Gamaleldin Elsayed Personal Website
Gamaleldin Fathy Elsayed

Gamaleldin Fathy Elsayed.

I am an AI resident at Google Brain Mountain View. I am interested in deep learning research with inspiration from neuroscience. In particular, I am interested in studying properties and problems of artificial neural networks, and investigate their overlap with neural systems. This investigations is critical and can help develop useful insights for direction for improvement in deep learning research.

I completed my B.S. from the American University in Cairo (AUC) with a major in Electronics Engineering and a minor in Computer Science. After my B.S., I joined the founding class of King Abdullah University of Science and Technology (KAUST), where I pursued my M.S. in Electrical Engineering under the supervision of Jürgen Kosel and studied EKG signals to extract respiratory-related information. In 2011, I moved to the US to pursue my PhD in Electrical Engineering at Washington University in Saint Louis, where I started to develop my research interest for machine learning and computational neuroscience. In 2013, I moved to Columbia University PhD program, where I joined the Center for Theoretical Neuroscience.

In 2017, I completed my PhD in Neuroscience from Columbia University in the lab of John P. Cunningham, where I worked on problems in computational neuroscience and machine learning. During my PhD I contributed to the field of computational neuroscience through designing methods for identifying and validating structures from complex neural data.

Contact: gamaleldin.elsayed at gmail dot com

Research

Research

I am interests in research topics in deep learning and computational neuroscience.

Adversarial Examples for Humans

Machine learning models are vulnerable to adversarial examples: small changes to images can cause computer vision models to make mistakes such as identifying a school bus as an ostrich. However, it is still an open question whether humans are prone to similar mistakes. Here, we address this question by leveraging recent techniques that transfer adversarial examples from computer vision models with known parameters and architecture to other models with unknown parameters and architecture, and by matching the initial processing of the human visual system. We find that adversarial examples that strongly transfer across computer vision models influence the classifications made by time-limited human observers.

Large Margin Deep Networks

The notion of margin, minimum distance to a decision boundary, has served as the foundation of several successful results for both classification and regression tasks. However, most large margin algorithms are applicable only to shallow models; and conventional margin methods for neural networks only enforce margin at the output layer. Such methods are therefore not well suited for deep networks. In this work, we propose a novel loss function to impose a margin on any chosen set of layers of a deep network (including input and hidden layers). Our formulation allows choosing any norm on the metric measuring the margin. We demonstrate that the decision boundary obtained by our loss has nice properties compared to standard classification loss functions. Specifically, we show improved empirical results on the MNIST, CIFAR-10 and ImageNet datasets on multiple tasks.

Adversarial Reprogramming

Deep neural networks are susceptible to adversarial attacks. Previous adversarial examples have been designed to degrade performance of models or cause machine learning models to produce specific outputs chosen ahead of time by the attacker. We introduce adversarial attacks that instead reprogram the target model to perform a task chosen by the attacker---without the attacker needing to specify or compute the desired output for each test-time input. This attack is accomplished by optimizing for a single adversarial perturbation that can be added to all test-time inputs to a machine learning model in order to cause the model to perform a task chosen by the adversary when processing these inputs---even if the model was not trained to do this task. These perturbations can be thus considered a program for the new task. We demonstrate adversarial reprogramming on six ImageNet classification models, repurposing these models to perform a counting task, as well as two classification tasks.

Reorganization between preparatory and movement population responses

Single neuron responses are highly complex yet are able to flexibly perform different computations underlying different behaviors through their collective activity. Here I study structure of preparatory and movement computations in motor cortex. I demonstrate that collective activity patterns of motor cortical neurons are orthogonal during successive task epochs yet are linearly linked, indicating a degree of flexibility that was not displayed or predicted by existing cortical models. These results reveal a population-level strategy for performing separate but linked computations.

Structure an expected byproduct of simpler phenomena?

Collective patterns of neural responses are increasingly studied from various neural systems. These population structures give concise pictures of the activity of populations beyond the complexity of single-neurons, and are relied upon to support various hypotheses about neural mechanisms. However, the increasing use of population structures without proper validation pose serious problems that could undermine the conclusions drawn from population data. In particular, we should be genuinely concerned that population findings may be outcomes of a powerful data analysis method rather than a fundamental structure as any identified population structure may be a trivial consequence of primary features we already know exist in data. In this work, I use ideas from Fisher randomization, optimization and maximum entropy to design novel methods that to generate random surrogate data while preserving the data primary features, thus overcoming the shortcomings of previous validation methods. With this method, one can perform statistical tests, and be confident that epiphenomenal structures can be identified or ruled out, providing a general-purpose platform for validating population structures for the whole scientific community.

Publications

Journal Papers

Conference Papers and Abstracts

Contact Me


Find me on ...