Zoom Logo

CEEDS Seminar: Machine Learning for the Natural Environment - Shared screen with speaker view
richard
16:12
There is 2 of us here if that help you get to 150!
Zhang, Ce
16:36
looks really good!
Rebecca Sudworth
20:46
Hi can we get a copy of the slides? Or do I need to take notes!
He, Vera
21:28
want to ask that as well
Sally Keith
21:46
the whole session is being recorded :)
Zhang, Ce
21:55
You will have recording for sure. If speakers are happy, I can share the slides
Gordon Blair
22:05
We can also collate slides and make hem available.
Tom August
22:11
I'm happy to share my slides
He, Vera
22:26
Thanks :)
Adam
26:47
Nice explanation Chris!
Dr Khalid Mahmood
26:56
Hi every one.
Dr Khalid Mahmood
27:04
Zoom asking passcode as few colleagues wanted to joi
Harrison, Samuel
27:53
If they use this link they shouldn't need the password: https://ukri.zoom.us/j/94105570319?pwd=dVNZRWNSczR6eC93SENYclcwejFFdz09
Dr Khalid Mahmood
30:13
Thank you Harrison
Zhang, Ce
30:17
Password: 838989
Rebecca Sudworth
31:53
That was fantastic thanks
Rebecca Sudworth
32:15
I can now amaze my friends and colleagues with my expert understanding ...
Gordon Blair
32:46
Chris - as a researcher in this area - what excites you in terms of pushing the boundaries of machine learning?
Alex Bush
33:16
???Could unsupervised classification indicate what number of classes/labels could be reasonable be distinguished with supervised learning?Could unsupervised guide more efficient labelling?
Garry Hayman
35:42
??? Application to process understanding rather than image processing
Stelian Curceac
36:11
I assume AI refers to reinforcement learning, how about deep learning?
Rebecca Sudworth
36:16
Are there ethical or other issues related to unsupervised learning due to characteristics of the input data?
Larry
36:59
How do you try to avoid over-training or narrowing your machine learning models too much?
Isaac Allenby
37:14
could you teach the machine using supervised learning to label complex data such as supervised learning?
Harrison, Samuel
38:12
Not a question but a comment: Even if you're a mechanistic modeller who couldn't ever imagine writing models without mechanistic processes at their heart, then ML methods can still help you with your model data analysis, e.g. looking for trends in model sensitivity analysis data. Great talk Chris, thanks.
Pocock, Michael J.O.
57:34
AI Naturalists Might Hold the Key to Unlocking Biodiversity Data in Social Media Imagery. Patterns 1(7):100116DOI: 10.1016/j.patter.2020.100116
Sally Keith
57:55
clap clap clap
Rebecca Sudworth
58:27
Apologies I have to go to another meeting now - great talks thanks.
Sally Keith
58:46
?
Zhang, Ce
59:00
that's alright no problem
Thackeray, Stephen J.
01:00:16
question here - if that's ok?
Levy, Peter E.
01:02:29
?Could you use existing BRC data on UK species distributions? This tells you if the species has been identified in the local 5-km square previously, so useful prior information.
Tom August
01:03:13
Yes Pete, we are looking to do this too, like a prior as you say
Fry, Matthew
01:03:20
For Tom: That sounds like a massive research project, but you said you did it for a hackathon! Does this drive you towards machine-accessible datasets, and what lessons does that bring out for how we need to make datasets (esp training data)and ML models available?
Dr Khalid Mahmood
01:05:30
Rothamsted Research North Wyke Farm Platform has millions of data sets and be open for access to scientists and companies, we will be hosting insight. Please mark your dairies 3rd November
Dr Khalid Mahmood
01:05:32
https://www.eventbrite.co.uk/e/digital-agriculture-environment-solutions-sme-collaborations-tickets-124977931357
Tomlinson, Sam J.
01:05:49
For Tom: do the species identification algorithms risk dismissing real-life outliers, with regards to a species creeping north in latitude, for example, especially if you link it to habitat and climate?
Tom August
01:07:03
Matt: Yes, I believe the biggest challenge we face is access to labelled training data. At the BRC we are making moves to make our images more accessible for others use to train models. It is also worth saying that most models for image classification of biodiversity are NOT open, this is a real shame as there would be a big benefits to conservation and research if they were.
Tom August
01:13:00
Sam: Yes, as an extreme example an albino peacock will not be identified correctly if your training data only has normally coloured peacocks. In my experience, in these cases the classifier normally has low confidence, but I could imagine cases where that wasn't the case. Regards range shifts, I agree, you would not want to have priors so strict that they lead to confirmation bias.
Tomlinson, Sam J.
01:13:53
Cheers Tom
Jamie Alison
01:17:00
Tom: Great talk! A philosophical question: My gut tells me that in order to be really really sure about our what's happening in the wider environment, there is no alternative to systematic sampling and "rigorous" data collection (i.e. following a specific protocol, and very carefully documenting deviations from that protocol).However, what can be done with opportunistic datasets is clearly amazing - is my gut feeling somehow out of date?
Fry, Matthew
01:21:59
To Ce: we're interested in merging datasets, e.g. hydrological forecasts from a range of models, which is often mentioned as a problem suitable for ML / AI. Could you explain a bit more about this and what AI is bringing to this? Would it be training a forecast-merging algorithm against the real world outcomes? How do you merge datasets but keep consistency, e.g. spatially, as opposed to producing an output that is as good as possible for any individual location but not necessarily spatially consistent between locations? Maybe one for a future workshop if we don't have time now!
Tom August
01:23:22
Jamie: I don't think this text box is big enough...! Briefly, yes I think you gut feeling is out of date, citizen collected data (without a protocol), has been used effectively to monitor environmental change and underpins numerous national and international assessments, such as the state of nature report. I note there are a number of participants to this meeting who have authored papers and reports on this topic. I agree that the data I introduced today is messier still, and we certainly need to consider carefully if it can or should be used in these sorts of analyses.
Jamie Alison
01:24:23
Thanks :-)
Zhang, Ce
01:25:23
Hi Matt: This is quite difficult question to answer. Are you looking for model ensemble? Maybe we should discuss more in details through a workshop
Vasileios Antoniou
01:25:38
Jamie: No, your feelings are not out of date. All data can be categorised with different "confidence". For example dual channel gps system are mm accurate, simple gps is 3m. It is possible to get to near mm with simple gps if you have many inaccurate readings.
Fry, Matthew
01:26:19
Ce: I think so. Thanks.
Harrison, Samuel
01:27:39
Great talk Clare. How was land cover image classification done before you moved to random forest methods?
Harrison, Samuel
01:30:15
Thanks!
Morton, Daniel
01:30:21
Also it enables us to use categorical data in the classifier, which maximum likelhood cannot
David Fletcher
01:30:27
?
David Fletcher
01:30:37
What is segmentation, please?
Martin Wain
01:30:45
As a practical project would an integrated monitoring scheme using satellites imagery be able to identify local areas of flowering primrose in Spring, around Morecambe Bay to help us to find sites that may hold Duke of Burgundy Butterfly populations, and to help us to work with farmers to connect up good habitat? In spring primrose flowers clearly stand out very clearly in the landscape!
Rich Burkmar
01:34:13
Thanks to organisers and speakers for an excellent session. I've got to go to another meeting now.
Jamie Alison
01:34:42
Thanks to all the speakers.
Thackeray, Stephen J.
01:35:28
sounds great Pete
Fry, Matthew
01:35:43
Thanks everyone, very good set of talks.
Philip Donkersley
01:35:55
It does seem that a lot of the training data sets for designing ML algorithms are based on satelite imagery/photography. Does anyone have examples of alternative data used to train their systems? :)
Tso, Chak Hau
01:36:07
Thanks very much guys!
Tom August
01:36:50
Philip: there are many examples using acoustic data
Cooper, Jonathan M.
01:37:00
Really good talks, thanks speakers
Sally Keith
01:37:07
agreed, thanks all
Harrison, Samuel
01:37:12
Thanks all, a great session!
Garry Hayman
01:37:18
Thanks to all
Dr Khalid Mahmood
01:37:19
Great talks
Zhang, Ce
01:37:27
thanks to all
David Fletcher
01:37:34
Thanks everyone! :)
Guillaume
01:37:35
Thanks to all speakers and organisers!
Sally Keith
01:37:48
YES, JOOOIIIIN USSSSSSSS
Dr Khalid Mahmood
01:37:51
plz share a link
Philip Donkersley
01:37:58
Cheers Tom! Thanks all!
Rathod, Biren
01:38:04
Thanks all.
Mike O'Malley
01:38:18
@phillip Donkersley having a look on "Kaggle" can be a good idea. Hundreds of competitions all roughly related to Machine Learning on loads of different applications, some environmental some business focused..
Sally Keith
01:38:31
https://ceeds.ac.uk/join-ceeds
Philip Donkersley
01:39:04
Thanks Mike! WIll do! :)