Incremental Learning Demo, page-57

First 57 Last

Share
Profile

Follow
uiux

6,267 Posts.

9114

03/10/21

15:17

Post #: 56530406
Share

Originally posted by hopalongpetrovski: ↑

Hi @uiux .
Thank you so much for putting this together and enabling us to have a play.
Am very impressed but am wondering why it seems to struggle somewhat with text?
It very easily distinguishes between say live images of Grass, yellow flower, white flower and buttercup (also yellow) and can easily discern my foot from a croc from a milk bottle from some bamboo.
But when I show it some words, black text on a white background such as couldnt, phenomenal, of, ability, etc it's recognition rate is much degraded. I would have guessed that the spike patterns generated by any particular word would be as unique and recognisable as any other and much more complicated data package of say the image of a buttercup?
Is that maybe the problem.... that not enough data difference from one word to the next?
Sorry in advance if these are dumb questions or a dumb experiment not within the parameters of your demo.
Appreciate any thoughts.
Hop

Expand

yeah, I'm not an expert so I can't give you the best answer

but this is mobilenet, and the weights have been created from training with the imagenet dataset

https://www.image-net.org/

so the network has been optimised to recognise shapes like this:

to recognise words and letters, you would start with a different network

so its kind of using the wrong tool for the job

all imo, still learning

BRN Price at posting: 39.5¢ Sentiment: Buy Disclosure: Held
22 Upvote

2 Great analysis

Reply

Top Stories

EBR

EBR Systems' heart pacing device on-target to reach the US market

WGX

Mining start at South Junction adds to Westgold's hopes for WA gold complex

GDP grows 0.2% in June quarter, but annual growth the slowest since the 1990s

Is the ASX getting its groove back? Here's 5 IPOs in September to watch

SKS

SKS secures $20M for extension works on Melbourne data centre