21 Matching Annotations
  1. Jul 2016
    1. A well-known fact is that transferring data to and from GPUs is quite slow. For this reason, we decide to store and update all model parameters on the CPU (see green box).

      Not very clear.

    1. True.

      a = np.array(1) # 0-d, but different from np.int64(1)

      b = np.array([1]) # 1-d, same as a.reshape((1))

      c = np.array([[1]]) # 2-d, same as a.reshape((1,1))

    1. Unsupervised Learning of 3D Structure from Images Authors: Danilo Jimenez Rezende, S. M. Ali Eslami, Shakir Mohamed, Peter Battaglia, Max Jaderberg, Nicolas Heess (Submitted on 3 Jul 2016) Abstract: A key goal of computer vision is to recover the underlying 3D structure from 2D observations of the world. In this paper we learn strong deep generative models of 3D structures, and recover these structures from 3D and 2D images via probabilistic inference. We demonstrate high-quality samples and report log-likelihoods on several datasets, including ShapeNet [2], and establish the first benchmarks in the literature. We also show how these models and their inference networks can be trained end-to-end from 2D images. This demonstrates for the first time the feasibility of learning to infer 3D representations of the world in a purely unsupervised manner.

      The 3D representation of a 2D image is ambiguous and multi-modal. We achieve such reasoning by learning a generative model of 3D structures, and recover this structure from 2D images via probabilistic inference.

    1. When building a unified vision system or gradually adding new capabilities to a system, the usual assumption is that training data for all tasks is always available. However, as the number of tasks grows, storing and retraining on such data becomes infeasible. A new problem arises where we add new capabilities to a Convolutional Neural Network (CNN), but the training data for its existing capabilities are unavailable. We propose our Learning without Forgetting method, which uses only new task data to train the network while preserving the original capabilities. Our method performs favorably compared to commonly used feature extraction and fine-tuning adaption techniques and performs similarly to multitask learning that uses original task data we assume unavailable. A more surprising observation is that Learning without Forgetting may be able to replace fine-tuning as standard practice for improved new task performance.

      Learning w/o Forgetting: distilled transfer learning

  2. Jun 2016
    1. Remember that a slicing tuple can always be constructed as obj and used in the x[obj] notation. Slice objects can be used in the construction in place of the [start:stop:step] notation. For example, x[1:10:5,::-1] can also be implemented as obj = (slice(1,10,5), slice(None,None,-1)); x[obj] . This can be useful for constructing generic code that works on arrays of arbitrary dimension.
    1. Dynamic Filter Networks

      "... filters are generated dynamically conditioned on an input" Nice video frame prediction experiments.

    1. Atl=xtifl= 0MAXPOOL(RELU(CONV(Etl1)))l >0(1)^Atl=RELU(CONV(Rtl))(2)Etl= [RELU(Atl^Atl);RELU(^AtlAtl)](3)Rtl=CONVLSTM(Et1l;Rt1l;Rtl+1)(4)

      Very unique network structure. Prediction results look promising.

    1. Two 140mm fan (front)

      Why are the front fans not designed to be center aligned...