Join devRant
Do all the things like
++ or -- rants, post your own rants, comment on others' rants and build your customized dev avatar
Sign Up
Pipeless API
From the creators of devRant, Pipeless lets you power real-time personalized recommendations and activity feeds using a simple API
Learn More
-
On reccomendation of @chabad360 I made this its own rant.
I switched from marketing to CS (complete with a three year degree, no Bootcamp). I still went to interviews as you'd expect a marketing man to go; in a suit. Commence the weirdest interview.
$I: interviewer
$M: Me
I: "You're not the typical engineer. Can you talk to real engineers?"
M: "could you elaborate"
I:"you're dressed in a suit. That leads me to think you're a MS user. Do you think you could talk to real, ie. Linux using engineers?"
M: " well, I haven't used windows in about a year soo..
I: "Mac isn't Linux."
M: "I'm aware. I've switched to Ubuntu so I could use KVM-QEMU android emulator with GPU pass through to train Deep Convolution Networks on mobile devices. Also had to compile Google's internal build tool because it had bugs I had to fix so I could compile the APK."
I: "ah, Ubuntu eh? **Insert Smirk** How about a follow-up?"
M: "no, I'm switching to Gentoo this week and would like to talk to real engineers about that."
I thanked him for the coffee and left.1 -
Holy duck, I lost two days on a convolutional autoencoder splitted in two separate neural networks to encode and decode separately, it reconstruction had some strange behaviours. I was giving as input an image and then saving the encoded compressed representation in a new image, in this way I could decode it with the decoder whenever I want saving space.
How much retarded am I?
The internal layer's weights hadn't constraints so in learning phase the convolutional filters can contain any number, positive > 255 or even negative and I cannot save it in a new image as they are so they were clipped automatically between 0 and 255 with an huge information loss.
It's so frustrating when you rewrite the code in any possible way, you obtain the same wrong result and then you realize that was a borderline behaviour of a third part library.1 -
I had the idea that part of the problem of NN and ML research is we all use the same standard loss and nonlinear functions. In theory most NN architectures are universal aproximators. But theres a big gap between symbolic and numeric computation.
But some of our bigger leaps in improvement weren't just from new architectures, but entire new approaches to how data is transformed, and how we calculate loss, for example KL divergence.
And it occured to me all we really need is training/test/validation data and with the right approach we can let the system discover the architecture (been done before), but also the nonlinear and loss functions itself, and see what pops out the other side as a result.
If a network can instrument its own code as it were, maybe it'd find new and useful nonlinear functions and losses. Networks wouldn't just specificy a conv layer here, or a maxpool there, but derive implementations of these all on their own.
More importantly with a little pruning, we could even use successful examples for bootstrapping smaller more efficient algorithms, all within the graph itself, and use genetic algorithms to mix and match nodes at training time to discover what works or doesn't, or do training, testing, and validation in batches, to anneal a network in the correct direction.
By generating variations of successful nodes and graphs, and using substitution, we can use comparison to minimize error (for some measure of error over accuracy and precision), and select the best graph variations, without strictly having to do much point mutation within any given node, minimizing deleterious effects, sort of like how gene expression leads to unexpected but fitness-improving results for an entire organism, while point-mutations typically cause disease.
It might seem like this wouldn't work out the gate, just on the basis of intuition, but I think the benefit of working through node substitutions or entire subgraph substitution, is that we can check test/validation loss before training is even complete.
If we train a network to specify a known loss, we can even have that evaluate the networks themselves, and run variations on our network loss node to find better losses during training time, and at some point let nodes refer to these same loss calculation graphs, within themselves, switching between them dynamically..via variation and substitution.
I could even invision probabilistic lists of jump addresses, or mappings of value ranges to jump addresses, or having await() style opcodes on some nodes that upon being encountered, queue-up ticks from upstream nodes whose calculations the await()ed node relies on, to do things like emergent convolution.
I've written all the classes and started on the interpreter itself, just a few things that need fleshed out now.
Heres my shitty little partial sketch of the opcodes and ideas.
https://pastebin.com/5yDTaApS
I think I'll teach it to do convolution, color recognition, maybe try mnist, or teach it step by step how to do sequence masking and prediction, dunno yet.6 -
Walking past a conversation with people and over hearing the term CNN, thinking its about Convolution Neural Networking. Long story short. Im standing here between people discussing news resources. Fml
-
Figure I can simplify the code if I have the compiler handle *some* of the register allocation.
Eh? What do you mean "NP-hard"? Dafuq's an ENN-PEE?
**frantically reads wiki**
I can proudly say that I understood absolutely nothing; CS stands for cocksucker or rather abysmal failure at the most basic forms of communication, I don't just sit here all day expecting you to flawlessly prove my point with every swallow of breath you draw, yet here we are.
Perhaps one factor involved in producing the generalized cluelessness of my colleagues, I mean their "imposter s*ndrome", has a bit to do with how fucking thick you've formulated this glorified bollocks you call theory. Were not for your incompetence, arcane crackheads like me would simply __not__ be capable of rising to the top of this field entirely via determination and a big salami, therefore I owe you both a debt of gratitude as well as every last word and sign of total disrespect.
As interesting as the study of computational complexity can be, if done correctly that is, you idiots are stuck in a mathematician's abstract mindset in a field entirely devoted to application of ideas rather than *just* the ideas themselves.
To answer my own question, it means there's no known efficient solution. That's it. The part about nondeterministic polynomial convolution of an irreductible rectosigmoid junction can apparently be skipped altogether. Anyway, I solved the problem with the computational equivalent of pizza sticks while you were out in the field mentally jacking off to λ.
Lecture is over, now go clean up the ethereal masturbatory residue if you will, I have mystical el Khwarizmi type-shit to solve via further clubbing of abstraction through liverwurst bologna of immense proportions. ^D3 -
I have to implement DSP using Convolution, FFT. And for the past 2 days I have like a million tabs open in Chrome and half of the titles end with ".... for dummies" :D
At least I'm learning something new and when I understand what the math behind those terms does I feel awesome :D -
Didn't really know how to categorize, bit of a question/discussion/curiosity, so I put it here.🤷
Just today I read an article that stated about the Netherlands, where the police will use an "AI surveillance camera" (yey buzzwords incoming 🙄, but it would actually make sense(?)🤷) to detect and punish drivers, holding a smartphone. Pictures without smartphone shall be deleted. How would this system work without having non-smartphone pictures? It needs to build a classifier, doesn't it? (To be clear, the system only reports those images to an officer for further analysis and actions.)
I mean let's consider that the images are somehow pre-processed, then some convolution(s) for feature extraction, then maybe some more intermediate steps and at the end apply the results on a classifier. How would that classifier work? Would a probability between 0 and 1 suffice? And if so, report those from 0,5 and above? Or would there be better techniques?9 -
I JUST WANT TO FUCKING EXCLUDE A DIRECTORY....
I run the code cleaner tool, OH CHRIST it's trying to sanitise the automatically generated code, I don't want this.
I try to exclude... takes ages to work out that while specifying the dirs is absolute you can only exclude relative but from what? I want to block a/b but not a/c/b but no it's all you can only block all b b it a/b, b/b, c/b, c/b, a/c/b, etc.
I google for other solutions, nothing but trash, docs a trash, here's some examples but we don't tell you the actual behaviour. All I want is to get everything in /home/hilldog/emails but not /home/hilldog/emails/topsekret how hard can it be?
I use the source but what's this, BeefJerkyIteratorIteratorBananaSpliterator all over the shop how much convolution and LOC does it take to provide a basic find facility?
Screw this...
$finder->in(explode("\n",trim(exec('find '.escape_args(...$good).' -type d ' . implode('
-o ', prefix('\! -wholename', escape_args(..$bad))) . ' -etc | grep -vETC \'pretty_patterns\''))))
3
Top Tags
Weekly Rant
View