Brandon's Notes

Intro:
In CS department, informatics; clinical faculty member
Co-op with Intel, IU in 1998, clinical faculty; looking for design; teach operating systems, 11 years building design/development; Configured Analog Neural Network; (CANN)

How is clinical faculty at the college?
Built the position for clinical faculty at IU, did not exist prior to him. Helped that he was director of IT previously to build up the area

Opened up options for others?
Many in medical fields… less common in other fields…

Does working with analog systems overlap with other areas? Such as biology and chemistry or healthcare?
Opportunities in learning new domain and applying a new domain. To build Extended Analog systems applied to hearing aid…ended in neural prosthetic field. Works better of you help other to utilizes the technology. Biological sciences fertile ground for transferring tech

Creating new tech? Working together or is it just building for a new client?
Most of the time they can’t give you a requirement in the way science sees it… more so as they see it. Unsure about constraints and variables. Must work closely to develop a way to solve problem around such variables important to bridge the gap

Extended analog computer
Configured analog neural network:
In general:
Weighs options for you. In order to classify by shape. Noisy signals and power requirements. Can use EAC to process these signals…built it into a match filter; EAC that attached to host thru USB used the conductive foam as substrate to do voltage…. To develop lookup table of injected voltage on the foam. A technology in search of a problem. (Dangerous place to be) took 4-5 generations to get where they are now…

Feed forward…the way EAC works with voltage is recurrent…isn’t recurrent and not feed forward?
Possible to get recurrence… they hang out there/not fleshed out. Project the currents get results in voltage manifold and that is static until that changes. Future input not intend on output. If you want recurrence you can get it; resolved to the construction to analog DAC (digital to Analog Converter)

No marketed purpose yet…what drove its construction vs nanotech? Why use it?
Looking for a way to build a match filter and commercialize it. Where can we get advantage with an analog computer over a digital one? Biotech did not use nanotech used DSPs… going to classify signals by how they go down the wire and the shape of that signal. Looking for a low power way to do this. Expensive to used digital system. Analog can wake up when finds what is being looked for. Looked for complex signature of variable to wake up someone with the correct amount of power.

In training the EAC/CAN to learn the network, any existing networks to compare your own networks against to see how they perform?
Very small network, larger networks are expensive to run power wise (ie voice) can do it well. Their (hearing) market for things that run off of tiny amounts of power (50 microwatts)
Inhibitive to restrict the network? (Dimensions/size)
Trying to minimize networks size to meet power constraints. Looking at nerves to work in that area and trying to expand the business to other areas. Looked at processing audio, (ie Texas Instruments). Network inhibited by ability to process signals but enough to attract other areas

How does backpropagation work in extended analog computer?
Trained offline, do it digitally, took signals, made training set (few thousand signals) trained against those, digital micro controller (online).
Analog beats digital where power is an issue; analog does low power- difficult for digital systems.

Combine analog and digital to get both?
Best of both worlds; analog front end as a wake up device; watches the world, the microcontroller can then wake up and take action. Have an external signal to wake up controller –cannot be too discriminatory. The analog can be very selective to wake up (too noisy for a digital system it would never sleep). Hybrid system great way to go

Tech before problem? Why do the research in analog in the first place?
Professor Mills was doing research in it already. Original think of plug boards to model the equations that one would express. Early on was building interesting systems (as engineering background). Came from mills building these systems and Bryce wanting to build them.
microEAC

Is a digital system at its heart… a large matrix of circuits.. Analog cells, current output or voltage input, has a microcontroller, hookup for USB and other converters. Where analog blocks is there is foam. Current inputs to sheet, and used currents to project manifold of voltage to sheet and get the answer of manifold off of the sheet.
(Combinational logic blocks, programmable data arrays)

How to determine the substrate for this project?
To generate currents that could generate potentials on sheet. Needed a higher resistance so projected currents they could be easily read. Dictated by needs of the construction of the device. In this case to show proof of principle. Wanted to use silicon. Came down to what he could actually build.

Using pins that moved around?
Have a sea of analog cells, want to set the weight on the analog cell, gain? Position? Look using SWARM to identify how to solve the particular problem. Used swarm to synthesis and map all at once. Better not to move things around all at once. Better to have the DACs to wait. Moved away from doing the SWARM

Consider yourself computer scientist/software? Or more engineer?
Works in a category called systems. Lives at operating systems and below. Can include hardware, virtual memory systems/file systems. Does not work in domains of databases and compiler performance.

Share experience..work for big? Small? Startup?
Intel –tech support, co-op. allowed him to interact with lots of people. Met future boss thru Intel. Can make amazing contacts, good way to make connections. Cummins-learned process, good at engineering process. How to solve problems in regular way. Documentation. Hard at college level…not lifetime of code. Process taught how to deal with documenting for long periods of time. Managed changed, (integrated circuit packets) process of doing all of that (tests, reviews, documentation). Working at IU like a small business. Need to be a jack of all trades. You’re going to do a lot of the work. Don’t have as much time to learn. Have a special focus to do that right out of school, cannot drift from whatever the goal may be. Have the right focus; things big company can offer are around the ability to teach about an industry, learning on their dime.

Clarify less time to learn at a small business, need to be more focused?
Tasked with something to learn a new language. In small business 1st milestone going to be very quick. Going to have to reach that milestone quickly to get commercial feedback to learn mistakes and fix them. Anything that distracts form that objective must be discarded. Have to learn really fast in small business. Going to get exposed to business realities more so than in big businesses.

Using DACs as weights between nodes instead of nodes on the foam?
The weights are done by DACs, a way to output a current based on digital information. The analog signal comes along and is multiplied by the DAC. Sheet was acting as the multiplier…

Does the digital to analog slow down the system?
The weights are not dynamic. The analog signal flows thru. The conversation done at transistor

Industry application of EAC? Best route to acquire job related to EAC?
Jobs that are related to EAC are low power and complex signal processing task. Ie. Hearing aids.

Circuitry training to work with analog? Go to analog computing from software perspective??
Been exposed to machine learning class… in that area that’s where a programmer comes at these issues in that way. How to understand algorithms of the training comes from programmer background. More interesting problems is where system changes over time- CS type of issues

Do you need hardware knowledge to get into it?
Like any other domain, come at it with perspective how trained with past experience. Will be working on a team with hardware engineers. Don’t have the experience to do everything that is necessary to use the chip. But would still be able to collaborate with others in building this system.

Common mistake that students have when getting a job/going into the professional world? In job hunting/weighing options?
Not having any experience at all going out of college. Helps narrow down what you don’t want to do. Get some kind of internship. Helps refine options, and discuss real world projects with prospective employers. Keep an open mind about the jobs they ask you do perform, look at everything as an opportunity for experience, a lot to learn for the future. Ride out the job for a few years to see what can be learned.

After hours:
What caused to fail?
Get something out there to learn from. Had a difficult time finding that partner. If profit cannot be returned in time, will not be funded… need to be profitable return on money in order to stay afloat. Much like Sharktank- start on the road to make money. Was a matter of tracking investment; can’t pause in business, need funding to survive.

Was problem from marketing or what it can do? Difficult to explain brilliance?
Most problems already have solution, need to make your offer sweeter than the old product. Solves the same problem with better performance. If how that will happen is not clear will not be used in market. Give them a number of how much money it’ll save or make them.

Problem was so inventive needed a better example to demo to customers?
Always better to have a walk around demo to show people how it can work; show it is much better more power efficient. Viable prototype is much better.