I recently had a conversation with a friend where the subject of our summer
internships came up. My friend is interning at a financial institution and she
had great things to say about her company and the people she worked with.
However, her one major complaint was that the work she did was very specialized
and narrowly focused. She worried that if she ultimately decided to pursue a
career outside of finance, the skills she had sharpened during her internship
wouldn’t be widely applicable.
To me, this sounded like the opposite of what I felt about my time with Numenta.
As more than one of my old bosses used to say, “working in a startup is like
building the airplane as you’re heading down the runway.” In a situation like
this, roles are somewhat amorphous, projects are fast paced, and coworkers
contribute wherever they can. Furthermore, the skills and mindset developed in
this type of environment are relevant across a wide range of functions.
As I think back on the conversation I had with my friend, I realize that the
differences between her internship experience and my own also provide a strong
metaphor for comparing the Hierarchical Temporal Memory (HTM) machine
intelligence technologies we are developing at Numenta against the typical
programmed tools we are all familiar with today. Almost all of the software we
use today has been programmed to do one specific set of tasks. This enables each
individual program to perform the jobs it was designed for extremely well, but
limits the usefulness of each tool to a very limited range of use cases. That’s
why you only need one brain, but your smart phone would be pretty dumb if you
only had one app.
On the other hand, HTM machine intelligence – like your brain – comprises a
single underlying processing method that is extensible across a wide range of
applications. Over the course of this summer alone, we have used our underlying
machine intelligence technology to deliver interesting new capabilities across a
wide variety of potential applications and data sets. Here are a few of the
applications that we have been working on this summer to demonstrate the
extensibility of HTM:
Rogue Behavior Detection
Using the data generated by human actions to alert to abnormal employee actions
and detect when computers or other devices are accessed by unauthorized users.
Using streaming GPS and speed data to automatically model location and travel
patterns, then use this model to identify subsequent irregularities in speed,
location and route pattern.
Natural Language Processing
Improving the ability of machines to understand the semantic meaning and context
of natural language.
Sensorimotor Learning and Prediction
Actively exploring an environment to:
- Predict the results of motor actions and control signals, and
- Form stable and discriminative representations of each environment
Just as the tools used during an internship only represent a few of the gadgets
we each hold in our tool bags, the applications shown above only demonstrate a
thin slice of the potential that HTM machine intelligence holds. As such, I’m
really excited to see how HTM will continue to grow over the next few years to
deepen its capabilities and broaden the ways in which it impacts our world.