871–880 of 1001 posts
Want to build a platform to alleviate chronic suffering that can't be understood by one's local doctor. Suggestions by this platform wouldn't interfere with treatment protocol straight away; it wouldn't ask the patient to stop medicines their doctor has prescribed, or itself prescribe scheduled drugs. It will suggest complementary interventions. Case in point: anxiety, depression, brain degeneration & other related d…
Finishing up my PhD thesis on low-resource audio classification for ecoacoustics. Our partners deployed 98 recorders in remote Arctic/sub-Arctic regions, collecting a massive (~19.5 years) dataset to monitor wildlife and human noise. Labeled data is the bottleneck, so my work focuses on getting good results with less data. Key parts: - Created EDANSA [1], the first public dataset of its kind from these areas, using a…
oh man that's awesome. I have been working for quite some time on big taxonomy/classification models for field research, espec for my old research area (pollination stuff). the #1 capability that I want to build is audio input modality, it would just be so useful in the field-- not only for low-resource (audio-only) field sensors, but also just as a supplemental modality for measuring activity out of the FoV of an im…
On the labeled audio data front: our Arctic dataset (EDANSA, linked in my original post) is open source. We've actually updated it with more samples since the initial release, and getting the new version out is on my to-do list.
Polli.ai looks fantastic! It's genuinely exciting to see more people tackling the ecological monitoring challenge with hardware/software solutions. While I know the startup path in this space can be tough financially, the work is incredibly important for understanding and protecting biodiversity. Keep up the great work!
I am working on nothing which by the is something important and precious because it is something to works on nothing. By the way, nothing else matter, which mean that everything keep my mind on working on the thing of nothing.
Finishing up my PhD thesis on low-resource audio classification for ecoacoustics. Our partners deployed 98 recorders in remote Arctic/sub-Arctic regions, collecting a massive (~19.5 years) dataset to monitor wildlife and human noise. Labeled data is the bottleneck, so my work focuses on getting good results with less data. Key parts: - Created EDANSA [1], the first public dataset of its kind from these areas, using a…
You should train a GPT on the raw data, and then figure out how to reuse the DNN for various other tasks you're interested in (e.g. one-shot learning, fine-tuning, etc). This data setting is exactly the situation that people faced in the NLP world before GPT. I would guess that some people from the frontier labs would be willing to help you, I doubt even your large dataset would cost very much for their massive GPU f…
The main hurdle we've hit is honestly the scale of relevant data needed to train such large models from scratch effectively. While our ~19.5 years dataset duration is massive for ecoacoustics, a significant portion of it is silence or ambient noise. This means the actual volume of distinct events or complex acoustic scenes is much lower compared to the densely packed information in the corpora typically used to train foundational speech or general audio models, making our effective dataset size smaller in that context.
We also tried leveraging existing pre-trained SSL models (like Wav2Vec 2.0, HuBERT for speech), but the domain gap is substantial. As you can imagine, raw ecoacoustic field recordings are characterized by significant non-stationary noise, overlapping sounds, sparse events we care about mixed with lots of quiet/noise, huge diversity, and variations from mics/weather.
This messes with the SSL pre-training tasks themselves. Predicting masked audio doesn't work as well when the surrounding context is just noise, and the data augmentations used in contrastive learning can sometimes accidentally remove the unique signatures of the animal calls we're trying to learn.
It's definitely an ongoing challenge in the field! People are trying different things, like initializing audio transformers with weights pre-trained on image models (ViT adapted for spectrograms) to give them a head start. Finding the best way forward for large models in these specialized, data-constrained domains is still key. Thanks again for the suggestion, it really hits on a core challenge!
I’m rebuilding a learning library. Webpage bookmarks (with big help from raindrop.io), PDFs (hundreds of them, mostly book chapters), images, jupyter notebooks, markdowns, etc. using a Jekyll static site generator to minimize the tech stack hairballs.
The interesting part is that nearly all of the content is tagged to associate with two or more domains ex:
Goodreads-history-truecrime
PyTorch-Jupyter
Rubyonrails-Testing
Behavior-Gaslighting
Semiconductors-GPUs
And so on.
It’s an exercise in taxonomy creation. Searching by tag doesn’t quite get it. And we all know PDF auto summaries are tough.
I’m not expecting anybody to be impressed. It keeps me occupied while learning how to be (unfortunately) retired.
Earlier quoted context omitted.
Something like this has been on my mind for a while now -- take the useful, positive elements from across the socials (network of connections, media sharing, events, etc) and create mini-nets that let people who want to, stay in touch. How do you envision onboarding? Do I join, and then try to convince a handful of people to join as well?
Glad this resonates with you! That’s exactly the goal—keeping the useful parts of social networking while removing the noise and AI-driven manipulation. > How do you envision onboarding? Do I join, and then try to convince a handful of people to join as well? Yes, that's been the idea so far for onboarding. But we’re also exploring ways to make the platform more organically discoverable and valuable from day one (wit…
The need is there (at least for some of us!) so the sell shouldn't be so hard, but I feel like I'm missing the "a-ha!" differentiator here. It's not enough to pull the good/useful remnants from the sludge socials are today; it would need an extra something to excite people enough to make the effort to engage with yet another online service.
Finishing up my PhD thesis on low-resource audio classification for ecoacoustics. Our partners deployed 98 recorders in remote Arctic/sub-Arctic regions, collecting a massive (~19.5 years) dataset to monitor wildlife and human noise. Labeled data is the bottleneck, so my work focuses on getting good results with less data. Key parts: - Created EDANSA [1], the first public dataset of its kind from these areas, using a…
I'm working on pure.md[1], which lets your scripts, APIs, apps, agents, etc reliably access web content in markdown format. Simply prefix any URL with `pure.md/` and you get the unblocked markdown content of that webpage. It avoids bot detection and renders JavaScript-heavy websites, and can convert HTML, PDFs, images, and more into pure markdown. pure.md acts as a global caching layer between LLMs and web content. I…
Recently discussed, too: https://news.ycombinator.com/item?id=43462894 (10 comments)
I wrote a small business preventive maintenance and calibration tracker ( https://pmcal.net ) as a side project. A few manufacturing companies that I have a close relationship with are using it and love it, but I have kind of hit a wall with other growth avenues (Google Ads, organic promotion on the web). I have been thinking of marketing directly to ISO 9001 auditors, because “can you get email reminders” is a quest…
Nice. A calibration tracker is actually on my list of possible future projects :-) I'm surprised to see it here since it's not typical HN fare!
It's a worthwhile project to build yourself. If nothing else I found out that I definitely do not like the date-fns library in JavaScript. I built it using AWS Amplify, and although I like that it scales to zero, but I think there are too many gotchas to Amplify, and especially DynamoDB, for a startup app that you want to move quickly on. I wrote up one of the major ones after I got really frustrated. [1]
Like I said in my original post, I am trying to figure out how to get it in front of the right people (who are less likely to be on HN). I have kind of decided that the B2C sales experience is not great unless you get a critical mass; my experience doing sales in manufacturing is working the booth at trade shows, talking to people about engineering, and using our process tools to develop a solution to the customer's problem. The more scattered "compete for attention" advertising/promotion sales model doesn't seem great unless you have a lot of money behind it.
I'm rambling, but if anyone likes this or feels it needs a certain feature, feel free to reach out. If you're in Boston / Providence I'll happily grab a drink with you.
[1] https://gist.github.com/rchowe/1db32f1f26d74688a9b4083a19f6a...