Epic
785 karma · joined May 14, 2017
YC Badge: 0x0c51bf0090235f2604af81df021023a3d62aa4d3
E: alex [@@@] numberboost.com
L: linkedin.com/in/alxcnwy/
X: @alxcnwy
Consulting: www.numberboost.com
Here is a vid of one my talks on deep learning for computer vision: https://www.youtube.com/watch?v=X4Q6C915sUY
-----BEGIN PGP PUBLIC KEY BLOCK-----
xsDNBGUfNa8BDAC1U6epAyxNbnBiVVa/5xg/VqviJmBvAFw7IMuRTXXpmmnUdPehwFsZTIsVxOQs jSZgB58d7V7Ps0NyDbpThGdV6D/Rs9O0RQ7NzBhPWtgXMjL05CMzQQKwAQwYiOXoMmdIFx+dDs4Q sEuTwhN7ZVfq1wCwvKaGaIa9YNs+84NP9nC4+0i1mb3kY4YFwsRUuiX1UuKaSVkF47ZGAdm+QyCz 4r12/eiRlfyR4Tn2MoKTcWjWcpON0Jt2c6NZAmae1ofBQd11AdenItwMcIz/YB388I+WSUtXal6i Ou1kHfkZD0cVBnKWSxbbdkTvlJwOTROq1PqfUxoJ7lNDse9t4UqWSLnxcChAw7G6cusvemEfyUBs bpbehD756bwrQDD4bc1GX4wHG7iERDrraUHEado+D4hcZJeE3hQ+hLQLbVUkfIOo8QANu/nEc533 utazWsUdzPB9ZhNXGESBU6FD1Dq3T7F88GQwgLhMWpLyw2bVIrHE1K+VFlWKeeTuS6LUXDcAEQEA Ac0AwsD6BBMBCAAkBQJlHzWvAhsDBBUICQoHCwkIBwMCAQQWAwIBAh4BAheAAhkBAAoJEL0wgBjJ Evk4fY8MAKEwfG6n8MSCdH/PYJdZCorgFe3Si+5XUAKuRVzXpIihwz+imuaNqabSvKt3BwvAJ3WQ F7xbWCSveIeTKv893gv5kglFi0pxQhCd2kpWY96O/zKThXZGbBQG19/276M0V9PnubnO0EOpNzxi QYjnU4UxmH32mjwx42Xb+/44Ird1iBW6TU4oE87iugjqbMREvsuRerur/jViUhCCkAwohM4OWTP6 HvHj57KH01wcqWROyiTj2amdV+lpizzRyvF9cC68p5zakGG7tyJ7fQMXIhgV72hj37mMDSq/qcsB lB3PFjc8mu/9zKwWlVNhV3G+8fFZ+dTuNvdsP0v2XbLqrONMYFrbnH3lsG/SjTzP0Ea5Mt3Wq8JV UQvA70xlS7YaI8juAbTYLFmssNVDQ77G+gMVXiKxRtPNvH4sc6HOsemsDpA3Rl86iY5MGDexT/oH 3FxslNSHGCy77WPRe8HsipoWLiWRiHuQqc5afsMExtgnoUIpU6r3WPG5Be00Pg0dxs7AzQRlHzWv AQwAxdgbC3umULrQGh423QVZZh+vwv5k9K2c/67JzuErFo02ld+9+0B23xbulUSw9igz1udmBFtW sU8212O40TMDACOdxMljYhB99zVnoo1zW17wce+ZrR/F+Qgdc8cJ0mb+V66/0VFzdxgVMEwCKVS+ TweUhrDBwnxVAr+zTPvTBHohIq9S2Md4+r5fM5uhy67WQFJ+nYe8mVEVHhG7NETOCSqVx1WCF/vs +4d5aUTZ/A8OVZxiEe2PlxeWeH4FVYsQe7qAroM0nkawtSPMGcHPfsM3R49x2XEiKeSssyXNiypX mIc8Z2MiIt8Yu264C8BXllL9LVKLgoSdgs/zoDK0VoLthLh+PNrz/FUsU+lvQaPUVL9MywP4FWyt K/Y82ObuUDglLs74HsT1y7jPV0G2HqmHe9jkdvFsFe6pqIILoOEucMyvYsrTfwFlAHuYQxjjCcIm L95F+AZQLMoD0gyTKuIlOvfX65bWSNp+5msPmWHP98eFvwSWvwhYf8CfxpFtABEBAAHCwn4EGAEI AagFAmUfNbACGwzA3SAEGQEIAAYFAmUfNbAACgkQDnm6M0ernGu5qgwAw0Xss6TGEWIkuBKgTkpl aN3iKETevBhJ+wp7uPeUnEZZdUwIEnZI4Vy7yiRUmhxH5BF78xueE+qOdiHVGzeLeEUj/mg6y1Mf LL5peoyjZLWOULOUec5vLKpLtUKuWrIoaZmQjFoJerWDpyLb0j/OpdSf077hrvNtK+a7pCGl++GT IFqiPHIibrsC02zLy0Oc4PjJgHI02vGhJLoEnml3dbuG/QoE1RCsnKY52bjA9qrKklS4ymwURpN6 jODnCTjpJZv+6wEDbf+9ks8S+XUfQyl3B+s+6Wp9SdSjIMsewQWiZUAnro6s72PdoBDPu8KX4N5/ itJUkqfKRtJDgU8geKgbnqy5Gw1CnDfs2ER1GlNoFFdsgGeSzNTmu4hzrtS9JcemkmBsYnxNLMjD //RWR4xiS6mEeZtEHj5/+8p4UqIWHIn0mXDWxwpZx0vsgtceoGAtS+KPuZIGHJSBkhnBkIg48mq5 oHGABF3W8s78LQT1F6Wf+r1TNARxXhLKRAfVAAoJEL0wgBjJEvk4xhcL/iZTPV+/mu5N3VA5zqYc uzXt4Z599ki57YJ0/ciZIlbjXog2Sqso78pxEw3PfBffJznamyqbchUiarjR5CMCZbTPJ7vJVPoF pPAQCRLfAVop6Jdq3fyAQMURFkgpMwfmSzZ3Fri/4TP4DRCRAEjzvCchfbTGI1H8ifz3bAoAFlG/ b2ttZug0baKb2gzFfsG+XPr3JfXG/at+95M1wbO3Kiw2y8y5mct4xwRUnNnkuEdoswcgeIe+cvNm JS3mxUL8PdmOLY+Wk1+tGjbgBiiH+6HE/uSteXYPDFklq5AtJ4/62nddHiZ4uisK+Pxy8TnToCaV 7Wbg0V2l6nJPPBmWQqgMQtPMrD0Blr5FkVND/+4cBH8dB2a7yi/MmJqaydxlc7hrKR5dvnrIRNJg WG5gq0Vn5z4FLr1Crp4uu5ymUWF/mdxSa4hTL1kboRfrL0GTEKB8p42GgJJdr5evHlkFv/GEY/MS sKod3avN+QOFXfpvRxbXssHClpuFunAPU+QQisXEmARlHzWvAQwAtVOnqQMsTW5wYlVWv+cYP1ar 4iZgbwBcOyDLkU116Zpp1HT3ocBbGUyLFcTkLI0mYAefHe1ez7NDcg26U4RnVeg/0bPTtEUOzcwY T1rYFzIy9OQjM0ECsAEMGIjl6DJnSBcfnQ7OELBLk8ITe2VX6tcAsLymhmiGvWDbPvODT/ZwuPtI tZm95GOGBcLEVLol9VLimklZBeO2RgHZvkMgs+K9dv3okZX8keE59jKCk3Fo1nKTjdCbdnOjWQJm ntaHwUHddQHXpyLcDHCM/2Ad/PCPlklLV2peojrtZB35GQ9HFQZylksW23ZE75ScDk0TqtT6n1Ma Ce5TQ7HvbeFKlki58XAoQMOxunLrL3phH8lAbG6W3oQ++em8K0Aw+G3NRl+MBxu4hEQ662lBxGna Pg+IXGSXhN4UPoS0C21VJHyDqPEADbv5xHOd97rWs1rFHczwfWYTVxhEgVOhQ9Q6t0+xfPBkMIC4 TFqS8sNm1SKxxNSvlRZVinnk7kui1Fw3ABEBAAEAC/90T2K+QPDMAklIaiEpG10sU7YRbDhzOzHu I40weRVncMrtotGtCavlwHl+zzrlcOzqsDjag4ZLiFt8711TcRyvuDgnlUf43Eswqvf0/vqZdnPF Lgl+OPc2WpWdr0J5/MrPLTX0EqkPf2YbM2Q3xLvuQVeKBphyot264EQCT0Utp5CKt3TeN+3ccBJp 45L4Hid3A2GjQeU+p+Ala852Aq0Zdhm+GAR0TlciY+xABgpH9XkgBYKsduEhy+89ixK0VDrPe2yy eiUAkQjGoDISoi4NpgnkA5g6YntBLTtde4Fk6AvQeEjHWjqawKclDrtXeuFyKHmj1Nn3uqOHTImJ AJwFbab23dl2mEYeakcL8ihnpvuggcakAO1rgXRWVmibItjdPWw1uMo+ugXAwm8irW/vAQe5uLMH eLAD2lKci/gGp8/e2rcv1eTKjtfq+L4VH8BoNWmN6e38EBox1Fai5hRMtRtWOdlBSuCP5u2K+sg6 xfaCvm2opBDDdkQYQ3c/dIEGAOItRO0ezZV7Fy1zCXzKz5JGLHblm+2SdTSJKJBaJzt4OIWhOpKG jmWIAeOMk0hV55PNarieoX6aNwm9wnEtoB7hbp+fmq994hjzkm7/irtwWLH33d+XUNlmGjj3Pvua Ri0sSD5Uoa7gxfizekGN7xH3GeWUwEuOSHAi1E0RICfMH3KV5z5ezVpFbq+GtDxuFrOcXmEUh8k1 BL0XCyhIwUNlmBPGs5TdswB5ZfPpLcDTQoqTAQ4pKVfoOMoqYkENVwYAzTxyJ2iEZSUyc6J30BFA oWyBlUcJo5T2q23aWA2JZTMVYKNKNomoZsJ3F3G8wpDriFLFgEjwjp0CQpnPrt/MSd3SgiZkTP8L SBkK4yKemVsjb2fInHgrU334WUg1cLFaJ9zRf1qVJk6jo60rSmp81fsWNQpMK4OcU85j34uw8h5F dIQnUxcjRwVCQj5umWGYQrebF0qj1xLSCtA5D6pdzPZKZo+4aHU6ELPlW0Sm4T2braOXBsBkdRMk i0SpMPwhBgCY/5Hoj3z12zIOSqRMzrfzKbNJTqyz+8IHtWOm/6/RTlA1m9nR4l8dKWi/A0YRbZ1Q fhzPI8L4To3Ta2Rb1vPPPZF67biMZs/YVC3Vd6ZdaT3aXjswOtKaqNxYJsur37oCUSIPBmsGr7ys Itif0oO7dsJ3bwq4YErDSaiWjl9RmGsZpr0hPBHQ4snLyHRgmwPnr5IivIA0o83FFCM6xv1R3TOS PSG4XRH8D2MLrj5GU6oZDI4uXFNz1bUYesjOxK3RNs0AwsD3BBMBCAAhBQJlHzWvAhsDBBUICQoH CwkIBwMCAQQWAwIBAh4BAheAAAoJEL0wgBjJEvk4eo4L/2v0fEQxSdYwSHnuH9htH/EJZNadRnJc 6UjUv1Tg2XAxm3W5WWF9DCTufjnpEYGA0aT3fsYiAcBSqDUIe/r4d83K6BnmbzDHqnD+L03HRPIq xsmiP7u2mxoxnjxEjuzILQ9TemRNkcQwWqkRK7RNsJLcOdvMH1ydWRmYaDpKYsqd127b8qR6w7tx ro2JduFrV/LYEZt2anKUGZ0Q5aWPYHsmlM+8BuIoCjPewPOHgFLk9RFRLLuR9AlvMCU5tNToVJwr NM9m6yf4yKUAW2oVfPv0E79z9Lsu32B/dJAIweTeYxymRUDK+4/wya7BFdesUZ9TI5YMs4pSOaMx UZbcEIYs0Ji6BPR3q6Dnaon1rQxrulYRFEfUfb6XIOjCDSyLjI84chsU7asbeBpVy5ER9PxYhdaM gi06yK6KBzX/llSIc5sK9eQpyB0rAI3ffZfB5iOcgX+X1eREm0y3m46dI7nSXtgIUiIlGtLKf6yZ ALlQLXSqS+19rHL7aWuZ2I8BfcfEmARlHzWvAQwAxdgbC3umULrQGh423QVZZh+vwv5k9K2c/67J zuErFo02ld+9+0B23xbulUSw9igz1udmBFtWsU8212O40TMDACOdxMljYhB99zVnoo1zW17wce+Z rR/F+Qgdc8cJ0mb+V66/0VFzdxgVMEwCKVS+TweUhrDBwnxVAr+zTPvTBHohIq9S2Md4+r5fM5uh y67WQFJ+nYe8mVEVHhG7NETOCSqVx1WCF/vs+4d5aUTZ/A8OVZxiEe2PlxeWeH4FVYsQe7qAroM0 nkawtSPMGcHPfsM3R49x2XEiKeSssyXNiypXmIc8Z2MiIt8Yu264C8BXllL9LVKLgoSdgs/zoDK0 VoLthLh+PNrz/FUsU+lvQaPUVL9MywP4FWytK/Y82ObuUDglLs74HsT1y7jPV0G2HqmHe9jkdvFs Fe6pqIILoOEucMyvYsrTfwFlAHuYQxjjCcImL95F+AZQLMoD0gyTKuIlOvfX65bWSNp+5msPmWHP 98eFvwSWvwhYf8CfxpFtABEBAAEADACguxgFtTgCIDNO9LsHxv/BOv71AO4N5jpFcsAbZc02rhTp n/NdexrX1W6RMFRAN5sd/bJg2b/Nnxhj1m1g7/uBEm5rYF1wXwunWejPI/HTCod3WPvVrZmN3yzm UY7GM8xwedT2fiC7YYJdUXL0FxfhZrwgzNstYQtDh3Bom7Hs0je3UQ6SdUPxGdaIO+kva+Q+L4cT /iGXwUrRugAHz9o/kixK3EorqAz4sDBP+JLL94NixADcpc5hJtWlIGS/AZE91ufRKPjybrppgiGe HR7GliXYc4fgcXlPs/9uWgfDgf2biRDTQ73DFIAWqRh42WEvnUs0avIKgKdkoIOmUb0mZT5jGxXe 5BQDjrHm9RvAyuZ2nnZJVWDIgRKOR1azx0z7PGNL0qyEcJ0VydzLKjrzqFFJOdISRgPXUl76sQiF rM1sL4fWbgdogINwSAblnKg2sJuIapwN9IAFglNqrZXBfIIlYYYW1K/EUG/zpLxR6RzJnUXuo8d3 C0JEEsiSIRUGAOcNGH3OEdsogHsTcV4dp1bzcm4XA9TY4x2P3UNwS/XiEARj0TmlHi0J6Q7AjHk5 65KivwjSwWTfe7XOg6oVUoobbTTn9jXmveTac8rZfbxasvHfQqGiEPFEaxyWAjjhJEBIp7twHu2e bS2IcqKxsN/48d3K1M8eBriybqF7gkPghnxfcmRnYIcuZ6/7uSDO/FplHyl8Z19DTCVHbS5fEyck TQlllEPfnmEQqo1NZSzLfPWtzsS15Dw+A2kSLK6hSwYA2zUTQkcMSrHY4/snm1dlNy3FWL6361OD 2hM4xYDYymig1GfTw4rD7SD2Ni2OeKms0Y+tlRJPiOqMNpo+XvlFRqFW5Iro2BRg1Ag6I/2AM+oQ BEe0J5EKl52sMX63sGxVcCn20Rr8EJ40sdSXUj7bw/wSlBUTJpAurzLc5K+Ks4wu4YfrJRXGOSwK murvBs62ZmQb2BWBGr4StWPhEcURRyVLIgJL7GavFH+Td3TvB0mCw5Ws+iyWfC40pX4hI50nBf9E ZQHXYiL6eYpuCE3hor/955qw8HyvrNDK9YHI9sv77uEz4t+DvHbrIN/oPQ7wuE05cYBBdSrVe1LK Tnue9vY9FkDLSI9x2Igdii/DP7S+WST40xTr8xrHR2O8n+IdvXLgu99PsvP6OdKTxxVSnt4pvzh+ EX2r+CqxF8cwUYdrUJLFoEowBdhkTM9h41duWz7YvtPJ02f59GI5Ev6dT8UJRCWuj53TPwEh5hXO dF8rp+EJACx+yVUDksnovhT6VXniF8LCfgQYAQgBqAUCZR81sAIbDMDdIAQZAQgABgUCZR81sAAK CRAOebozR6uca7mqDADDReyzpMYRYiS4EqBOSmVo3eIoRN68GEn7Cnu495ScRll1TAgSdkjhXLvK JFSaHEfkEXvzG54T6o52IdUbN4t4RSP+aDrLUx8svml6jKNktY5Qs5R5zm8sqku1Qq5asihpmZCM Wgl6tYOnItvSP86l1J/TvuGu820r5rukIaX74ZMgWqI8ciJuuwLTbMvLQ5zg+MmAcjTa8aEkugSe aXd1u4b9CgTVEKycpjnZuMD2qsqSVLjKbBRGk3qM4OcJOOklm/7rAQNt/72SzxL5dR9DKXcH6z7p an1J1KMgyx7BBaJlQCeujqzvY92gEM+7wpfg3n+K0lSSp8pG0kOBTyB4qBuerLkbDUKcN+zYRHUa U2gUV2yAZ5LM1Oa7iHOu1L0lx6aSYGxifE0syMP/9FZHjGJLqYR5m0QePn/7ynhSohYcifSZcNbH ClnHS+yC1x6gYC1L4o+5kgYclIGSGcGQiDjyarmgcYAEXdbyzvwtBPUXpZ/6vVM0BHFeEspEB9UA CgkQvTCAGMkS+TjGFwv+JlM9X7+a7k3dUDnOphy7Ne3hnn32SLntgnT9yJkiVuNeiDZKqyjvynET Dc98F98nOdqbKptyFSJquNHkIwJltM8nu8lU+gWk8BAJEt8BWinol2rd/IBAxREWSCkzB+ZLNncW uL/hM/gNEJEASPO8JyF9tMYjUfyJ/PdsCgAWUb9va21m6DRtopvaDMV+wb5c+vcl9cb9q373kzXB s7cqLDbLzLmZy3jHBFSc2eS4R2izByB4h75y82YlLebFQvw92Y4tj5aTX60aNuAGKIf7ocT+5K15 dg8MWSWrkC0nj/rad10eJni6Kwr4/HLxOdOgJpXtZuDRXaXqck88GZZCqAxC08ysPQGWvkWRU0P/ 7hwEfx0HZrvKL8yYmprJ3GVzuGspHl2+eshE0mBYbmCrRWfnPgUuvUKuni67nKZRYX+Z3FJriFMv WRuhF+svQZMQoHynjYaAkl2vl68eWQW/8YRj8xKwqh3dq835A4Vd+m9HFteywcKWm4W6cA9T5BCK =Iwr2 -----END PGP PUBLIC KEY BLOCK-----
Major question:
Why isn’t Tesseract using neural networks. I know it just introduced LSTM based models but they suck.
Why is GCP vision text recognition API so much better than open source alternatives?!
Why isn’t there an open source OCR engine even half as powerful as the Google Cloud Platform API?!
It reminds me of "trap streets" mapmakers add to their maps to "trap" potential copyright violators: https://en.wikipedia.org/wiki/Trap_street
Their “new method for finding the latent code that reproduces a given image” is really interesting and I’m curious to see if it plays a role in the new $1 million Kaggle DeepFakes Detection competition.
It feels like we’re almost out of the uncanny valley. It’s interesting to place this in context and think about where this technology will be a few years from now - see this Tweet by Ian Goodfellow on 4.5 years of GAN progress for face generation: https://twitter.com/goodfellow_ian/status/108497359623614464...
His points on grounded language learning and meta learning are really interesting...
Interesting paper on an unsupervised pre-training approach using overlapping image fields to massively reduce the amount of labeled data required to train image classification models.
Reminds me a lot of the way word2vec is trained with overlapping word vectors and negative sampling...
I'm not suggesting you can use models in all situations or an all current setups, just that it's possible (and surprisingly easy) to use models to automate a lot of analysis on video.
A lot of current commodity $100 cameras are good enough resolution for many use cases but of course you do need sufficiently high res video and good enough lighting conditions.
You also don't necessarily need to do facial recognition to track people. I'm nervous of the privacy implications of all this surveillance tech but it's possible to do useful tracking (e.g. I've worked on a project counting people to prevent over-filling busses which is a big issue in Africa) without needing to use facial recognition. In the counting case, you don't need to recognize people to count them and can do edge computing on quite low resolution video so that the raw footage never leaves the device (only computed count statistics).
* identifying vehicles & reading license plates * identifying people and tracking them across cameras * identifying products on the back of trucks * counting people in queues or getting in/out of vehicles / stores * detecting suspicious activity / shoplifting
Generally you don't even need to completely eliminate humans watching footage, you can just present the human with cases detected by the object and let the human decide whether it's a true positive / what to do.
It's scary but useful technology because there's no way humans can watch all the footage being produced by a billion cameras...
[1] https://speakerdeck.com/alxcnwy/gdg-devfest-2019-deep-neural...
I’ve had an employee tell me they’re exhausted and need time out and I definitely didn’t “propagate need to my network” about them. That would be a serious breach of trust. I’ve also hired an engineer who told me they just came off a burnout and they were great.
This fear of vulnerability is part of the problem - it’s ok to be burned out and need a bit of a rest.
The insidious thing about burnout is you can’t think clearly when you’re burned out. I’ve been heavily burned out twice and I can relate to just wanting to work in a bar or something but after a couple weeks off you’ll be able to think clearly again. Worst case it’s 2 weeks of paid leave before you quit, best case you come back grateful for the opportunity and recharged.
Your team lead sounds like a pain in the ass, the founders would probably get rid of them instead of you if you gave them an ultimatum but my advice would be to take time out before you do anything rash.
Good luck!
There looks like a lot of overfitting the validation set going on in that post.
It's also a mistake to conclude that "there was no subtle underlying pattern" just because the author couldn't find one.
Throwing XGBoost at a bunch of technical indicators isn't gonna cut it but I have had some solid real-world success (as have several people I know) applying ensembles of deep learning models (with regime switching based on model residuals) to profit from "subtle underlying patterns".
Funnily enough I think the ML hobbyist problem is most pervasive in the "predict the stock market" domain. There was a post on HN a few days ago [1] that was overfitting the validation set and hand-waving away fees and spreads. The author concluded that "there was no subtle underlying pattern" because they failed to find one.
If you run 10'000 experiments over thousands of indicators, you're bound to find some that idiosyncratically perform well on the validation set but underperform out-of-sample.
You also can't conclude that "there was no subtle underlying pattern" just because you couldn't find one...
I've done a lot of work in the space and would love to chat - just emailed you :)
"we realized that there will always been a better resourced or more dedicated team willing to fight you for your alpha" - that's part of what makes financial markets such a fun and interesting challenge but agreed, intermediary risk at the timescales you were operating at in this kind of unregulated market is real and not fun
I'd love to hear what problem you moved to that you believe you can build a long term competitive advantage on if you can talk about it?
What deep learning gives you that’s really useful and valuable (beyond better OCR) is that you can use graph convolutional networks to automatically parse the OCR output and convert it into structured data. You could hand-write a parser or use a template matching approach but you’ll have to create a new parser/template for every ID card type whereas the GCN approach can be used to learn the parser…
It's also very hand-wavey on the details of how to actually use graph convolutional networks to extract structured ID card data. For example what "bounding box information" is used in your node representations? What is the architecture of your biLSTM?
This seems very much more like a promotion for your API than useful information on how to build a system that extracts data from ID cards.
You don’t carry “all” the risk but your risk still goes up.
Sure, the capital injection helps grow quicker but many businesses could have been comfortably bringing in a few million dollars per year but end up dying because they were forced to grow too quickly.