desired assignment of subtasks with high accuracy. We implement the algorithms on a group of 25 iRobot SwarmBots, and collect and analyze performance data.

Hyperopt: a Python library for model selection and hyperparameter optimization

Sequential model-based optimization (also known as Bayesian optimization) is one of the most efficient methods (per function evaluation) of function minimization. This efficiency makes it appropriate for optimizing the hyperparameters of machine learning algorithms that are slow to train. The Hyperopt library provides algorithms and parallelization infrastructure for performing hyperparameter optimization (model selection) in Python. This paper presents an introductory tutorial on the usage of the Hyperopt library, including the description of search spaces, minimization (in serial and parallel), and the analysis of the results collected in the course of minimization. This paper also gives an overview of Hyperopt-Sklearn, a software project that provides automatic algorithm configuration of the Scikit-learn machine learning library. Following Auto-Weka, we take the view that the choice of classifier and even the choice of preprocessing module can be taken together to represent a single large hyperparameter optimization problem. We use Hyperopt to define a search space that encompasses many standard components (e.g. SVM, RF, KNN, PCA, TFIDF) and common patterns of composing them together. We demonstrate, using search algorithms in Hyperopt and standard benchmarking data sets (MNIST, 20-newsgroups, convex shapes), that searching this space is practical and effective. In particular, we improve on best-known scores for the model space for both MNIST and convex shapes. The paper closes with some discussion of ongoing and future work.

StarFlow: A Script-Centric Data Analysis Environment

We introduce StarFlow, a script-centric environment for data analysis. StarFlow has four main features: (1) extraction of control and data-flow dependencies through a novel combination of static analysis, dynamic runtime analysis, and user annotations, (2) command-line tools for exploring and propagating changes through the resulting dependency network, (3) support for workflow abstractions enabling robust parallel executions of complex analysis pipelines, and (4) a seamless interface with the Python scripting language. We describe real applications of StarFlow, including automatic parallelization of complex workflows in the cloud.

A performance-optimized model of neural responses across the ventral visual stream

Human visual object recognition is subserved by a multitude of cortical areas. To make sense of this system, one line of research focused on response properties of primary visual cortex neurons and developed theoretical models of a set of canonical computations such as convolution, thresholding, exponentiating and normalization that could be hierarchically repeated to give rise to more complex representations. Another line or research focused on response properties of high-level visual cortex and linked these to semantic categories useful for object recognition. Here, we hypothesized that the panoply of visual representations in the human ventral stream may be understood as emergent properties of a system constrained both by simple canonical computations and by top-level, object recognition functionality in a single unified framework (Yamins et al., 2014; Khaligh-Razavi and Kriegeskorte, 2014; Güçlü and van Gerven, 2015). We built a deep convolutional neural network model optimized for object recognition and compared representations at various model levels using representational similarity analysis to human functional imaging responses elicited from viewing hundreds of image stimuli. Neural network layers developed representations that corresponded in a hierarchical consistent fashion to visual areas from V1 to LOC. This correspondence increased with optimization of the model’s recognition performance. These findings support a unified view of the ventral stream in which representations from the earliest to the latest stages can be understood as being built from basic computations inspired by modeling of early visual cortex shaped by optimization for high-level object-based performance constraints. Significance Statement Prior work has taken two complimentary approaches to understanding the cortical processes underlying our ability to visually recognize objects. One approach identified canonical computations from primary visual cortex that could be hierarchically repeated and give rise to complex representations. Another approach linked later visual area responses to semantic categories useful for object recognition. Here we combined both approaches by optimizing a deep convolution neural network based on canonical computations to preform object recognition. We found that this network developed hierarchically similar response properties to those of visual areas we measured using functional imaging. Thus, we show that object-based performance optimization results in predictive models that not only share similarity with late visual areas, but also intermediate and early visual areas.

A Task-Optimized Neural Network Replicates Human Auditory Behavior, Predicts Brain Responses, and Reveals a Cortical Processing Hierarchy

A core goal of auditory neuroscience is to build quantitative models that predict cortical responses to natural sounds. Reasoning that a complete model of auditory cortex must solve ecologically relevant tasks, we optimized hierarchical neural networks for speech and music recognition. The best-performing network contained separate music and speech pathways following early shared processing, potentially replicating human cortical organization. The network performed both tasks as well as humans and exhibited human-like errors despite not being optimized to do so, suggesting common constraints on network and human performance. The network predicted fMRI voxel responses substantially better than traditional spectrotemporal filter models throughout auditory cortex. It also provided a quantitative signature of cortical representational hierarchy-primary and non-primary responses were best predicted by intermediate and late network layers, respectively. The results suggest that task optimization provides a powerful set of tools for modeling sensory systems.

Neural Mechanisms Underlying Visual Object Recognition

Invariant visual object recognition and the underlying neural representations are fundamental to higher-level human cognition. To understand these neural underpinnings, we combine human and monkey psychophysics, large-scale neurophysiology, neural perturbation methods, and computational modeling to construct falsifiable, predictive models that aim to fully account for the neural encoding and decoding processes that underlie visual object recognition. A predictive encoding model must minimally describe the transformation of the retinal image to population patterns of neural activity along the entire cortical ventral stream of visual processing and must accurately predict the responses to any retinal image. A predictive decoding model must minimally describe the transformation from those population patterns of neural activity to observed object recognition behavior (i.e., subject reports), and, given that population pattern of activity, it must accurately predict behavior for any object recognition task. To date, we have focused on core object recognition-a remarkable behavior that is accomplished with image viewing durations of <200 msec. Our work thus far reveals that the neural encoding process is reasonably well explained by a largely feed-forward, highly complex, multistaged nonlinear neural network-the current best neuronal simulation models predict approximately one-half of the relevant neuronal response variance across the highest levels of the ventral stream (areas V4 and IT). Remarkably, however, the decoding process from IT to behavior for all object recognition tasks tested thus far is very accurately predicted by simple direct linear conversion of the inferior temporal neural population state to behavior choice. We have recently examined the behavioral consequences of direct suppression of IT neural activity using pharmacological and optogenetic methods and find them to be well-explained by the same linear decoding model.