Jason Yosinski
Jason Yosinski is a founding member of Uber AI Labs and there leads
the Deep Collective research group.
He is known for contributions to understanding neural network modeling, representations, and training. Prior to Uber, Jason worked on robotics at Caltech, co-founded two web companies, and started a robotics program in Los Angeles middle schools that now serves over 500 students. He completed his PhD working at the Cornell Creative Machines Lab, University of Montreal, JPL, and Google DeepMind. He is a recipient of the NASA Space Technology Research Fellowship, has co-authored over 50 papers and patents, and was VP of ML at Geometric Intelligence, which Uber acquired. His work has been profiled by NPR, the BBC, Wired, The Economist, Science, and the NY Times. In his free time, Jason enjoys cooking, reading, paragliding, and pretending he's an artist.
Engineering Blog Articles
Controlling Text Generation with Plug and Play Language Models
Plug and Play Language Model, introduced by Uber AI Labs, gives NLP practitioners the flexibility to plug in one or more simple attribute models into a large, unconditional language model.
Uber Goes to NeurIPS 2019
Uber is presenting 11 papers at the NeurIPS 2019 conference in Vancouver, Canada, as well as sponsoring workshops including Women in Machine Learning (WiML) and Black in AI.
Introducing LCA: Loss Change Allocation for Neural Network Training
Uber AI Labs proposes Loss Change Allocation (LCA), a new method that provides a rich window into the neural network training process.
Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask
Uber builds upon the Lottery Ticket Hypothesis by proposing explanations behind these mechanisms and deriving a surprising by-product: the Supermask.
Faster Neural Networks Straight from JPEG
Uber AI Labs introduces a method for making neural networks that process images faster and more accurately by leveraging JPEG representations.
How to Get a Better GAN (Almost) for Free: Introducing the Metropolis-Hastings GAN
Metropolis-Hastings Generative Adversarial Networks (GANs) leverage the discriminator to pick better samples from the generator after ML model training is done.
An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution
As powerful and widespread as convolutional neural networks are in deep learning, AI Labs’ latest research reveals both an underappreciated failing and a simple fix.
Measuring the Intrinsic Dimension of Objective Landscapes
Curious about what it is like to traverse the high-dimensional loss landscapes of modern neural networks? Check out Uber AI Labs’ latest research on measuring intrinsic dimension to find out.
Introducing the Uber AI Residency
Interested in accelerating your career by tackling some of Uber’s most challenging AI problems? Apply for the Uber AI Residency, a research fellowship dedicated to fostering the next generation of AI talent.
Research Papers
First-Order Preconditioning via Hypergradient Descent
T. Moskovitz, R. Wang, J. Lan, S. Kapoor, T. Miconi, J. Yosinski, A. Rawal
Standard gradient descent methods are susceptible to a range of issues that can impede training, such as high correlations and different scaling in parameter space.These difficulties can be addressed by second-order approaches that apply a pre-conditioning matrix to the gradient to improve convergence. Unfortunately, such algorithms typically struggle to scale to high-dimensional problems, in part because the calculation of specific preconditioners such as the inverse Hessian or Fisher information matrix is highly expensive. We introduce first-order preconditioning (FOP), a fast, scalable approach that generalizes previous work on hypergradient descent (Almeida et al., 1998; Maclaurin et al., 2015; Baydin et al.,2017) to learn a preconditioning matrix that only makes use of first-order information. [...] [PDF]
Conference on Neural Information Processing Systems (NeurlPS), 2019
Standard gradient descent methods are susceptible to a range of issues that can impede training, such as high correlations and different scaling in parameter space.These difficulties can be addressed by second-order approaches that apply a pre-conditioning matrix to the gradient to improve convergence. Unfortunately, such algorithms typically struggle to scale to high-dimensional problems, in part because the calculation of specific preconditioners such as the inverse Hessian or Fisher information matrix is highly expensive. We introduce first-order preconditioning (FOP), a fast, scalable approach that generalizes previous work on hypergradient descent (Almeida et al., 1998; Maclaurin et al., 2015; Baydin et al.,2017) to learn a preconditioning matrix that only makes use of first-order information. [...] [PDF]
Conference on Neural Information Processing Systems (NeurlPS), 2019
Estimating Q(s,s’) with Deep Deterministic Dynamics Gradients
A. Edwards, Himanshu Sahni, R. Liu, J. Hung, A. Jain, R. Wang, A. Ecoffet, T. Miconi, C. Isbell, J. Yosinski
In this paper, we introduce a novel form of value function, Q(s,s′), that expresses the utility of transitioning from a state s to a neighboring state s′ and then acting optimally thereafter. In order to derive an optimal policy, we develop a forward dynamics model that learns to make next-state predictions that maximize this value. [...] [PDF]
International Conference on Machine Learning (ICML), 2020
In this paper, we introduce a novel form of value function, Q(s,s′), that expresses the utility of transitioning from a state s to a neighboring state s′ and then acting optimally thereafter. In order to derive an optimal policy, we develop a forward dynamics model that learns to make next-state predictions that maximize this value. [...] [PDF]
International Conference on Machine Learning (ICML), 2020
Plug and Play Language Models: A Simple Approach to Controlled Text Generation
S. Dathathri, A. Madotto, J. Lan, J. Hung, E. Frank, P. Molino, J. Yosinski, R. Liu
Large transformer-based language models (LMs) trained on huge text corpora have shown unparalleled generation capabilities. However, controlling attributes of the generated language (e.g. switching topic or sentiment) is difficult without modifying the model architecture or fine-tuning on attribute-specific data and entailing the significant cost of retraining. We propose a simple alternative: the Plug and Play Language Model (PPLM) for controllable language generation, which combines a pretrained LM with one or more simple attribute classifiers that guide text generation without any further training of the LM. [PDF]
International Conference on Learning Representations (ICLR), 2020
Large transformer-based language models (LMs) trained on huge text corpora have shown unparalleled generation capabilities. However, controlling attributes of the generated language (e.g. switching topic or sentiment) is difficult without modifying the model architecture or fine-tuning on attribute-specific data and entailing the significant cost of retraining. We propose a simple alternative: the Plug and Play Language Model (PPLM) for controllable language generation, which combines a pretrained LM with one or more simple attribute classifiers that guide text generation without any further training of the LM. [PDF]
International Conference on Learning Representations (ICLR), 2020
Hamiltonian Neural Networks
S. Greydanus, M. Dzamba, J. Yosinski
Even though neural networks enjoy widespread use, they still struggle to learn the basic laws of physics. How might we endow them with better inductive biases? In this paper, we draw inspiration from Hamiltonian mechanics to train models that learn and respect exact conservation laws in an unsupervised manner. [...] [PDF]
Conference on Neural Information Processing Systems (NeurIPS), 2019
Even though neural networks enjoy widespread use, they still struggle to learn the basic laws of physics. How might we endow them with better inductive biases? In this paper, we draw inspiration from Hamiltonian mechanics to train models that learn and respect exact conservation laws in an unsupervised manner. [...] [PDF]
Conference on Neural Information Processing Systems (NeurIPS), 2019
LCA: Loss Change Allocation for Neural Network Training
J. Lan, R. Liu, H. Zhou, J. Yosinski
Neural networks enjoy widespread use, but many aspects of their training, representation, and operation are poorly understood. In particular, our view into the training process is limited, with a single scalar loss being the most common viewport into this high-dimensional, dynamic process. We propose a new window into training called Loss Change Allocation (LCA), in which credit for changes to the network loss is conservatively partitioned to the parameters. [...] [PDF]
Conference on Neural Information Processing Systems (NeurIPS), 2019
Neural networks enjoy widespread use, but many aspects of their training, representation, and operation are poorly understood. In particular, our view into the training process is limited, with a single scalar loss being the most common viewport into this high-dimensional, dynamic process. We propose a new window into training called Loss Change Allocation (LCA), in which credit for changes to the network loss is conservatively partitioned to the parameters. [...] [PDF]
Conference on Neural Information Processing Systems (NeurIPS), 2019
Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask
H. Zhou, J. Lan, R. Liu, J. Yosinski
Optical Character Recognition (OCR) approaches have been widely advanced in recent years thanks to the resurgence of deep learning. The state-of-the-art models are mainly trained on the datasets consisting of the constrained scenes. Detecting and recognizing text from the real-world images remains a technical challenge. [...] [PDF]
Conference on Neural Information Processing Systems (NeurIPS), 2019
Optical Character Recognition (OCR) approaches have been widely advanced in recent years thanks to the resurgence of deep learning. The state-of-the-art models are mainly trained on the datasets consisting of the constrained scenes. Detecting and recognizing text from the real-world images remains a technical challenge. [...] [PDF]
Conference on Neural Information Processing Systems (NeurIPS), 2019
Metropolis-Hastings Generative Adversarial Networks
R. Turner, J. Hung, Y. Saatci, J. Yosinski
We introduce the Metropolis-Hastings generative adversarial network (MH-GAN), which combines aspects of Markov chain Monte Carlo and GANs. The MH-GAN draws samples from the distribution implicitly defined by a GAN's discriminator-generator pair, as opposed to sampling in a standard GAN which draws samples from the distribution defined by the generator. [...] [PDF]
International Conference on Machine Learning (ICML), 2019
We introduce the Metropolis-Hastings generative adversarial network (MH-GAN), which combines aspects of Markov chain Monte Carlo and GANs. The MH-GAN draws samples from the distribution implicitly defined by a GAN's discriminator-generator pair, as opposed to sampling in a standard GAN which draws samples from the distribution defined by the generator. [...] [PDF]
International Conference on Machine Learning (ICML), 2019
Understanding Neural Networks via Feature Visualization: A survey
A. Nguyen, J. Yosinski, J. Clune
A neuroscience method to understanding the brain is to find and study the preferred stimuli that highly activate an individual cell or groups of cells. Recent advances in machine learning enable a family of methods to synthesize preferred stimuli that cause a neuron in an artificial or biological brain to fire strongly. [...] [PDF]
Interpretable AI: Interpreting, Explaining and Visualizing Deep Learning, 2019
A neuroscience method to understanding the brain is to find and study the preferred stimuli that highly activate an individual cell or groups of cells. Recent advances in machine learning enable a family of methods to synthesize preferred stimuli that cause a neuron in an artificial or biological brain to fire strongly. [...] [PDF]
Interpretable AI: Interpreting, Explaining and Visualizing Deep Learning, 2019
Faster Neural Networks Straight from JPEG
L. Gueguen, A. Sergeev, B. Kadlec, R. Liu, J. Yosinski
The simple, elegant approach of training convolutional neural networks (CNNs) directly from RGB pixels has enjoyed overwhelming empirical success. But can more performance be squeezed out of networks by using different input representations? In this paper we propose and explore a simple idea: train CNNs directly on the blockwise discrete cosine transform (DCT) coefficients computed and available in the middle of the JPEG codec. [...] [PDF]
Advances in Neural Information Processing Systems (NeurIPS), 2018
The simple, elegant approach of training convolutional neural networks (CNNs) directly from RGB pixels has enjoyed overwhelming empirical success. But can more performance be squeezed out of networks by using different input representations? In this paper we propose and explore a simple idea: train CNNs directly on the blockwise discrete cosine transform (DCT) coefficients computed and available in the middle of the JPEG codec. [...] [PDF]
Advances in Neural Information Processing Systems (NeurIPS), 2018
An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution
R. Liu, J. Lehman, P. Molino, F.i Such, E. Frank, A. Sergeev, J. Yosinski
Few ideas have enjoyed as large an impact on deep learning as convolution. For any problem involving pixels or spatial representations, common intuition holds that convolutional neural networks may be appropriate. In this paper we show a striking counterexample to this intuition via the seemingly trivial coordinate transform problem, which simply requires learning a mapping between coordinates in (x,y) Cartesian space and one-hot pixel space. [...] [PDF]
Advances in Neural Information Processing Systems (NeurIPS), 2018
Few ideas have enjoyed as large an impact on deep learning as convolution. For any problem involving pixels or spatial representations, common intuition holds that convolutional neural networks may be appropriate. In this paper we show a striking counterexample to this intuition via the seemingly trivial coordinate transform problem, which simply requires learning a mapping between coordinates in (x,y) Cartesian space and one-hot pixel space. [...] [PDF]
Advances in Neural Information Processing Systems (NeurIPS), 2018
Measuring the Intrinsic Dimension of Objective Landscapes
Chunyuan Li, Heerad Farkhoor, R. Liu, J. Yosinski
Many recently trained neural networks employ large numbers of parameters to achieve good performance. One may intuitively use the number of parameters required as a rough gauge of the difficulty of a problem. But how accurate are such notions? How many parameters are really needed? In this paper we attempt to answer this question by training networks not in their native parameter space, but instead in a smaller, randomly oriented subspace. [...] [PDF]
International Conference on Learning Representations (ICLR), 2018
Many recently trained neural networks employ large numbers of parameters to achieve good performance. One may intuitively use the number of parameters required as a rough gauge of the difficulty of a problem. But how accurate are such notions? How many parameters are really needed? In this paper we attempt to answer this question by training networks not in their native parameter space, but instead in a smaller, randomly oriented subspace. [...] [PDF]
International Conference on Learning Representations (ICLR), 2018
The surprising creativity of digital evolution: A collection of anecdotes from the evolutionary computation...
J. Lehman, J. Clune, D. Misevic, C. Adami, L. Altenberg, J. Beaulieu, P. Bentley, S. Bernard, G. Beslon, D. Bryson, P. Chrabaszcz, N. Cheney, A. Cully, S. Doncieux, F. Dyer, K. Ellefsen, R. Feldt, S. Fischer, S. Forrest, A. Frénoy, C. Gagné, L. Goff, L. Grabowski, B. Hodjat, F. Hutter, L. Keller, C. Knibbe, P. Krcah, R. Lenski, H. Lipson, R. MacCurdy, C. Maestre, R. Miikkulainen, S. Mitri, D. Moriarty, J. Mouret, A. Nguyen, C. Ofria, M. Parizeau, D. Parsons, R. Pennock, W. Punch, T. Ray, M. Schoenauer, E. Shulte, K. Sims, K. Stanley, F. Taddei, D. Tarapore, S. Thibault, W. Weimer, R. Watson, J. Yosinski
Biological evolution provides a creative fount of complex and subtle adaptations, often surprising the scientists who discover them. However, because evolution is an algorithmic process that transcends the substrate in which it occurs, evolution's creativity is not limited to nature. [...] [PDF]
2018
Biological evolution provides a creative fount of complex and subtle adaptations, often surprising the scientists who discover them. However, because evolution is an algorithmic process that transcends the substrate in which it occurs, evolution's creativity is not limited to nature. [...] [PDF]
2018
Automated Identification of Northern Leaf Blight-Infected Maize Plants from Field Imagery Using Deep Learning
C. DeChant, T. Wiesner-Hanks, S, Chen, E. Stewart, J. Yosinski, M. Gore, R. Nelson, and H. Lipson
Northern leaf blight (NLB) can cause severe yield loss in maize; however, scouting large areas to accurately diagnose the disease is time consuming and difficult. We demonstrate a system capable of automatically identifying NLB lesions in field-acquired images of maize plants with high reliability. [...] [PDF]
Phytopathology, 2017
Northern leaf blight (NLB) can cause severe yield loss in maize; however, scouting large areas to accurately diagnose the disease is time consuming and difficult. We demonstrate a system capable of automatically identifying NLB lesions in field-acquired images of maize plants with high reliability. [...] [PDF]
Phytopathology, 2017
Time-series extreme event forecasting with neural networks at Uber
N. Laptev, J. Yosinski, L. Li, S. Smyl
Accurate time-series forecasting during high variance segments (e.g., holidays), is critical for anomaly detection, optimal resource allocation, budget planning and other related tasks. At Uber accurate prediction for completed trips during special events can lead to a more efficient driver allocation resulting in a decreased wait time for the riders. [PDF]
International Conference on Machine Learning (ICML), 2017
Accurate time-series forecasting during high variance segments (e.g., holidays), is critical for anomaly detection, optimal resource allocation, budget planning and other related tasks. At Uber accurate prediction for completed trips during special events can lead to a more efficient driver allocation resulting in a decreased wait time for the riders. [PDF]
International Conference on Machine Learning (ICML), 2017
SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability
M. Raghu, J. Gilmer, J. Yosinski, J. Sohl-Dickstein
We propose a new technique, Singular Vector Canonical Correlation Analysis (SVCCA), a tool for quickly comparing two representations in a way that is both invariant to affine transform (allowing comparison between different layers and networks) and fast to compute (allowing more comparisons to be calculated than with previous methods). [...] [PDF]
Neural Information Processing Systems (NIPS), 2017
We propose a new technique, Singular Vector Canonical Correlation Analysis (SVCCA), a tool for quickly comparing two representations in a way that is both invariant to affine transform (allowing comparison between different layers and networks) and fast to compute (allowing more comparisons to be calculated than with previous methods). [...] [PDF]
Neural Information Processing Systems (NIPS), 2017
Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space
A. Nguyen, J. Clune, Y. Bengio, A. Dosovitskiy, J. Yosinski
Generating high-resolution, photo-realistic images has been a long-standing goal in machine learning. Recently, Nguyen et al. (2016) showed one interesting way to synthesize novel images by performing gradient ascent in the latent space of a generator network to maximize the activations of one or multiple neurons in a separate classifier network. In this paper we extend this method by introducing an additional prior on the latent code, improving both sample quality and sample diversity, leading to a state-of-the-art generative model that produces high quality images at higher resolutions (227x227) than previous generative models, and does so for all 1000 ImageNet categories. [...] [PDF]
Computer Vision and Pattern Recognition (CVPR), 2017
Generating high-resolution, photo-realistic images has been a long-standing goal in machine learning. Recently, Nguyen et al. (2016) showed one interesting way to synthesize novel images by performing gradient ascent in the latent space of a generator network to maximize the activations of one or multiple neurons in a separate classifier network. In this paper we extend this method by introducing an additional prior on the latent code, improving both sample quality and sample diversity, leading to a state-of-the-art generative model that produces high quality images at higher resolutions (227x227) than previous generative models, and does so for all 1000 ImageNet categories. [...] [PDF]
Computer Vision and Pattern Recognition (CVPR), 2017
