Talk the Talk: Optimization’s Untold Gift to Learning

by Li Yang Ku (Gooly)

In this post I am going to talk about a fascinating talk by Nati Srebro at ICML this June. Srebro have given similar talks at many places but I think he really nailed it this time. This talk is interesting not only because he provided a different view of the role of optimization in deep learning but also because he clearly explained why many researcher’s argument on the reason that deep learning works doesn’t make sense.

Srebro first look into what we know about deep learning (typical feed forward network) based on three questions. The first question is regarding the capacity of the network. How many samples do we need to learn certain network architecture? The short answer is that it should be proportional to the number of parameters in the network, which is the total number of edges. The second question is about the expressiveness of the network. What can we express with certain model class? What type of questions can we learn? Since a two layer neural network is a universal approximator, it can learn any continuous function, this is however not a very useful information since it may require an exponentially large network and exponential amount of samples to learn. So the more interesting question is what can we express with a reasonable sized network? Many recent research more or less focuses on this question. However, Srebro argues that since there is another theory that says any function that can be executed within a reasonable amount of time can be captured by a network of reasonable size (please comment below if you know what theory this is), all problems that we expect to be solvable can be expressed by a reasonable sized network.

The third question is about computation. How hard is it to find optimal parameters? The bad news is that finding the weights for even tiny networks is NP-Hard. Theories (link1 link2) show that even if the training data can be perfectly expressed by a small neural network there are no polynomial time algorithm to find such set of weights. This means that neural network’s expressiveness described in question 2 doesn’t really do much good since we aren’t capable of finding the optimal solution. But we all know that in reality neural network works pretty well, it seems that there are some magical property that allows us to learn neural networks. Srebro emphasizes that we still don’t know what is the magical property that makes neural networks learnable, but we do know it is not because we can represent the data well with the network. If you ask vision folks why neural networks work, they might say something like the lower layers of the network matches low level visual features and the higher layers match higher level visual features. However, this answer is about the expressiveness of the network described in question 2 which is not sufficient for explaining why neural networks work and provides zero evidence since we already know neural networks have the power to express any problem.

Srebro then talked about the observed behavior that neural networks usually don’t overfit to the training data. This is an unexpected property quite similar to the behavior of Adaboost, which was invented in 1997 and quite popular in the 2000s. It was only after the invention that people discovered that the reason Adaboost doesn’t overfit is because it is implicitly minimizing the L-1 norm that limits the complexity. So the question Srebro pointed out was whether the gradient decent algorithm for learning neural networks are also implicitly minimizing certain complexity measure that would be beneficial in reaching a solution that would generalize. Given a set of training data, a neural network can have multiple optimal solutions that are global minima (zero training error). However, some of these global minima perform better than the others on the test data. Srebro argues that the optimization algorithm might be doing something else other than just minimizing the training error. Therefore, by changing the optimization algorithm we might observe a difference in how well can a neural network generalize to test data, and this is exactly what Srebro’s group discovered. In one experiment they showed that even though using Adam optimization achieves lower training error then stochastic gradient decent, it actually performs worse on the test data. What this means is that we might not be putting enough emphasize on optimization in the deep learning community where a typical paper looks like the following:

The contributions are on the model and loss function, while the optimization is just a brief mention. So the main point Srebro is trying to convey is that different optimization algorithms would lead to different inductive biases, and different inductive biases would lead to different generalization properties. “We need to understand optimization algorithm not just as reaching some global optimum, but as reaching a specific optimum.”

Srebro further talked about a few more works based on these observations. If you are interested by now, you should probably watch the whole video (You would need to fast forward a bit to start.) I am however going to put in a little bit of my own thoughts here. Srebro emphasizes the importance of optimization a lot in this talk and said the deep models we use now can basically express any problem we have, therefore the model is not what makes deep learning work. However, we also know that the model does matter based on claims of many papers that invented new model architectures. So how could both of these claims be true? We have to remember that the model architecture is also part of the optimization process that shapes the geometry which the optimization algorithm is optimizing on. Hence, if the nerual network model provides a landscape that allows the optimization algorithm to reach a desired minimum more easily, it will also generalize better to the test data. In other words, the model and the optimization algorithm have to work together.

The Deep Learning Not That Smart List

by Li Yang Ku (Gooly)

Deep learning is one of the most successful scientific story in modern history, attracting billions of investment money in half a decade. However, there is always the other side of the story where people discover the less magical part of deep learning. This post is about a few research (quite a few published this year) that shows deep learning might not be as smart as you think (most of the time they would came up with a way to fix it, since it used to be forbidden to accept paper without deep learning improvements.) This is just a short list, please comment below on other papers that also belong.

a) Szegedy, Christian, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. “Intriguing properties of neural networks.”, ICLR 2014

The first non-magical discovery of deep learning has to go to the finding of adversarial examples. It was discovered that images added with certain unnoticeable perturbations can result in mysterious false detections by a deep network. Although technically the first publication of this discovery should go to the paper “Evasion Attacks against Machine Learning at Test Time” by Battista Biggio et al. published in September 2013 in ECML PKDD, the paper that really caught people’s attention is this one that was put on arxiv in December 2013 and published in ICLR 2014. In addition to having bigger names on the author list, this paper also show adversarial examples on more colorful images that clearly demonstrates the problem (see image below.) Since this discover, there have been continuous battles between the band that tries to increase the defense against attacks and the band that tries to break it (such as “Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples” by Athalye et al.), which leads to a recent paper in ICLR 2019 “Are adversarial examples inevitable?” by Shafahi et al. that questions whether it is possible that a deep network can be free of adversarial examples from a theoretical standpoint.

b) Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. “Deep Image Prior.” CVPR 2018

This is not a paper intended to discover flaws of deep learning, in fact, the result of this paper is one of the most magical deep learning results I’ve seen. The authors showed that deep networks are able to fill in cropped out images in a very reasonable way (see image below, left input, right output) However, it also unveils some less magical parts of deep learning. Deep learning’s success was mostly advertised as learning from data and claimed to work better than traditional engineered visual features because it learns from large amount of data. This work, however, uses no data nor pre-trained weights. It shows that convolution and the specific layered network architecture, (which may be the outcome of millions of grad student hours through trial and error,) played a significant role in the success. In other words, we are still engineering visual features but in a more subtle way. It also raises the question of what made deep learning so successful, is it because of learning? or because thousands of grad students tried all kinds of architectures, lost functions, training procedures, and some combinations turned out to be great?

c) Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. “ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness.” ICLR 2019.

It was widely accepted in the deep learning community that CNNs recognize objects by combining lower level filters that represent features such as edges into more complex shapes layer by layer. In this recent work, the authors noticed that contrary to what the community believes, existing deep learning models seems to have a strong bias towards textures. For example, a cat with elephant texture is often recognized as an elephant. Instead of learning how a cat looks like, CNNs seem to take the short cut and just try to recognize cat fur. You can find a detailed blog post about this work here.

d) Wieland Brendel, and Matthias Bethge. “Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet.” ICLR 2019.

This is a paper from the same group as the previous paper. Based on the same observations, this paper claims that CNNs are not that different from bag of feature approaches that classifies based on local features. The authors created a network that only looks at local patches in an image without high level spatial information and was able to achieve pretty good result on ImageNet. The author further shuffled features in an image and existing deep learning models seems to be not sensitive to these changes. Again CNNs seem to be taking short cuts by making classifications based on just local features. More on this work can be found in this post.

e) Azulay, Aharon, and Yair Weiss. “Why do deep convolutional networks generalize so poorly to small image transformations?.” rejected by ICLR 2019.

This is a paper that discovered that modern deep networks may fail to recognize images shifted 1 pixel apart, but got rejected because reviewers don’t quite buy-in on the experiments nor the explanation. (the authors made a big mistake of not providing an improved deep network in the paper.) The paper showed that when the image is shifted slightly or if a sequence of frames from a video is given to a modern deep network, jaggedness appear in the detection result (see example below where the posterior probability of recognizing the polar bear varies a lot frame by frame.) The authors further created a dataset from ImageNet with the same images embedded in a larger image frame at a random location and showed that the performance dropped about 30% when the embedded frame is twice the width of the original image. This work shows that despite modern networks getting close to human performance on image classification tasks on ImageNet, it might not be able to generalize to the real world as well as we hoped.

f) Nalisnick, Eric, Akihiro Matsukawa, Yee Whye Teh, Dilan Gorur, and Balaji Lakshminarayanan. “Do Deep Generative Models Know What They Don’t Know?.” ICLR 2019

This work from DeepMind looks into tackling the problem that when tested on data with a distribution different from training, deep neural network can give wrong results with high confidence. For example, in the paper “Multiplicative Normalizing Flows for Variational Bayesian Neural Networks” by Louizos and Welling, it was discovered that on the MNIST dataset a trained network can be highly confident but wrong when the input number is tilted. This makes deploying deep learning to critical tasks quite problematic. Deep generative models were thought to be a solution to such problems, since it also models the distribution of the samples, it can reject anomalies if it does not belong to the same distribution as the training samples. However, the authors short answer to the question is no; even for very distinct datasets such as digits versus images of horse and trucks, anomalies cannot be identified, and many cases even wrongfully provide stronger confidence than samples that does come from the trained dataset. The authors therefore “urge caution when using these models with out-of-training-distribution inputs or in unprotected user-facing systems.”

Paper Picks: CVPR 2018

by Li Yang Ku (Gooly)

I was at CVPR in salt lake city. This year there were more then 6500 attendances and a record high number of accepted papers. People were definitely struggling to see them all. It was a little disappointing that there were no keynote speakers, but among the 9 major conferences I have been to, this one has the best dance party (see image below). You never know how many computer scientists can dance until you give them unlimited alcohol.

In this post I am going to talk about a few papers that were not the most popular ones but were what I personally found interesting. If you want to know the papers that the reviewers though were interesting instead, you can look into the best paper “Taskonomy: Disentangling Task Transfer Learning” and four other honorable mentions including the “SPLATNet: Sparse Lattice Networks for Point Cloud Processing” from collaborations between Nvidia and some people in the vision lab at UMass Amherst which I am in.

a) Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. “Learning and Using the Arrow of Time.”

I am quite fond of works that explore cues in the world that may be useful for unsupervised learning. Traditional deep learning approaches requires large amount of labeled training data but we humans seem to be able to learn from just interacting with the world in an unsupervised fashion. In this paper, the direction of time is used as a clue. The authors train a neural network to distinguish the direction of time and show that such network can be helpful in action recognition tasks.

b) Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. “Learning to Localize Sound Source in Visual Scenes.”

This is another example of using cues available in the world. In this work, the authors ask whether a machine can learn the correspondence between visual scene and sound, and localize the sound source only by observing sound and visual scene pairs like humans? This is done by using a triplet network that tries to minimize the difference between visual feature of a video frame and the sound feature generated in a similar time window, while maximizing the difference between the same visual feature and a random sound feature. As you can see in the figure above, the network is able to associate different sounds with different visual regions.

c) Edward Kim, Darryl Hannan, and Garrett Kenyon. “Deep Sparse Coding for Invariant Multimodal Halle Berry Neurons.”

This work is inspired by experiments done by Quiroga et al. that found a single neuron in one human subject’s brain that fires on both pictures of Halle Berry and texts of Halle Berry’s name. In this paper, the authors show that training a deep sparse coding network that takes a face image and a text image of the corresponding name results in learning a multimodal invariant neuron that fires on both Halle Berry’s face and name. When certain modality is missing, the missing image or text can be generated. In this network, each sparse coding layer is learned through the Locally Competitive Algorithm (LCA) that uses principles of thresholding and local competition between neurons. Top down feedback is also used in this work through propagating reconstruction error downwards. The authors show interesting results where adding information to one modality changes the belief of the other modality. The figure above shows that this Halle Berry neuron in the sparse coding network can distinguish between cat women acted by Halle Berry versus cat women acted by Anne Hathaway and Michele Pfeiffer.

d) Assaf Shocher, Nadav Cohen, and Michal Irani. “Zero-Shot Super-Resolution using Deep Internal Learning.”

Super resolution is a task that tries to increase the resolution of an image. The typical approach nowaday is to learn it through a neural network. However, the author showed that this approach only works well if the down sampling process from the high resolution to the low resolution image is similar in training and testing. In this work, no training is needed beforehand. Given a test image, training examples are generated from the test image by down sampling patches of this same image. The fundamental idea of this approach is the fact that natural images have strong internal data repetition. Therefore, from the same image you can infer high resolution structures of lower resolution patches by observing other parts of the image that have higher resolution and similar structure. The image above shows their results (top row) versus state of the art results (bottom row).

e) Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. “Deep Image Prior.”

Most modern approaches for denoising, super resolution, or inpainting tasks use an image generation network that trains on a large dataset that consist of pairs of images before and after the affect. This work shows that these nice outcomes are not just the result of learning but also the effect of the convolutional structure. The authors take an image generation network, feed random noise as input, and then update the network using the error between the outcome and the test image, such as the left image shown above for inpainting. After many iterations, the network magically generates an image that fills the gap, such as the right image above. What this works says is that unlike common belief that deep learning approaches for image restoration learns image priors better then engineered priors, the deep structure itself is just a better engineered prior.

Deep Learning Approaches For Object Detection

by Li Yang Ku

In this post I am going to talk about the progression of a few deep learning approaches for object detection. I will start from R-CNN and OverFeat (2013) then gradually move to more recent approaches such as the RetinaNet which won the best student paper in ICCV 2017. Object detection here refers to the task of identifying a limited set of object classes (20 ~ 200) in a given image by giving each identified object a bounding box and a label. This is one of the main stream challenges in Computer Vision which requires algorithms to output the locations of multiple object in addition to corresponding class. Some of the most well known datasets are the PASCAL visual object classes challenge (2005-2012) funded by the EU (20 classes ~10k images), the ImageNet object detection challenge (2013 ~ present) sponsored by Stanford, UNC, Google, and Facebook (200 classes ~500k images) , and the COCO dataset (2015 ~ current) first started by Microsoft (80 classes ~200K images). These datasets provide hand labeled bounding boxes and class labels of objects in images for training. Challenges for these datasets happen yearly; teams from all over the world submit their code to compete on an undisclosed test set.

In December 2012, the success of Alexnet on the ImageNet classification challenge was published. While many computer vision scientist around the world were still scratching their head trying to understand this result, several groups quickly harvested techniques implemented in Alexnet and tested it out. Based on the success of Alexnet, in November 2013 the vision group in Berkeley published (on arxiv) an approach for solving the object detection problem. This proposed R-CNN is a simple extension that extends the Alexnet that was designed to solve the classification problem to handle the detection problem. R-CNN is composed of 3 parts, 1) region proposal: where selective search is used to generate around 2000 possible object location bounding boxes, 2) feature extraction: Alexnet is used to generate features, 3) classification: a SVM (support vector machine) is trained for each object class. This hybrid approach successfully outperformed previous algorithms on the PASCAL dataset by a significant margin.

R-CNN architecture

Around the same time (December 2013), the NYU team (Yann LeCun, Rob Fergus) published an approach called OverFeat. OverFeat is based on the idea that convolutions can be done efficiently on dense image locations in a sliding window fashion. The fully connected layers in the Alexnet can be seen as 1×1 convolution layers. Therefore, instead of generating a classification confidence for a cropped fix size image, OverFeat generates a map of confidence on the whole image. To predict the bounding box a regressor network is added after the convolution layers. OverFeat was at the 4th place during the 2013 ImageNet object detection challenge but claimed to have better then 1st place result with longer training time which wasn’t ready in time for the competition.

Since then, a lot of researches expanded based on concepts introduced in these work. The SPP-net is an approach that speeds up the R-CNN approach up to 100x by performing the convolution operations just once on the whole image. (note that OverFeat does convolution on images of different scale) The SPP-net adds a spatial pyramid pooling layer before the fully connected layers. This spatial pyramid pooling layer transforms an arbitrary size feature map into a fixed size input by pooling from areas separated by grids of different scale. However, similar to R-CNN, SPP-net requires multistep training on feature extraction and the SVM classification. Fast R-CNN came across to address this problem. Similar to R-CNN, Fast R-CNN uses selective search to generate a set of possible region proposals and by adapting the idea of SPP-net, feature map is generated once on the whole image and a ROI pooling layers extracts a fixed size features for each region proposal. A multi task loss is also used so that the whole network can be trained together in one stage. The Fast R-CNN can speed up R-CNN up to 200x and produce better accuracy.

Fast R-CNN architecture

At this point, the region proposal process have become the computation bottleneck for Fast R-CNN. As a result, the “Faster” R-CNN addresses this issue by introducing the region proposal network that generates region proposals based on the same feature map used for classification. This requires a four stage training that alternates between these two networks but achieves a 5 frames per second speed.

Image pyramid where images of multiple scales are created for feature extraction was a common approach used in features such as SIFT features to handle scale invariant. So far, most R-CNN based approaches does not use image pyramids due to the computation and memory cost during training. The feature pyramid network shows that since deep convolution neural networks are by natural multi-scale, a similar effect can be achieved with little extra cost. This is done by combining top-down information with lateral information for each convolution layer as shown in the figure below. By restricting the feature maps to have the same dimension, the same classification network can be used for all scales; this has a similar flavor to traditional approaches that use the same detector for images of different scales in the image pyramid.

Till 2017, most of the high accuracy approaches on object detection are extensions of R-CNN that have a region proposal module separate from classification. Single stage approaches although faster, were not able to out perform in accuracy. The paper “Focal Loss for Dense Object Detection” published in ICCV 2017 discovers the problem with single stage approaches and proposed an elegant solution that results in faster and more accurate models. The lower accuracy among single stage approaches was a consequence of imbalance between foreground and background training examples. By replacing the cross entropy loss with the focal loss that down weights examples the network already has high confidence, the network improves substantially on accuracy. The figure below shows the difference between the cross entropy loss (CE) and the focal loss (FL). A larger gamma parameter puts less weight on high confidence examples.

The references of approaches I mentioned is listed below. Note that I only talked about a small part of a large body of work on object detection and the current progress on object detection have been moving in a rapid speed. If you look at the current leader board for the COCO dataset, the numbers have already surpassed the best approach I have mentioned by a substantial margin.

• Girshick, Ross, Jeff Donahue, Trevor Darrell, and Jitendra Malik. “Rich feature hierarchies for accurate object detection and semantic segmentation.” In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 580-587. 2014.
• Sermanet, Pierre, David Eigen, Xiang Zhang, Michaël Mathieu, Rob Fergus, and Yann LeCun. “Overfeat: Integrated recognition, localization and detection using convolutional networks.” arXiv preprint arXiv:1312.6229 (2013).
• He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. “Spatial pyramid pooling in deep convolutional networks for visual recognition.” In european conference on computer vision, pp. 346-361. Springer, Cham, 2014.
• Girshick, Ross. “Fast r-cnn.” arXiv preprint arXiv:1504.08083(2015).
• Ren, Shaoqing, Kaiming He, Ross Girshick, and Jian Sun. “Faster r-cnn: Towards real-time object detection with region proposal networks.” In Advances in neural information processing systems, pp. 91-99. 2015.
• Lin, Tsung-Yi, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. “Feature pyramid networks for object detection.” In CVPR, vol. 1, no. 2, p. 4. 2017.
• Lin, Tsung-Yi, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. “Focal loss for dense object detection.” arXiv preprint arXiv:1708.02002 (2017).

 

Machine Learning, Computer Vision, and Robotics

By Li Yang Ku (Gooly)

Having TA’d for Machine Learning this semester and worked in the field of Computer Vision and Robotics for the past few years, I always have this feeling that the more I learn the less I know. Therefore, its sometimes good to just sit back and look at the big picture. This post will talk about how I see the relations between these three fields in a high level.

First of all, Machine Learning is more a brand then a name. Just like Deep Learning and AI, this name is used for getting funding when the previous name used is out of hype. In this case, the name popularized after AI projects failed in the 70s. Therefore, Machine learning covers a wide range of problems and approaches that may look quite different at first glance. Adaboost and support vector machine was the hot topic in Machine Learning when I was doing my master’s degree, but now it is deep neural network that gets all the attention.

Despite the wide variety of research in Machine Learning, they usually have this common assumption on the existent of a set of data. The goal is then to learn a model based on this set of data. There are a wide range of variations here, the data could be labeled or not labeled resulting in supervised or unsupervised approaches; the data could be labeled with a category or a real number, resulting in classification or regression problems; the model can be limited to a certain form such as a class of probability models, or can have less constraints in the case of deep neural network. Once the model is learned, there are also a wide range of possible usage. It can be used for predicting outputs given new inputs, filling missing data, generating new samples, or providing insights on hidden relationships between data entries. Data is so fundamental in Machine Learning, people in the field don’t really ask the question of why learning from data. Many datasets from different fields are collected or labeled and the learned models are compared based on accuracy, computation speed, generalizability, etc. Therefore Machine Learning people often consider Computer Vision and Robotics as areas for applying Machine Learning techniques.

Robotics on the other hand comes from a very different background. There are usually no data to start with in robotics. If you cannot control your robot or if your robot crashes itself at first move, how are you going to collect any data. Therefore, classical robotics is about designing models based on physics and geometries. You build models that model how the input and current observation of the robot changes the robot state. Based on this model you can infer the input that will safely control the robot to reach certain state.

Once you can command your robot to reach certain state, a wide variety of problems emerge. The robot will then have to do obstacle avoidance and path planning to reach certain goal. You may need to to find a goal state that satisfies a set of restrictions while optimizing a set of properties. Simultaneous localization and mapping (SLAM) may be needed if no maps are given. In addition, sensor fusion is required when multiple sensors with different properties are used. There may also be uncertainties in robot states where belief space planning may be helpful. For robots with a gripper, you may also need to be able to identify stable grasps and recognizing the type and pose of an object for manipulation. And of course, there is a whole different set of problems on designing the mechanics and hardware of the robot.  Unlike Machine Learning, a lot of approaches of these problems are solved without a set of data. However, most of these robotics problems (excluding mechanical and hardware problems) share a common goal of determining the robot input based on feedback. (Some) Roboticists view robotics as the field that has the ultimate goal of creating machines that act like humans, and Machine Learning and Computer Vision are fields that can provide methods to help accomplish such goal.

The field of Computer Vision started under AI in the 60s under the goal of helping robots to achieve intelligent behaviors, but left such goal behind after the internet era when tons of images on the internet are waiting to be classified. In this age, computer vision applications are no longer restricted to physical robots. In the past decade, the field of Computer Vision is driven by datasets. The implicit agreement on evaluation based on standardized datasets helped the field to advance in a reasonably fast pace (under the cost of millions of grad student hours on tweaking models to get a 1% improvement.) Given these datasets, the field of Computer Vision inevitably left the Robotics community and embraced the data-driven Machine Learning approaches. Most Computer Vision problems have a common goal of learning models for visual data. The model is then used to do classification, clustering, sample generation, etc. on images or videos. The big picture of Computer Vision can be seen in my previous post. Some Computer Vision scientists consider vision different from other senses and believe that the development of vision is fundamental to the evolution of intelligence (which could be true… experiments do show 50% of our brain neurons are vision related.) Nowadays, Computer Vision and Machine Learning are deeply tangled; Machine Learning techniques help foster Computer Vision solutions, while successful models in Computer Vision contribute back to the field of Machine Learning. For example, the successful story of Deep Learning started from Machine Learning models being applied to the ImageNet challenge, and end up with a wide range of architectures that can be applied to other problems in Machine Learning. On the other hand, Robotics is a field where Computer Vision folks are gradually moving back to. Several well known Computer Vision scientists, such as Jitendra Malik, started to consider how Computer Vision can help the field of Robotics ,since their conversation with Robotics colleagues were mostly about vision not working, based on the recent success on data-driven approaches in Computer Vision.

Paper Picks: ICRA 2017

by Li Yang Ku (Gooly)

I was at ICRA (International Conference on Robotics and Automation) in Singapore to present one of my work this June. Surprisingly, the computer vision track seems to gain a lot of interest in the robotics community. The four computer vision sessions are the most crowded ones among all the sessions that I have attended. The following are a few papers related to computer vision and deep learning that I found quite interesting.

a) Schmidt, Tanner, Richard Newcombe, and Dieter Fox. “Self-supervised visual descriptor learning for dense correspondence.”

In this work, a self-supervised learning approach is introduced for generating dense visual descriptors with convolutional neural networks. Given a set of RGB-D videos of Schmidt, the first author, wandering around, a set of training data can be automatically generated by using Kinect Fusion to track feature points between frames. A pixel-wise contrastive loss is used such that two points belong to the same model point would have similar descriptors.

Kinect Fusion cannot associate points between videos, however with just training data within the same video, the authors show that the learned descriptors of the same model point (such as the tip of the nose) are similar across videos. This can be explained by the hypothesis that with enough data, a model point trajectory will inevitably come near to the same model point trajectory in another video. By chaining these trajectories, clusters of the same model point can be separated even without labels. The figure above visualizes the learned features with colors. Note that it learns a similar mapping across videos despite with no training signal across videos.

b) Pavlakos, Georgios, Xiaowei Zhou, Aaron Chan, Konstantinos G. Derpanis, and Kostas Daniilidis. “6-dof object pose from semantic keypoints.”

In this work, semantic keypoints predicted by convolutional neural networks are combined with a deformable shape model to estimate the pose of object instances or objects of the same class. Given a single RGB image of an object, a set of class specific keypoints is first identified through a CNN that is trained on labeled feature point heat maps. A fitting problem that maps these keypoints to keypoints on the 3D model is then solved using a deformable model that captures different shape variability. The figure above shows some pretty good results on recognizing the same feature of objects of the same class.

The CNN used in this work is the stacked hourglass architecture, where two hourglass modules are stacked together. The hourglass module was introduced in the paper “Newell, Alejandro, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. ECCV, 2016.” An hourglass module is similar to a fully convolutional neural network but with residual modules, which the authors claim to make it more balanced between down sampling and up sampling. Stacking multiple hourglass modules allows repeated bottom up, top down inferences which improves on the state of the art performances.

c) Sung, Jaeyong, Ian Lenz, and Ashutosh Saxena. “Deep Multimodal Embedding: Manipulating Novel Objects with Point-clouds, Language and Trajectories.”

In this work, point cloud, natural language, and manipulation trajectory data are mapped to a shared embedding space using a neural network. For example, given the point cloud of an object and a set of instructions as input, the neural network should map it to a region in the embedded space that is close to the trajectory that performs such action. Instead of taking the whole point cloud as input, a segmentation process that decides which part of the object to manipulate based on the instruction is first executed. Based on this shared embedding space, the closest trajectory to where the input point cloud and language map to can be executed during test time.

In order to learn a semantically meaningful embedding space, a loss-augmented cost that considers the similarity between different types of trajectory is used. The result shows that the network put similar groups of actions such as pushing a bar and moving a cup to a nozzle close to each other in the embedding space.

d) Finn, Chelsea, and Sergey Levine. “Deep visual foresight for planning robot motion.”

In this work, a video prediction model that uses a convolutional LSTM (long short-term memory) is used to predict pixel flow transformation from the current frame to the next frame for a non-prehensile manipulation task. This model takes the input image, end-effector pose, and a future action to predict the image of the next time step. The predicted image is then fed back into the network recursively to generate the next image. This network is learned from 50000 pushing examples of hundreds of objects collected from 10 robots.

For each test, the user specifies where certain pixels on an object should move to, the robot then uses the model to determine actions that will most likely reach the target using an optimization algorithm that samples actions for several iterations. Some of the results are shown in the figure above, the first column indicates the interface where the user specifies the goal. The red markers are the starting pixel positions and the green markers of the same shape are the goal positions. Each row shows a sequence of actions taken to reach the specified target.

Generative Adversarial Nets: Your Enemy is Your Best Friend?

by Li Yang Ku (gooly)

Generating realistic images with machines was always one of the top items on my list of difficult tasks. Past attempts in the Computer Vision community were only able to get a blurry image at best. The well publicized Google Deepdream project was able to generate some interesting artsy images, however they were modified from existing images and were designed more to make you feel like on drugs then realistic. Recently (2016), a work that combines the generative adversarial network framework with convolutional neural networks (CNNs) generated some results that look surprisingly good. (A non vision person would likely not be amazed though.) This approach was quickly accepted by the community and was referenced more then 200 times in less then a year.

This work is based on an interesting concept first introduced by Goodfellow et al. in the paper “Generative Adversarial Nets” at NIPS 2014 (http://papers.nips.cc/paper/5423-generative-adversarial-nets). The idea was to have two neural networks compete with each other. One would try to generate images as realistic as it can and the other network would try to distinguish them from real images at its best. By theory this competition will reach a global optimum where the generated image and the real image will belong to the same distribution (Could be a lot trickier in practice though). This work in 2014 got some pretty good results on digits and faces but the generated natural images are still quite blurry (see figure above).

In the more recent work “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks” by Radford, Metz, and Chintala, convolutional neural networks and the generative adversarial net framework are successfully combined with a few techniques that help stabilize the training (https://arxiv.org/abs/1511.06434). Through this approach, the generated images are sharp and surprisingly realistic at first glance. The figures above are some of the generated bedroom images. Notice that if you look closer some of them may be weird.

The authors further explored what the latent variables represents. Ideally the generator (neural network that generates image) should disentangle independent features and each latent variable should represent a meaningful concept. By modifying these variables, images that have different characteristics can be generated. Note that these latent variables are what given to the neural network that generates images and is randomly sampled from a uniform distribution in the previous examples. In the figure above is an example where the authors show that the latent variables do represent meaningful concepts through arithmetic operations. If you subtract the average latent variables of men without glasses from the average latent variables of men with glasses and add the average latent variables of women without glasses, you obtain a latent variable that result in women with glasses when passed through the generator. This process identifies the latent variables that represent glasses.

 

 

 

Convolutional Neural Network Features for Robot Manipulation

by Li Yang Ku (Gooly)

In my previous post, I mentioned the obstacles when applying deep learning techniques directly to robotics. First, training data is harder to acquire; Second, interacting with the world is not just a classification problem. In this post, I am gonna talk about a really simple approach that treats convolutional neural networks (CNNs) as a feature extractor that generates a set of features similar to traditional features such as SIFT. This idea is applied to grasping on Robonaut 2 and published in arXiv (Associating Grasp Configurations with Hierarchical Features in Convolutional Neural Networks) with more details. The ROS package called ros-deep-vision that generates such features using a RGB-D sensor is also public.

 

When we look at these deep models such as CNNs, we should keep in mind that these models work well because how the layers stack up hierarchically matches how the data is structured. Our observed world is also hierarchical, there are common shared structures such as edges that can be used to represent more complex structures such as squares and cubes when combined in meaningful ways. A simple view of CNN is just a tree structure, where a higher level neuron is a combination of neurons in the previous layer. For example, a neuron that represents cuboids is a combination of neurons that represent the corners and edges of the cuboid. The figures above show such examples on neurons that found to activate consistently on cuboids and cylinders.

By taking advantage of this hierarchical nature of CNN, we can turn a CNN into a feature extractor that generates features that represents local structures of a higher level structure. For example, such hierarchical features can represent the left edge of the top face of a box while traditional edge detectors would find all edges in the scene. Instead of representing a feature with a single filter (neuron) in one of the CNN layers, this feature, which we call hierarchical CNN feature, uses a tuple of filters from different layers. Using backpropagation that restricts activation to one filter per layer allows us to locate the location of such feature precisely. By finding features such as the front and back edge of the top face of a box we can learn where to place robot fingers relative to these hierarchical CNN features in order to manipulate the object.

 

The most cited papers in computer vision and deep learning

by Li Yang Ku (Gooly)

In 2012 I started a list on the most cited papers in the field of computer vision. I try to keep the list focus on researches that relate to understanding this visual world and avoid image processing, survey, and pure statistic works. However, the computer vision world have changed a lot since 2012 when deep learning techniques started a trend in the field and outperformed traditional approaches on many computer vision benchmarks. No matter if this trend on deep learning lasts long or not I think these techniques deserve their own list.

As I mentioned in the previous post, it’s not always the case that a paper cited more contributes more to the field. However, a highly cited paper usually indicates that something interesting have been discovered. The following are the papers to my knowledge being cited the most in Computer Vision and Deep Learning (note that it is “and” not “or”). If you want a certain paper listed here, just comment below.

Cited by 5518

Imagenet classification with deep convolutional neural networks

A Krizhevsky, I Sutskever, GE Hinton, 2012

Cited by 1868

Caffe: Convolutional architecture for fast feature embedding

Y Jia, E Shelhamer, J Donahue, S Karayev…, 2014

Cited by 1681

Backpropagation applied to handwritten zip code recognition

Y LeCun, B Boser, JS Denker, D Henderson…, 1989

Cited by 1516

Rich feature hierarchies for accurate object detection and semantic segmentation

R Girshick, J Donahue, T Darrell…, 2014

Cited by 1405

Very deep convolutional networks for large-scale image recognition

K Simonyan, A Zisserman, 2014

Cited by 1169

Improving neural networks by preventing co-adaptation of feature detectors

GE Hinton, N Srivastava, A Krizhevsky…, 2012

Cited by 1160

Going deeper with convolutions

C Szegedy, W Liu, Y Jia, P Sermanet…, 2015

Cited by 977

Handwritten digit recognition with a back-propagation network

BB Le Cun, JS Denker, D Henderson…, 1990

Cited by 907

Visualizing and understanding convolutional networks

MD Zeiler, R Fergus, 2014

Cited by 839

Dropout: a simple way to prevent neural networks from overfitting

N Srivastava, GE Hinton, A Krizhevsky…, 2014

Cited by 839

Overfeat: Integrated recognition, localization and detection using convolutional networks

P Sermanet, D Eigen, X Zhang, M Mathieu…, 2013

Cited by 818

Learning multiple layers of features from tiny images

A Krizhevsky, G Hinton, 2009

Cited by 718

DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition

J Donahue, Y Jia, O Vinyals, J Hoffman, N Zhang…, 2014

Cited by 691

Deepface: Closing the gap to human-level performance in face verification

Y Taigman, M Yang, MA Ranzato…, 2014

Cited by 679

Deep Boltzmann Machines

R Salakhutdinov, GE Hinton, 2009

Cited by 670

Convolutional networks for images, speech, and time series

Y LeCun, Y Bengio, 1995

Cited by 570

CNN features off-the-shelf: an astounding baseline for recognition

A Sharif Razavian, H Azizpour, J Sullivan…, 2014

Cited by 549

Learning hierarchical features for scene labeling

C Farabet, C Couprie, L Najman…, 2013

Cited by 510

Fully convolutional networks for semantic segmentation

J Long, E Shelhamer, T Darrell, 2015

Cited by 469

Maxout networks

IJ Goodfellow, D Warde-Farley, M Mirza, AC Courville…, 2013

Cited by 453

Return of the devil in the details: Delving deep into convolutional nets

K Chatfield, K Simonyan, A Vedaldi…, 2014

Cited by 445

Large-scale video classification with convolutional neural networks

A Karpathy, G Toderici, S Shetty, T Leung…, 2014

Cited by 347

Deep visual-semantic alignments for generating image descriptions

A Karpathy, L Fei-Fei, 2015

Cited by 342

Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

K He, X Zhang, S Ren, J Sun, 2015

Cited by 334

Learning and transferring mid-level image representations using convolutional neural networks

M Oquab, L Bottou, I Laptev, J Sivic, 2014

Cited by 333

Convolutional networks and applications in vision

Y LeCun, K Kavukcuoglu, C Farabet, 2010

Cited by 332

Learning deep features for scene recognition using places database

B Zhou, A Lapedriza, J Xiao, A Torralba…,2014

Cited by 299

Spatial pyramid pooling in deep convolutional networks for visual recognition

K He, X Zhang, S Ren, J Sun, 2014

Cited by 268

Long-term recurrent convolutional networks for visual recognition and description

J Donahue, L Anne Hendricks…, 2015

Cited by 261

Two-stream convolutional networks for action recognition in videos

K Simonyan, A Zisserman, 2014

 

Convolutional Neural Networks in Robotics

by Li Yang Ku (Gooly)

As I mentioned in my previous post, Deep Learning and Convolutional Neural Networks (CNNs) have gained a lot of attention in the field of computer vision and outperformed other algorithms on many benchmarks. However, applying these technics to robotics is non-trivial for two reasons. First, training large neural networks requires a lot of training data and collecting them on robots is hard. Not only do research robots easily have network or hardware failures after many trials, the time and resource needed to collect millions of data is also significant. The trained neural network is also robot specific and cannot be used on a different type of robot directly, therefore limiting the incentive of training such network. Second, CNNs are good for classification but when we are talking about interacting with a dynamic environment there is no direct relationship. Knowing you are seeing a lightsaber gives no indication on how to interact with it. Of course you can hard code this information, but that would just be using Deep Learning in computer vision instead of robotics.

Despite these difficulties, a few groups did make it through and successfully applied Deep Learning and CNNs in robotics; I will talk about three of these interesting works.

• Levine, Sergey, et al. “End-to-end training of deep visuomotor policies.” arXiv preprint arXiv:1504.00702 (2015). 
• Finn, Chelsea, et al. “Deep Spatial Autoencoders for Visuomotor Learning.” reconstruction 117.117 (2015): 240. 
• Pinto, Lerrel, and Abhinav Gupta. “Supersizing Self-supervision: Learning to Grasp from 50K Tries and 700 Robot Hours.” arXiv preprint arXiv:1509.06825 (2015).

Traditional policy search approaches in reinforcement learning usually use the output of a “computer vision systems” and send commands to low-level controllers such as a PD controller. In the paper “end-to-end training of deep visuomotor policies”, Sergey, et al. try to learn a policy from low-level observations (image and joint angles) and output joint torques directly. The overall architecture is shown in the figure above. As you can tell this is ambitious and cannot be easily achieved without a few tricks. The authors first initialize the first layer with weights pre-trained on the ImageNet, then train vision layers with object pose information through pose regression. This pose information is obtained by having the robot holding the object with its hand covered by a cloth similar to the back ground (See figure below). 

In addition to that, using the pose information of the object, a trajectory can be learned with an approach called guided policy search. This trajectory is then used to train the motor control layers that takes the visual layer output plus joint configuration as input and output joint torques. The results is better shown then described; see video below.

The second paper, “Deep Spatial Autoencoders for Visuomotor Learning”, is done by the same group in Berkeley. In this work, the authors try to learn a state space for reinforcement learning. Reinforcement learning requires a detailed representation of the state; in most work such state is however usually manually designed. This work automates this state space construction from camera image where the deep spatial autoencoder is used to acquire features that represent the position of objects. The architecture is shown in the figure below.

The deep spatial autoencoder maps full-resolution RGB images to a down-sampled, grayscale version of the input image. All information in the image is forced to pass through a bottleneck of spatial features therefore forcing the network to learn important low dimension representations. The position is then extracted from the bottleneck layer and combined with joint information to form the state representation. The result is tested on several tasks shown in the figure below.

As I mentioned earlier gathering a large amount of training data in robotics is hard, while in the paper “Supersizing Self-supervision: Learning to Grasp from 50K Tries and 700 Robot Hours” the authors try to show that it is possible. Although still not comparable to datasets in the vision community such as ImageNet, gathering 50 thousand tries in robotics is significant if not unprecedented. The data is gathered using this two arm robot Baxter that is (relatively) mass produced compared to most research robots.

 

The authors then use these collected data to train a CNN initialized with weights trained on ImageNet. The final output is one out of 18 different orientation of the gripper, assuming the robot always grab from the top. The architecture is shown in the figure below.