# Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling. NIPS 2017

Access link

Other related resource not presented

• Fundamental paper on GAN [Access link] (https://arxiv.org/pdf/1406.2661v1.pdf)
• you tube video of From Facebook AI research - Soumith Chintala - Adversarial Networks Access link

Presented by Pooran Singh Negi

Summary: Papers propose 3D- Generative Adversial Network(3D-GAN) for 3D object creation (using sample from a probablistic space) and classification task. This latent space can also be inferred as shown in 2D image to 3D image generation( 3D-VAE-GAN section). A VAE is used along with 3D-VAN for this setup. Proposed method shows state of the art and above performance against supervised and unsupervied method. Another interesting work done in paper is showing latent space arithmetics where objects can be manipulated in latent space to build new novel and realistic objects. This is simialr to word to vector of Mikolov but is more impressive as manipuation is done in latent space not the actual learned representation space.

• Main architecture

credit:: All the figure realted to this papers are from ppt and paper from Jiajun Wu*, Chengkai Zhang*, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum

Network:

• Loss functions
  • 3D-GAN
  • 3D-VAE-GAN

Access link

Slides

Presented by Ali Mollahosseini

Summary: This paper proposes a Deep Neural Decision Forests that unifies classification trees with the representation learning functionality known from deep convolutional networks, by training them in an end-to-end manner. Experimental result on benchmark machine learning datasets like MNIST and ImageNet shows superior results when compared to state-of-the-art deep models. Top5-Errors of only 7.84%/6.38% on ImageNet validation data when integrating our forests in a single-crop, single/seven model GoogLeNet architecture, respectively. Thus, even without any form of training data set augmentation we are improving on the 6.67% error obtained by the best GoogLeNet architecture (7 models, 144 crops)

Random forests has been empirically demonstrated to outperform most state-of-the-art learners when it comes to handling high dimensional data problems, they are inherently able to deal with multiclass problems, are easily distributable on parallel hardware architectures while being considered to be close to an ideal learner

Slides on Random Forest

Video on Random Forest

Deep learning approaches is can learn feature representations together with their classifiers. This paper introduces a stochastic, differentiable, and therefore backpropagation compatible version of decision trees, guiding the representation learning in lower layers of deep convolutional networks.

# Scene Labeling with LSTM Recurrent Neural Networks. CVPR 2015 [Access Link](http://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Byeon_Scene_Labeling_With_2015_CVPR_paper.pdf)

Presented by Behzad Hasani

Summary:

This paper addresses the problem of pixel-level segmentation and classification of scene images with an entirelym learning-based approach using Long Short Term Memory (LSTM) recurrent neural networks, which are commonly used for sequence classification. They investigate two-dimensional (2D) LSTM networks for natural scene images taking into account the complex spatial dependencies of labels. Prior methods generally have required separate classification and image segmentation stages and/or pre- and post-processing. In this approach, classification, segmentation, and context integration are all carried out by 2D LSTM networks, allowing texture and spatial model parameters to be learned within a single model. The networks efficiently capture local and global contextual information over raw RGB values and adapt well for complex scene images. This approach, which has a much lower computational complexity than prior methods, achieved state-ofthe- art performance over the Stanford Background and the SIFT Flow datasets. In fact, if no pre- or post-processing is applied, LSTM networks outperform other state-of-the-art approaches. Hence, only with a single-core CPU, the running time of our approach is equivalent or better than the compared state-of-the-art approaches which use a GPU. Finally, the networks’ ability to visualize feature maps from each layer supports the hypothesis that LSTM networks are overall suited for image processing tasks.

What is LSTM?

Proposed Network Architecture:

Results:

Examples: