- 2017, September 7, Canstar, “Australian Small Businesses Confident In Cyber Security” — are you kidding me? Even though I doubt the reporting (given that it is a promo piece for MYOB), SMEs could only think they are secure because they might assume a hacker would have no interest in their data. Maybe… A better response would be: “SMEs expect to be hacked and to have some of their data breached. But, all good, at least they backed up and put some measures in place to keep customer and personal data secure.”
- 2017, September 6, SSRN, “The 7 Reasons Most Machine Learning Funds Fail (Presentation Slides)” — some of this stuff went over my head, but I think the gist is that investment managers need to re-think their forecasting practice. These modifications are then used to generate machine learning models. The most interesting idea – and the one that the author highlights as important – is the use of ‘meta labelling’. It does look like using this labelling can provide a measure of visibility about how the model makes its decisions (as opposed to a ‘black box’. There is also an interesting point about how labels are applied based on outcomes. And if you look at financial data across different time frames there are different outcomes, and therefore, different labels. I suspect ML models will struggle with such fuzzy labelling.
- 2017, September 6, SSRN, “Technology’s Continuum: Body Cameras, Data Collection, and Constitutional Searches” — this is always a tricky subject – government surveillance. My automatic response is to think – ‘hey, this is not Stalinist Russia or Stasi East Germany’; so why do we care so much about the potential for state surveillance? And if we had this surveillance, are we safer as a society? This is a completely different question to whether the body cameras serve their purpose to hold police officers accountable.
- 2017, September 6, SSRN, “A Guide to Commercial Innovation in Artificial Intelligence“
- 2017, September 6, arXiv – 1709.01921, “Distributed Deep Neural Networks over the Cloud, the Edge and End Devices” — my only question is: what about the latency between the input collection and the feedback from the model’s output? Unless the model’s computation is distributed as well? But then is that at the expense of learning in other parts of the data collection space?
- 2017, September 5, arXiv – 1709.01604, “The Unintended Consequences of Overfitting: Training Data Inference Attacks“
Now it’s Equifax’s turn … Who’s next (or, who’s left)?
Mountain Duck : a way to securely mount cloud disks as part of your normal file system. If it’s built on Cyberduck then I expect it will be a cool product. I’m about to install it…