Assistant Professor
Carnegie Mellon University (CMU)

Profesor Asistente. Ese será mi nuevo título, como ya habrán leído en una entrada anterior. Me quedo en Pittsburgh por unos años más, quién sabe cuántos. No había escrito nada al respecto porque para hacerlo tenía que estar completamente en paz conmigo mismo, haber asumido por completo la decisión y volver nuevamente a la rutina. Ahora que estoy listo, quiero escribir sobre dos cosas: lo que significa ser un Assitant Professor, y cómo llegué a la decisión.

Continúe leyendo…

There is an article about our electricity monitoring project on the latest issue of the Carnegie Mellon Engineering Magazine. The direct link to the PDF that contains the article is here:

Two minor things to note:

• They quoted me saying “desegregation“, when I actually said “disaggregation“. After looking up the definition of the former, it’s not awfully far.
• The article, in the third paragraph, somewhat overestimates the time we’ve spent working on this (e.g., I haven’t been here 4 years yet).

After exactly one year of making a video about our Non-Intrusive Load Monitoring (NILM) project , I finally got approval from our funding source to release it. Obviously, the current prototype is different from what is shown in the video, but nevertheless, it provides a quick overview of the main idea behind the project.

The goal of this project is to develop simple and inexpensive solutions for obtaining appliance-specific electricity consumption information from easy-to-obtain data sources like the overall power of the building. It’s the idea for which we recently received $1.5M from the National Science Foundation.

We attempt to detect and correctly identify state-transitions for individual appliances in a building (e.g., a light-bulb going from state “off” to state “on”), by carefully analyzing the overall voltage and current signals for the building’s electrical distribution system. We also explore how different statistical and machine learning tools can be applied to solve the problem.

NILM has been around for more than two decades, but mainly as a research topic. What we really want to do is to take it a step further and find out what are the obstacles that we need to overcome to bring it to homes.

I will be posting more about it now that I finally found a quick way to introduce it to all of you (i.e., the video). There are also some interesting links to the other project I am working on: Sensor Andrew.

When trying to configure the Digi Real Port daemon (a virtual serial port, or RS-232 over IP) on a Red Hat machine (RHEL Server release 5.4), I found that installing form the RPM did not work properly. Some kernel module was not loaded.

After doing a little research and tweaking, I solved it by compiling it from source, and following the steps below:

• 1) Modify your configure.in file, by adding:
  KERNEL_HEADERS=/usr/src/kernels/your_current_kernel
  And substituting “your_current_kernel” with yours (e.g. 2.6.18-164.11.1.el5xen-x86_64).
• 2) Place that line of code, at around line 405, right before:
  if test -z "$KERNEL_HEADERS";
• 3) From the shell, run the following to complete the installation:
  autoconf
./configure
make
sudo make install

Después de algunas súplicas y averiguaciones sobre los derechos de autor, conseguí que me hicieran llegar el vídeo de la charla que di en IBM hace dos meses. Aquí está una parte:

Algunas notas:
1) ¡Qué largo! Ya entiendo por qué estaba tan sediento al terminar la charla.
2) Me impresiona lo mucho que siento que he mejorado el inglés después de tres años aquí.
3) Hay muchas cosas que quiero mejorar la próxima vez que tenga que hacer esta presentación. Por ejemplo, muchos de los “slides” van en contra de mi política actual que limita el número de palabras en cada uno a aproximadamente 15 ó 20.

One of the databases I’m using for my thesis project is storing a series of data-streams that need micro-second or even finer grained precision for the timestamps. Since I’m using a MySQL database, it seems that there is no native support for these type of timestamps (if this is not the case anymore, please comment!), and the person who designed the schema decided to use two fields to describe the timestamps:

• dataTimeStamp, as a DATETIME type.
• fractionalTimeStamp, as a DECIMAL(6,6) type.

The first attribute would store the timeStamp with a precision of up to a second, and then the other attribute would store the fractional values. This clearly works, but is far from being optimal. Specially when you have a huge dataset (days or weeks of continuous data) and want to extract a specific region in time.

As a quick solution, I decided to consolidate those two fields into a single BIGINT type column (I’m using field, column and attribute interchangeably here). The new column would contain a 64-bit integer with the first 32-bits used for a traditional UNIX timestamp (number of seconds since 00:00:00 1/1/1970, also known as the epoch), and the other 32-bits would be used for the fractional part of the timestamp, only this time in nano-seconds (10^-9).

To do this conversion, I’m thinking about using something like:

UPDATE power_data
SET timeStamp =
(CAST(UNIX_TIMESTAMP(dataTimeStamp) as UNSIGNED) << 32 | CAST(fractionalTimeStamp * pow(10,9) AS UNSIGNED));

I'm still not sure if this is such a great idea. On the positive side, this new column should sort the table well (instead of having to sort by the two previous columns). On the negative side, I would still have to convert those numbers back into meaningful dates whenever I want to use them.

Another option would be to use the full 64 bits to encode the number of nano-seconds since the epoch. This would be OK for the next 500+ years, since 2^64 is approximately 1.85 x 10^19, and the number of nano-seconds since the epoch on the year 2510 is approximately 1.57 x 10^19. This approach would also sort well, and would need a less complicated parsing process in the end.

What do you think? I think I'm going to go with the second option.

This is the first post in English in this blog. Let’s see if it works out.

Sensor Andrew, one of the projects I’ve been working on for the past two years, is an experiment on the interoperability and semantic capabilities of cyber-physical systems. In essence, it’s a collection of hardware and software elements that together form a virtual instrument for large-scale sensing and actuation. But I don’t want to spend time explaining it. The interested reader can refer to our webpage or the techincal report.

What I do want to talk about is the relative ease with which new applications, that make use of the available sensor data, can be created.

Continúe leyendo…

Mis asesores me han sugerido intentar graduarme en Mayo de este año, en vez de Agosto como tenía planeado. Para por lo menos intentarlo, decidí no tomar clases este semestre y concentrarme sólo en la tesis. Pero parece que una parte de mí no está contenta con la decisión de no aprovechar al máximo las ofertas que había para este semestre, y he decidido comprar una serie de libros que, aunque sé que no leeré por completo, sirven de solaz.

Tenía US$400 que la universidad me había dado para estos fines, así que decidí gastarlos todos de un solo tiro. Aquí está el carrito de compras:

• Spikes: Exploring the Neural Code (Computational Neuroscience) – Fred Rieke; Hardcover
• Handbook of Statistical Analysis and Data Mining Applications – Robert Nisbet; Hardcover
• The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Second Edition (Springer Series in Statistics) – Trevor Hastie; Hardcover
• All of Statistics: A Concise Course in Statistical Inference (Springer Texts in Statistics) – Larry Wasserman; Hardcover
• Introduction to Linear Algebra, Fourth Edition – Gilbert Strang; Hardcover
• All of Nonparametric Statistics (Springer Texts in Statistics) – Larry Wasserman; Hardcover

El mismo Wasserman está dando la clase de “Statistical Machine Learning” aquí.

Parece mentira pero ya el año que viene, si todo sale como espero, estaré defendiendo la tesis doctoral.

El “Dr. Mario Bergés”. Bueno…

El caso es que ahora toca pensar en cuál será el próximo paso. De hecho, como los americanos son muy buenos en eso de hacer planes a largo plazo, hace tiempo que debí estar pensando en cuál será el próximo paso. Las solicitudes de empleo para una posición de profesor de universidad se hacen con 8-12 meses de anticipación. O sea, que si quiero solicitar para el semestre de otoño 2010, ya estoy un poco tarde. Pero ese no es el único camino.
Continúe leyendo…

El lunes, un día después de llegar de vacaciones en Rep. Dominicana, salí para el laboratorio de investigación T.J. Watson de IBM. Acompañaba al jefe del departamento de Ing. Civil en la universidad a dar una charla sobre Sensor Andrew, uno de los proyectos en los que he estado trabajando aquí.

El complejo está repartido en varias zonas al norte de la isla de Manhattan, en NY. La edificación en la que nos tocó dar la charla y pasarnos el día, queda en Yorktown Heights, y aunque fue construida en los 60s, la arquitectura sigue siendo moderna e imponente:

Continúe leyendo…