Profile cover photo
Profile photo
Jason Yosinski
Hacker, Cornell Ph.D. candidate in Computer Science
Hacker, Cornell Ph.D. candidate in Computer Science

Jason's posts

Post has attachment

Post has attachment

Post has attachment
Here's a quick video tour of our new open-source Deep Visualization Toolbox, which allows you to investigate what each neuron in a deep neural network does: 

We also have a blog post describing how we generate synthetic images for what each neuron in a neural net wants to see. We provide examples for all 1000 ImageNet classes: 

From our paper to appear at the ICML-DL Workshop on Friday: Yosinski, Clune, Nguyen, Fuchs, and Lipson. "Understanding Neural Networks through Deep Visualization."

Post has attachment
My niece Emery knows about 50 words by now. I'm proud to say her 51st word was "Gnu". Next step: teach her Emacs.

Post has attachment
Wired just posted a writeup on our Fooling paper. In my opinion it's well-written and presents a more balanced summary than the first few hype-laden articles. I especially liked the inclusion of both types of reactions to the paper:

"The reactions [at the NIPS conference] sorted into two rough groups. One group—generally older, with more experience in the field—saw how the study made sense. They might’ve predicated a different outcome, but at the same time, they found the results perfectly understandable. The second group, comprised of people who perhaps hadn’t spent as much time thinking about what makes today’s computer brains tick, were struck by the findings. At least initially, they were surprised these powerful algorithms could be so plainly wrong."

Post has attachment
I am pleased to announce a new paper called “Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images,” written by +Anh Nguyen, myself, and +Jeff Clune.

Trained neural nets "see" the world in a different way than humans; this disparity has been the topic of a few recent papers. Here we further investigate these differences and along the way come up with two potentially useful methods for visualizing what supervised neural nets are really looking for.

Both methods use optimization to find images that cause high activations, but with different priors in image space. One directly regularizes to explore parts of image space that contain less high-frequency information, and another employs a genetic algorithm with a strong direct bias towards images exhibiting compositionality. This latter method produces fooling images of a completely different character than have previously been found, almost looking like abstract art!

Congratulations to Anh on a great paper!

Post has attachment
Here's the arXiv preprint version of a paper by +Jeff Clune, +Yoshua Bengio, +Hod Lipson and myself that attempts to answer the question "How transferable are features in deep neural networks?"

Many people have noticed that the first layers of neural nets trained on images tend to produce Gabor features and color blobs, prompting the suspicion that such features are generic to many image datasets and tasks. But to what extent is this true? And to what extent are higher layers generic?

In this study we measure the generality of features as the extent to which they are transferrable from one task to another, and in the process come across a few interesting results:

 - Transferability is negatively affected by two distinct issues: not only the specialization of higher layer neurons to their original task, but also optimization difficulties encountered when chopping neural nets in half, severing connections between co-adapted neurons.
 - Which of these two effects dominates can depend on whether features are transferred from the bottom, middle, or top of the network.
 - Features in the middle of a network can transfer well to other semantically similar tasks but much more poorly to semantically distant tasks.
 - We also observe a surprising effect that initializing a network with transferred features from almost any number of layers can produce a boost to generalization that lingers even after extensive fine-tuning to the target dataset.

This paper will be presented at NIPS 2014.

Post has attachment
Protests in Hong Kong are burgeoning. There are, like, at least twice as many people around this bridge in Central now as when I was there two months ago!

Our paper on "Quantifying the transferability of features in deep neural networks" was accepted at NIPS for oral presentation! Abstract below, arXiv version soon. See you soon, Montreal :).

(with +Jeff Clune, +Yoshua Bengio, and +Hod Lipson)

Post has attachment
On Friday I had the chance to visit the folks at Nervana Systems ( in San Diego to chat about their ambitious goal of producing custom hardware for training and deploying deep neural nets. Any glimpse over the GPU hedge is very exciting stuff! I can't wait to have some prototype boards to play with (those are in the mail, right, +Arjun Bansal?)
Wait while more posts are being loaded