Past Events

I will talk about finite sample expressivity, aka memorization power of ReLU networks. Recent results showed (unsurprisingly) that arbitrary input data could be perfectly memorized using a shallow ReLU network with one hidden layer having N hidden nodes. I will describe a more careful construction that trades of width with depth to show that a ReLU network with 2 hidden layers, each with 2*sqrt(N) hidden nodes, can perfectly memorize arbitrary datasets. Moreover, we prove that width of Θ(sqrt(N)) is necessary and sufficient for having perfect memorization power.

no results

no results

Many modern neural networks are trained in an over-parameterized regime where the parameters of the model exceed the size of the training dataset. Due to their over-parameterized nature these models in principle have the capacity to (over)fit any set of labels including pure noise. Despite this high fitting capacity, somewhat paradoxically, models trained via first-order methods (often with early stopping) continue to predict well on yet unseen test data.

The role of image quality assessment in tasks such as (i) pan sharpening (PS) (i.e. merging high-resolution panchromatic and lower resolution multispectral imagery to create a single high-resolution color image) and (ii) super-resolution (SR) has not been researched extensively from the natural scene statistics (NSS) perspective. For instance, even though there are several well-known measures that quantify the quality of PS and SR images, there has been little work done on analyzing the statistics of PS and SR images and associated distortions.

Soft electronic devices that can acquire vital signs from the human body represent an important trend for healthcare. Combined strategies of materials design and advanced microfabrication allow the integration of a variety of components and devices on a stretchable platform, resulting in functional systems with minimal constraints on the human body. In this presentation, I will demonstrate a wearable multichannel patch that can sense a collection of signals from the human skin in a wireless mode.

The great progress achieved by communications in the last twenty years can be attested by the amount of audio-visual multimedia services available nowadays, such as digital television and IP-based video transmission. The success of these kind of services relies on their trustworthiness and the delivered quality of experience. Therefore, the development of efficient real-time quality monitoring tools that can quantify the audio-visual experience (as perceived by the end user) is key to the success of any multimedia service or application.

Information theory can characterize one-way, non-interactive communication, where one source sends a message to one destination, very well. When the communication is interactive, as in (1) channels with feedback, or (2) two-way channels where two users exchange messages over a shared channel, much less is understood. We outline what is understood about interactive communications as in (1) and (2) from an information theoretic perspective, and why they are so challenging to characterize. Many open problems and connections to related fields will be presented.

The following problem arose from the work of Fonio et al, a group of ecologists and computer scientists, who tried to understand the behaviour of longhorn crazy ants (Paratrechina longicomis) in navigating back to their nest after gathering food. Single ants were demonstrated to be laying pheromone ‘pointers’ to be followed by groups of ants carrying large loads. Sometimes the pointers are wrong. This leads to an optimization problem on networks with a destination node (the nest). A GPS or other system selects a direction (pointer) to the nest at every node.

In this talk, I will discuss some of my recent and surprising findings on the use of hashing algorithms for large-scale estimations. Locality Sensitive Hashing (LSH) is a hugely popular algorithm for sub-linear near neighbor search. However, it turns out that fundamentally LSH is a constant time (amortized) adaptive sampler from which efficient near-neighbor search is one of the many possibilities. Our observation adds another feather in the cap for LSH. LSH offers a unique capability to do smart sampling and statistical estimations at the cost of few hash lookups.