Well even though the trip started out with lots of promising projects it to take part in, the reality is you can only focus on so many. The three projects I stayed involved with were the Sensory Fusion “BabyBot”, Fast Card Identification, and the SMD USB Board project.
The USB board needs some attention. It was all going OK until the IC moved when I was soldering it. Consequently a track came off the board. I think it can be fixed by using a wire directly from one pin to the board, but I’ll need the lab facilities to continue.
Fast Card Identification for the most part is finished in APRON, The plan is in the last couple of days for Piotr to attempt it in Log Mode SCAMP. The Silicon Retina does a reasonable job, and thrives on the high speed. Obviously, we’re going to need several halogen lamps.
The BabyBot is going well, at least it’s not being slowed down by APRON! The relationships between different sensors and actuators are being learnt quite nicley, however, it does seem that people can’t get used to programming with array constraints. I’ve not really been allowed to get stuck in on any level other than theoretical. The next couple of days will see the addition of the silicon cochleas. Good, because Sanchez Da Mambo Bot can’t hear his moraccas yet…
So I’m contiuoing with APRON support for the baby-bot. The arm is now under some sort of control.
I also seem to have fallen into the role of “playing card identification” for the spiking robot project. This has been quite a challenging task, as they wanted to do it in a contrived environment, and I said I could do it from natural scenes! Anyway I now have an algorithm that can count the symbols on a card. This is quite tricky as black and red cards are very different to my camera. Now I can identify where the symbols are, I should be able to extract a single symbol and classify it.
So where are the blog posts? Well the truth is, not much had been going on! The schedule has been tightly packed with lots of talks, up to 12 hours worth per day. This has not left much time for any actual work. The talks have all been interesting, but many of them have been quite basic. Hopefully next week they’ll pick up a bit.
The most interesting talk so far was about bayesian belief probability. It’s a nice way of determining confidence in the information returned by sensors, and even multiple sensors combined.
APRON has been adopted as the platform of choice for the sensory fusion group, and a small army of APRON programmers is currently in training. The boss has forbidden me from actually doing any of the programming for the group.
I’m going to be implementing an APRON driver for the deathstar! This is a ball of 64 audio sensors and 5 cameras. I could not believe it when I discovered an XEM3001 at its heart. These guys have done a really good job of managing consistent data rates.
Another APRON driver will be to the Cappo Caccia arm. Now I know Kevin already has this, but I’m going to see if I can do it without the need for yarp.
One guy is working on implementing optic flow using APRON, and has ambitions of using CUDA. Another more odd project, is to use APRON for auditory sensor processing.
I’m working on a playing card recognition algorithm, which would be a nice demo when visitors come to the lab, and an auto bias controller to throttle events coming from the silicon retina. I’ll probably also build some APRON plugins for easy access to 3d arrays, using Alex’s parsing algorithm as a base, and also some rudimentary drawing library to annotate registers more easily.
So, lots of work to do, but given we don’t have any time to actually eat properly, I don’t know when it will all get done. We have a day off on Sunday for the 4th july.
Had a late night coding session last night. The multi modal sensory fusion guys have decided to adopt APRON as their implementation platform, so I had to make sure that it can do all the things we claimed it could.
In particular, this involved a complete overhaul of the plastic connectivity structures used in neural network learning. The libraries are nice now, and externalise the learning rule. Part of the project is to investigate different learning rules.
On top of this, I had to prepare a tutorial which I’m giving tonight. This project will stretch every aspect of APRON from network comms, to GPU acceleration.
So no pictures today. I’m slowly deciding which projects are going ot be interetsing for me. The first is the “BabyBot”, which will utilise topographic neural nets to fuse together data from multiple sensors. The plan is for the baby to become aware of its own environment, under the guidance of a mother entity. The robot has an arm, with a touch sensor, and a stereo head and neck. The aim of the project is to develop maps that fuse together input form both all the sensors.
The second project is building a system that can play the card game “Hearts”. The “Robot” will use vision and audition to identify the cards that are in front of it. This sensory information is then given to a cognition system. Two main streams of this project will involve a spike based implementation versus a non-spiking system. I think I’m going to have a go at implementing the card recognition in APRON.
Speaking of APRON, there are several collaborations on the go. The first is a project that will implement optic flow, as a prototype array processing solution to be later implemented with much higher dimensionality using CUDA. The aim is to process megapixel images, which I believe is quite ambitious (read challenging, borderline impossible).
I have to start preparing APRON tutorials for next tuesday. Progress on the new SCAMP box is ongoing, but proving to be really stable. For example, you can switch between algorithms many times a second without anything crashing!!!
Started to think about a dedicated APRON processor again. I envisage a digital setup with an array processor of the dimensions 1000×5. This will work alongside a sequential processor, that has access to the array memory as well. All a bit pie in the sky at the moment but it seems that there is a demand for a credit-card sized processing system that can be fitted with attachments such as vision sensors. I foresee such a devcice being of much use to the neuromorphic community, and possibly spin-off opportunities…
)