jimduk

I use this a bit - the You Only Look Once implementation is good & fast (should be - redmon wrote the paper http://arxiv.org/abs/1506.02640) and it's all C. I don't think it's competing to overtake Torch/Caffe/Tensorflow as 'the' framework. What I find it good for is looking at the source code and trying to understand how something works in code (often not easy even after reading the papers) - easier than with the big frameworks. Also may be a good place to start for simpler target platforms e.g. embedded. Kudos to pjreddie.

TL;DR - Universities don't provide employable skills, Companies don't train, therefore people (affluent parents?) pay for training which may lead to employment

Isn't this about the difference between no job and some job? AFAIK a lot of UK students aren't finding jobs and so there are various 'schemes' where (their parents) pay for skills and then maybe they get work. I think there is still a lot of decent paid dev work in London if you can get your foot in the door. I think it's hard to get your foot in the door for your first job if you have no differentiation - this provides differentiation. Of course github projects would be better - but this is probably not targeted at HN readers.

You need a certain amount of scale to stay close to the cutting edge (mobile) of imaging - GoPro uses the Ambarella chipsets and Sony sensors - if you want to talk to them directly (as opposed to 3rd parties), or the sensor manufacturers you need a level of scale/cash/backing. Also - in general its a tough market - if you look at the classic camera market their margins aren't great

This is not just your problem. Every company that you work in has an implicit social contract (e.g. we're quirky, we're obsessed ..). Often over time the culture changes, and at some point you re-examine the social contract, realise it has changed and then underperform/ quit. [this was a massive issue in big outsourcing deals in the last 10-20 yrs where Company A would take over Company B's employees and then wonder why they became worse at doing the same job].

I would be constructively explicit early about the things you are uncomfortable about and try and articulate them to get a story going about what is changing. This way either you can adapt, or find out if others share your feelings, or you can decide if there's still a fit. Culture is key - once the culture (often the engineering culture) goes/radically changes, it signifies big changes ahead. Also - for the new people (PMs especially) - talk to them constructively to see if they can explain what they are bringing to the table, and also what they see as culturally valuable.

This is why e.g. go pro uses a hybrid approach with an embedded rtos for the camera stuff and then virtualized Linux on top of this for less critical work where you want to use a stack e.g.http streaming. 3 yrs back I worked in a team which looked at booting Linux for a camera device in < 100ms and couldn't find a solution.

One comment from experience Slide 21 "The mobile supply chain dominates all tech" / Flood of smartphone components - Lego for technology

    The shiny Lego is only available for the major players. 

For the current key components - GPU/CPU/Camera sensor - you can't order them/get support/get docs unless you have scale or amazing connections. If you are a hardware startup your lego is 2/3 yrs behind the big players, and behind public perception.

This makes complete sense once you look at chip fab costs/profit models and the industry structure, but is not great for disruption from smaller players.

NB this was from a European perspective of doing things officially - it's possible there might be more 'unofficial' components and support if this is done in China with strong local support

I did a quick review of agencies 2 yrs back to use as we started to run out of organic references in London - Oxford Knight, Burns Sheehan, Recworks (Java and grads) seemed decent. Stack Overflow is also worth using carefully

What could also be valuable is having support accounts, escalation paths and relevant implementation docs/software stacks from the big CPU/Sensor/GPU/Bluetooth/Wireless companies. Getting information/support from these big companies if you have no track record can be excruciatingly inefficient.

Life and Fate - Vasily Grossman

Russian epic following multiple characters through WW2.

It shows how serious, good people can have very different and legitimate views on life; It characterises some eternal types of criminal or charlatan. It reminds you about fate and about how lucky we are now.

Aren't Zisserman and Simonyan in the Visual Geometry Group - which is part of Engineering - see http://www.robots.ox.ac.uk/~vgg/people.html.

Interesting paper. Any more details on the architecture of the feedback connections. Also I can't tell from the paper where and how weights are being updated e..g what does "train" mean in this context

Neophyte question - is FP necessary for a CNN as opposed to using 32 bits on some fixed encoding. If I used a 32 bit value on the order of +- 4.27 or +- 8.23, wouldn't that have enough accuracy. I'm assuming the weights and parameters don't go much above say 8 or 16 after all the Relu stuff.

Brilliant, though could you add Licking Summit ?