We recently began a *Visiting Scholar
Program*,

where researchers and professors can spend some time at our offices and

learn about HTM in depth. Designed to promote collaboration,

participants play an active part in our research meetings while

continuing their normal research. To give you a better understanding of

this new program, I interviewed our first visiting researcher, Mirko

Klukas, and asked him about his time spent at Numenta.

Mirko Klukas, Ph.D.

## Hi Mirko, can you introduce yourself and at a high level, your area of research and expertise?

My name is Mirko Klukas and I am a post-doctoral researcher at the

*Institute of Science and Technology Austria*. The

core of my current research is a combination of mathematics and computer

science, including topological data analysis and computational topology.

My background lies in pure mathematics in the field of geometric and

differential topology. Believe me, it sounds worse than it actually is. 🙂

## How did you learn about Numenta?

In 2013 I was on a vacation in Sweden, in a typical Swedish country

house with a pretty unique set-up: no running water, but WIFI. I read

the first Numenta white paper on that vacation – basically between

chopping wood and keeping the fire burning – and kept thinking about the

approach ever since. How the paper actually did end up in my hands in

the first place? I have no clue.

At my current position I enjoy a lot of freedom when it comes to my

research, and I devoted a significant amount of time to think about the

approach from the paper and related questions. Last year I reached out

to Subutai to see if there is any interest in my ideas.

## What was a typical day (or week) like for you during your time here?

It felt very familiar, the atmosphere was very calm and relaxed, and a

typical day didn’t differ too much from my usual research days with two

exceptions. I could discuss things with the team, and I tried to do that

as often as possible. This could happen in smaller one-on-one meetings

using a whiteboard, or by giving a talk on my ideas to more members of

the team. In addition I had the chance to attend basically all regular

meetings, non-research meetings included. Personally I found that to be

very interesting as it gave me an idea about the operations side and the

other facets of the company.

## What did you hope to get out of it?

The part of my research that is related to Numenta’s technology is only

very loosely connected to my mathematical research and it was great to

have a counterpart to bounce ideas back and forth. It was great to talk

and pitch ideas to people that speak the same language, and I am pretty

confident and excited that down the line this collaboration will result

in a publication.

## What was most memorable or valuable to you?

Numenta finds a nice balance between its research mission and being

product driven. The atmosphere and people at Numenta reflect that. In my

experience the two ends of this spectrum usually dance to a very

different beat, and it was great to experience a research environment

outside of academia, where I felt comfortable and could imagine myself

working. I felt really welcome, which I value a lot. Being able to

present my ideas to Jeff and the rest of team was a unique experience

as well.

## Was there anything you learned that you could not have learned without being here?

It was inspiring to be at the cutting edge of Numenta’s current

research, and it opened up new perspectives on previous ideas. Looking

over someone’s shoulders when they are in the process of solving a

problem is something that shouldn’t be undervalued. Talking to Jeff,

Subutai, and the rest of the team, and getting their perspectives is a

great asset. In particular, I appreciated getting a glimpse on where the

individual priorities and emphases lie.

## How will your time at Numenta shape your future work?

As mentioned earlier, I hope my time at Numenta wasn’t an isolated

event, and my visit was just the starting point for many future

discussions to come. It also encouraged me to pursue the direction

towards machine intelligence research even more.

## Can you describe for more technical readers the area of your specific research?

I will spare you a detailed explanation of my mathematical research. 🙂

If someone is interested in that, all my pure math papers can be found

online on

*arXiv* – some

keywords are: *low-dimensional topology*, *contact* and *symplectic
topology*,

*symplectic cobordisms*,

*open book decompositions*,

*Engel*

structures.

structures

My current research focus in the machine learning and intelligence realm

revolves around two main directions.

How can you generalize sequence learning approaches for discrete

alphabets, e.g. variable markov order models like *context tree
weighting* or

*prediction by partial match*, to non-discrete input

spaces, namely constant weight codes, or sparse distributed

representations respectively? This can be done, for instance, for the

*incremental parsing algorithm*, the main ingredient in the classical

1978 compression algorithm after Lempel and Ziv. The result is a

recurrent neural network structure very similar to Numenta’s sequence

memory. This construction could serve as a reference point to highlight

or contrast certain aspects of Numenta’s sequence memory, e.g. it could

help to proof theoretical bounds on its prediction performance.

The second focus revolves around *binary sparse coding,* and Numenta’s

spatial pooling algorithm in particular. The algorithm is closely

related to the *witness complex*, a construction of simplicial complexes

well-known in computational topology – you find a small note about the

connection of the spatial pooler and the witness complex on my blog:

*Maps, Functions and
Arrows*.

Generally I try to establish a meaningful role of entropy within binary

sparse coding and contrast it to other binary coding approaches. This is

loosely related to

*expressivity*of neural networks, and plays an

interesting part in the objective of the spatial pooler. The objective

balances a good

*reconstruction error*and maximal

*mean individual*

entropy. The interplay of these two quantities is something I want to understand better as well.

entropy

*If you’re interested in applying for the Visiting Scholar Program, [*click here*](/company/careers-and-team/careers/visiting-scholar-program/) to apply*.