eCortex – What Could Computers Do If They Could See?

One of my oldest (as in “known for the longest time”) friends – Dave Jilk – recently started a new company called eCortex with his business partner, CU Boulder Professor Randall O’Reilly.  I – along with several other friends and The University of Colorado (through their Proof-of-Concept Investment program)  – participated in the angel financing.

I’ve known Dave since my second day at MIT.  I was a lost and lonely freshman during the madness known as MIT Rush (MIT used to have fraternity rush the first week of school – this has since been modified for a variety of reasons – some good, some not-so-good) when Dave walked up to me and said something like “hey, want a beer?” (this was probably one of the “good” reasons since I was 17.) This started a 20+ year friendship that includes starting a successful company together (Feld Technologies), marriages and divorces, moves (we both ended up in Boulder), an unsuccessful entrepreneurial adventure (Wideforce Systems), several salvage operations (planetU, Xaffire) and lots of time spent exploring and learning new things. Dave was always one of the most intellectually capable people in the room (a 4.9/5.0 in Course 6 at MIT means something – I think the B was in psychology) and – while the startup company thing suited him well, he was never in love with the thing he was working to create.

Several years ago, Dave started taking some Cognitive Science classes at CU Boulder and hooked up with Randall O’Reilly. I remember seeing him light up one of the first times we talked about the research he was exploring and thought to myself – “this is it – he’s found what he wants to work on.”  Last summer when he was visiting me in Alaska we talked about this idea a lot and I continued to encourage him to just go for it.

He did.  Dave and Randy have started a new software company called eCortex to commercialize the visual object recognition technology that Randy and his team have been working on at CU Boulder.  The technology is essentially a neural network model of the human visual system, including numerous biologically realistic characteristics, which allows it to see things much like we see them. Applications include surveillance (security cameras as well as luggage scanning), satellite photo analysis, military target identification, online image or video search, optical character recognition, and manufacturing. In one test, this system was trained on a sampling of 100 different objects, and successfully recognized the objects 96% of the time, even when shifted in the image, scaled down to as small as one-tenth the size, and rotated within a small range.

Now – there have been many attempts to create object recognition systems.  While I don’t understand any of the science that Dave and Randy are working on, I do understand that much of what exists is basic pattern recognition algorithms rather than the magic of neural network simulations.  I sat in a conference room recently and watched as Randy walked me through some examples using PDP++ (a neural network simulation system written in C++) explaining how eCortex is approaching the problem differently.

A few weeks later I sat in the same conference room and listened as a visiting VC told me that he’d heard there are no real computer science people in Boulder.  I choked down my chai and mentioned a number of examples, including these two dudes working on this neural network thingy.

  • Brad,

    It’s very difficult to comment on specific cases, without knowing the details. The following are general comments, and may well not apply to what your friends are doing.

    Firstly, “the magic of neural networks”. In general, there are actually big problems using neural networks for machine-learning. Except in rather specific cases (e.g. probabilistic neural networks), neural nets tend not to have much general predictive power. Rather, they tend to simply learn the data sets they are trained on. Hence, it’s typical to obtain good results only in test cases, and much less good results in the real world. In the right hands, other approaches such as support vector machines, often give better results.

    Secondly, visual object recognition. You’re right that much of what exists out there is pretty basic pattern recognition. However, the thing that many people don’t realise is that the reason why computer vision is a hard general problem is that there fundamentally isn’t enough information in many images to be able to identify objects. Human perception of objects is a complex area. In computer vision, if the only information you’re using to identify objects in an image is information extracted from the image itself, you are *way* behind the state-of-the-art.

  • Dave Jilk


    Your first comment is very true of backpropagation networks, but we don’t use backpropagation – in fact, the learning rule(s) we use are one of the key innovations in the technology, and they have essentially the best features of both error-driven learning and associative learning.

    On your second comment, obviously there is enough information in photographs to identify objects – given appropriate prior knowledge, since humans are capable of doing it. The key is: what is that prior knowledge and how does it need to be organized? The cool thing about our technology is that we don’t need to analyze the form of this knowledge in detail, the network does it for us – just like in the brain.

    Some recent innovations in algorithmic machine vision are our best (academic) competition, and they have essentially analyzed and replicated some of these early stage visual representations and thrown an SVM on top of them. This actually does quite a good job, but it’s fundamentally limited in ways I’m not going to get into.


  • mbt mens sandals suit for the person who like running over walking, so covering the distances required in crowds is frustrating and uncomfortable, and standing for long periods is absolutely painful. Whether the mbt kisumu black sandal, for health, I think the mbt shoes may be overlooked mbt women's sandals still despite what I’d consider to be tangible benefits.Try to be the best, to see whether mbt chapa azul are comfortable and relaxed.mbt chapa black sale are the highest daily percentage of people wearing a shoe.MBT chapa caviar black is not just a good quality of beautiful

  • Pingback: cheap auto insurance nyc()

  • Pingback: car insurance comparisons()

  • Pingback: cheap auto insurance in va()