Typically, the energy minimization or snakes based object detection frameworks evolve a parametrized curve guided by some form of image gradient information. However due to heavy reliance on gradients, the approaches tend to fail in scenarios where this information is misleading or unavailable. This cripples the snake and renders it unusable as it gets stuck in a local-minima away from the actual object. Moreover, the parametrized snake lacks the ability to model multiple evolving curves in a single run. In order to address these issues, Chan and Vese introduced a new framework which utilized region based information to guide a spline, and tries to solve the minimal partition problem formulated by Mumford and Shah. 

The framework is built upon the following energy equation, where $C$ is a level-set formulation of the curve:$F(c1, c2, C) = \mu . \text{Length}(C) + v . \text{Area}(inside(C))\\
        \lambda_1 \int_{inside(C)}|u_0(x,y) - c_1|^2 dxdy + \lambda_2 \int_{outside(C)}|u_0(x,y) - c_2|^2 dxdy$
The framework essentially divides the image into two regions (per curve), which are referred to as inside and outside of the curve. The first two terms of the equation control the physical aspects of the curve, particularly the length, and area inside the curve, with their contributions controlled by two parameters $\mu$ and $v$. 

The image forces in this equation correspond to the third and fourth terms, which are identical but work in respective regions identified by the curve. The terms use $c_1$ and $c_2$, which are the mean intensity values inside and outside the curve respectively to guide the curve towards a minima where the both regions are consistent with respect to the mean intensity values. 

The proposed framework also utilizes an improved representation of the curve in the form of a level set function $\phi(x, y)$, which has many numerical advantages and naturally supports multiple curves evolved during a single run, as compared to the traditional snakes model where only one curve can be evolved in a single run.  The unknown function $\phi$ is computed using Euler-Lagrange equations formulated using modified Heaviside function $H$, and Dirac measure $\delta$.

The proposed framework was applied on numerous challenging 2D images with varying degree of difficulties. The framework was also capable of segmenting point clouds decomposed into 2D images, objects with blurred boundaries, and contours without gradients, all without requiring image denoising. 

Due to the formulation of $\phi$ approximation routine, the framework has tendencies to find actual global minima independent of the initial position of the curve. However the choice of multiple parameters namely $\lambda_{1,2}, \mu, v$ is done heuristically, and seem to be problem dependent. Also, the framework's implicit dependency on absolute image intensities in regions inside and outside of curve sometimes fail in very specific cases where the averages tend to be zero, though the authors proposed to use image curvature and orientation information from the initial image $u_0$.

Following one of the basic principles in evolutionary biology that complex life forms derive from more primitive ancestors, it has long been believed that the higher animals, the Bilateria, arose from simpler (diploblastic) organisms such as the cnidarians (corals, polyps, and jellyfishes). A large number of studies, using different datasets and different methods, have tried to determine the most ancestral animal group as well as the ancestor of the higher animals. Here, we use “total evidence” analysis, which incorporates all available data (including morphology, genome, and gene expression data) and come to a surprising conclusion. The Bilateria and Cnidaria (together with the other diploblastic animals) are in fact sister groups: that is, they evolved in parallel from a very simple common ancestor. We conclude that the higher animals (Bilateria) and lower animals (diploblasts), probably separated very early, at the very beginning of metazoan animal evolution and independently evolved their complex body plans, including body axes, nervous system, sensory organs, and other characteristics. The striking similarities in several complex characters (such as the eyes) resulted from both lineages using the same basic genetic tool kit, which was already present in the common ancestor. The study identifies Placozoa as the most basal diploblast group and thus a living fossil genome that nicely demonstrates, not only that complex genetic tool kits arise before morphological complexity, but also that these kits may form similar morphological structures in parallel.

- SIFT family is still the best local descriptor, outperforms novel CNN [SiamNet2015] approaches.
- (adaptive) Hessian-Affine is the best detector with broad applicability (not beaten yet)
- Affine view synthesis greatly helps for non-geometrical problems.
- Datasets and WxBS-Matcher available http://cmp.felk.cvut.cz/wbs/
-  We need more diverse datasets for learning local descriptors than Yosemite and Libert 

The method they use basically tells the robot to reason as follows:
1. The human gave me a reward function $\tilde{r}$, selected in order to get me to behave the way they wanted.
2. So I should favor reward functions which produce that kind of behavior. 

This amounts to doing RL (step 1) followed by IRL on the learned policy (step 2); see the final paragraph of section 4.

Meta learning is an area sparking a lot of research curiosity these days. It’s framed in different ways: models that can adapt, models that learn to learn, models that can learn a new task quickly. This paper uses a somewhat different lens: that of neural plasticity, and argues that applying the concept to modern neural networks will give us an effective, and biologically inspired way of building adaptable models. The basic premise of plasticity from a neurobiology perspective (at least how it was framed in the paper: I’m not a neuroscientist myself, and may be misunderstanding) is that plasticity performs a kind of gating function on the strength of a neural link being upregulated by experience. The most plastic a connection is, the more quickly it can get modified by new data; the less plastic, the more fixed it is. 

In concrete terms, this is implemented by subdividing the weight on each connection in the network into two parts: the “fixed” component, and the “plastic” component. (see picture). The fixed component acts like a typical weight: it gets modified during training, but stays fixed once training is done. The plastic component is composed of an alpha weight, multiplied by a term H. H is basically a decaying running average of the past input*output activations of this weight. Activations that are high in magnitude, and the same sign, for both the input and the output will lead to H being pushed higher. Note that that this H can continue to be updated even after the model is done training, because it builds up information whenever you pass a new input X through the network. The plastic component’s learned weight, alpha, controls how strong the influence of this is on the model. If alpha is near zero, then the connection behaves basically identically to a “typical” neural network, with weights that don’t change as a function of activation values. If alpha is positive, that means that strong co-activation within H will tend to make the connection weight higher. If alpha is negative, the opposite is true, and strong co-activation will make the connection weight more negative. (As an aside, I’d be really interested to see the distribution over alpha values in a trained model, relative to the weight values, and look at how often they go in the same direction as the weights, and increase magnitude, and how often they have the opposite direction and attenuate the weight towards zero).

These models are trained by running them for fixed size “episodes” during which the H value gets iteratively changed, and then the alpha parameters of H get updated in the way that would have reduced error over the episode.  One area in which they seem to show strong performance is that of memorization (where the network is shown an image once, and needs to reconstruct it later). The theory for why this is true is that the weights are able to store short-term information about which pixels are in the images it sees by temporarily boosting themselves higher for inputs and activations they’ve recently seen. 

There are definitely some intuitional gaps for me in this paper. The core one is: this framework just makes weights able to update themselves as a function of the values of their activations, not as a function of an actual loss function. That is to say: it seems like a potentially better analogy to neural plasticity is just a network that periodically gets more training data, and has some amount of connection plasticity to update as a result of that.