I know, I know it’s been a while since I posted on my blog, naughty Folknology… I am contemplating the idea of switching my blog to Medium or even creating a static blog on Gitlab with the various projects I am working on, let me know if you have any experiences/opinions with these options. I could of course just keep things as they are…
But I guess I ought to let you know what I am currently working on project wise as it’s been a while. Although I am still working on robotics projects based around Xmos technology which I need to pay the mortgage and school fees etc.. I have also been doing a lot of research over the last few years into machine learning (ML). Although this isn’t my first foray into this it’s been a couple of decades since my last exploration in the 90s.
In case you haven’t noticed machine learning having lived through a serious A.I. winter has emerged refreshed and with renewed energy over the last half decade. This time around it has found real applications in big data and the major cloud players Google, Amazon, Facebook, Microsoft et al are now using machine learning on very large scales. It will not be long before the machine learning process time of most cloud infrastructures from these vendors will pass from minority into majority, as these applications are consuming exponentially more bandwidth and processing resources. In order to get higher performance densities. We also see these vendors employing the latest GPU technologies from Nvida (Pascal) in order to try to keep up with the extraordinary demands of these algorithms and data. Recently Google has even developed there own silicon (TPU) to help run these machine learnt inference platforms. Counter intuitively these new processors are actually integer rather than floating point based in order to try to contain the performance per watt targets of these deployments into manageable envelopes. The killer bottleneck for all of these vendors is performance per watt for specific machine learning and inference applications and we are likely to see a great deal of research and development in this area.
Intel will likely be placing more inside the chip package alongside their venerable Von Neumann Architectures (VNA) in order to try to cut Nvidia off at the pass so to speak. We can expect them to integrate very high bandwidth DRAM and Interconnects with massive FPGA inside single chip packages in an attempt to reduce the power issues with their architectures in this critical application market place for them. IBM is moving in a radical direction with their Truenorth asics using spike based neurosynaptic corelets. Along with these big players expect many others to bring rapid innovation to satisfy both the big data players using machine learning and the emerging inference based embedded and robotics markets over the next decade.
The latter is of course is where my research and interest has been focused, I am working on hardware applications based around both hybrid embedded (VMA + VLIW cores) and emergent FPGA inference + VMA learning along with a number of higher level tools like Theano , Scikit-learn and others for the data analysis,ML and modelling etc.. Therefore you will likely see me posting within these areas at all different levels over the next few months, it’s good to be back and it’s great to live in such exciting times….