Hacker News Reader

alexcnwyon Fastmail – 20 Years Old

A lot of people probably don’t know that Fastmail was started by Jeremy Howard, one of the creators of https://www.fast.ai which is by far the best way to learn deep learning IMHO :)

alexcnwyon BodyPix: Real-Time Person Segmentation in the Browser with Tensorflow.js

That’s so epic! Can you please give a little bit more detail on your process?

alexcnwyon Ask HN: What will become possible by 2025 in your field of expertise?

I have no doubt that deepfakes will be impossible to tell apart from real videos (at least to a human) in 3-5 years.

In 10 years I think we’ll be able to generate entire movies programmatically.

alexcnwyon Can you sing like Freddie Mercury?

This is so cool!

From the page: "Your audio doesn’t get uploaded to servers to be analyzed, so your vocals stay private."

From their ProductHunt launch: "We made an AI-powered singing challenge to find out. This experiment uses new, on-device machine learning models to see how close your timbre, pitch, and melody are to Freddie’s."

alexcnwyon Announcing .NET Jupyter Notebooks

Very cool, thanks for the detailed reply - your JS tutorial looks great :)

Yeah automatic re-execution sounds like a weird design choice. Think I'll stick with Jupyter notebooks for now...

alexcnwyon Announcing .NET Jupyter Notebooks

Yeah I remember seeing that one a few years ago - it's also super cool :)

Especially crazy that these work as well and quickly on mobile!

alexcnwyon Announcing .NET Jupyter Notebooks

Mike Bostock is a beast - this d3.js example from 2013 still blows my mind: https://archive.nytimes.com/www.nytimes.com/interactive/2012...

I haven’t gotten properly stuck into observable - in what ways is it better than using a jupyter notebook?

alexcnwyon When your data doesn’t fit in memory: the basic techniques

There's a lot of room between a laptop with 16GB of RAM and a "Big Data cluster" - in my experience the easiest solution is to spin up a VM in GCP and crank the RAM way up.

It blew my mind when I realized I could add / remove RAM by turning off the instance, dragging a slider, then turning it back on.

I also find h5py really useful for creating massive numpy arrays on disk that are too large to fit in memory. I used it to precompute CNN features for a video classification model (much faster than computing on each gradient descent pass) and it makes it easy to read/write parts of a numpy array when the entire numpy array is too big to fit in memory.

alexcnwyon Modern Data Practice and the SQL Tradition

In my experience it's a bit of both.

Most of a data scientist's time is definitely spent getting data from various places, possibly enriching it, cleaning it, and converting it into a dataframe that can then be modeled (and surprisingly a lot of those tasks are often beyond pure actuaries / statisticians who often get nervous writing SQL queries with joins nevermind Spark jobs)...

But also once you've got that dataframe, it's often too big to be processed on a single machine in the office so you need to spin up a VM and move the data around which also requires a lot more computer science / hacking skills than you typically learn in a stats university degree so I think the term data scientist for someone who can do the computer sciency stuff in addition to the stats has its place...

alexcnwyon Modern Data Practice and the SQL Tradition

The ability to re-size GCP VMs totally blew me away when I first discovered it. Just power off the machine, drag the RAM slider way up and turn it back on. SOOO much easier than re-creating the VM on AWS.

I also way prefer to just crank up the RAM on a single instance and use Pandas/Dask instead of dealing with distributed computing headaches :)

alexcnwyon Modern Data Practice and the SQL Tradition

I totally get your skepticism but the bottom line is that throwing computing power at a problem generally leads to a much better solution even if only because you can do a much more extensive grid-search. This doesn’t have to be using some complex cluster like you could just have n copies of the same VM that each run experiments with parameters pulled from a Google sheet until all the experiments are done.

Sure, the cloud providers sell computing power but it’s a race to the bottom and makes much more sense than buying hardware for these kinds of bursty analytics workloads.

I don’t think “Big data” from a stats perspective is analogous to census data - in some cases yes but for applications like recommendation engines you lose a lot of valuable signal by sampling.

alexcnwyon Modern Data Practice and the SQL Tradition

The big difference is that historically you could run pretty much all analysis on a single computer (if not in Excel - I’ve seen and built many actuarial models in Excel myself) but consumer internet scale data (both in terms of the number of rows and the number of features available) is too big to model without distributed computing and more programming work on the part of the analyst.

Moreover it’s an iterative process where you start with hypotheses on what features will be useful then build a model and might need to pull more data. For example you could build a recommendation engine just using the set of products purchased by customers but then might explore adding user agent features like devices or might add data about the products like category or price information.

An analyst (“data scientist”) who can pull and interact with that data is much more effective than one who needs to wait for someone else to pull the data for them (most actuaries I know are weak programmers outside of R and wouldn’t know where to begin using distributed computing or getting a TensorFlow gridsearch running on a cluster of cloud GPUs :)

alexcnwyon We Asked People Around the World How They Feel About AI. Here’s What We Learned

The impacts in improved translation aren’t marginal - they improve people’s lives in real ways.

Advances in deep learning for NLP have also significantly improved search (Google and others) which has a nontrivial impact on people’s lives.

Multi-modal models that translate images into words have a significant impact for blind people.

Deep learning is significantly improving medical diagnosis, crop yield, assistive robotics, and many other specific scenarios.

Specific scenarios matter and again, it’s still very early days...

alexcnwyon We Asked People Around the World How They Feel About AI. Here’s What We Learned

A lot of the breakthroughs have been in image classification, object detection and image segmentation using deep convolutional models way way outperforming prior approaches and beating human level benchmarks (on constrained problems e.g. classify this image into one of 1000 categories)...

But there have also been significant breakthroughs in:

* image/video generation using generative adversarial networks

* game playing & robotics in constrained environments (AlphaGo, AlphaStar, etc) using reinforcement learning

* natural language processing using semantic word vectors and attentional recurrent models

* voice-to-text using recurrent models

Convolutional and recurrent neural networks were invented in the early 1990s and we’ve only recently got them properly working as tools. I think we’re going to see (and need) more algorithmic breakthroughs to push through the current limitations but I’d be very reluctant to bet against that happening, even if it takes a long time...

alexcnwyon Show HN: OpenCV based multi-object-tracker with various object detection models

I've built several object detection systems and you definitely want to fine-tune your model to get good results, even if the object class you want to track is available in a pre-trained model (often it isn't).

For example, pre-trained models don't work too well at detecting people in surveillance footage when the camera is mounted on the ceiling and people are at an angle in the footage. Off-the-shelf object detection models also don't work well in poor lighting conditions but can work well if you fine-tune one of the pre-trained models for your dataset and lighting conditions.

Here's a great and fun tutorial on how to fine-tune an object detection model using TensorFlow (it's a surprisingly fiddly process but worth the headache): https://towardsdatascience.com/how-to-train-your-own-object-...

I suspect you can use your fine-tuned model with this repo by replacing the `get_ssd_model.sh` compiled graph with your fine-tuned model graph.

It's also a fun exercise to implement the tracking yourself - here's a solid tutorial to get you started: https://www.pyimagesearch.com/2018/08/13/opencv-people-count...

alexcnwyon Few-Shot Video-to-Video Synthesis

The flicker would be a challenge but you can use something like DeOldify [1] or super resolution [2] on the frame sequence to improve the colour and resolution... although you may get some weird artifacts between frames using image models rather than models built for video.

Generally video models are quite far behind image models because they're so much higher dimensional (much more computationally expensive and expensive to annotate) but I'm sure it's a matter of time before someone releases a DeOldify type model for video (if one doesn't exist already).

[1] https://github.com/jantic/DeOldify [2] https://github.com/idealo/image-super-resolution

alexcnwyon Visualizing Music with GANs

This is so incredible and well executed. It's crazy to see how quickly GANs are moving (e.g. check out this tweet [1] by Ian Goodfellow on 4.5 years of GAN progress) ... excited to see what they can do a few years from now!

[1] https://twitter.com/goodfellow_ian/status/108497359623614464...

alexcnwyon DeepSORT: Deep Learning to Track Custom Objects in a Video

Often occlusions don’t matter in practice for dense object tracking e.g. counting people entering a bus or measuring gym equipment utilization you actually don’t care if the person IDs switch.

Centroid tracking works well in practice at 2 FPS for the bus use case and several others.

More complex approaches have their place but an important limiting factor (e.g. in the bus project I worked on) is that often the hardware required to run a more complex approach at a higher frame rate is too expensive to be feasible.

There are definitely limitations to centroid tracking and use cases where it doesn’t work like if you need to track specific faces moving through a crowd but it’s a useful tool that has its place in practice.

Your blog post is awesome, I’m just pointing out that a simpler method often works well in practice and is much less intimidating for beginners :)

alexcnwyon Ask HN: Successful one-person online businesses?

Super interesting, thanks for the reply! :)

alexcnwyon DeepSORT: Deep Learning to Track Custom Objects in a Video

Yes, of course!

Essentially centroid tracking works like this:

   1. Detect objects (e.g. using TensorFlow) in each video frame 

   2. Initialize an incrementing ID for each object at the first frame 

   3. Compute the centroid of each object 4. For each object: 

   4.1. Compute the Euclidean distance to every object centroid in the previous frame. The nearest object from the previous frame is the candidate assignment. 

   4.2.a) If the distance is less than a hand-set threshold, use the object from the previous frame's ID for the object in the current frame. 

   4.2.b) If the distance is above the threshold, assign a new ID

I explained it during my PyConZA keynote a few weeks ago. The vid isn't out yet but the slides are available here and it's much easier to explain visually: http://bit.ly/pycon_keynote_slides

Feel free to email me if you get stuck or need some help implementing it :)

alexcnwyon DeepSORT: Deep Learning to Track Custom Objects in a Video

This is an interesting and well-written post but in my experience for practical object tracking problems using a fine-tuned object detection model with a centroid tracking algorithm works well enough. Much more computationally efficient and you only have to annotate object examples rather than frame-by-frame object tracking.

I would have loved to have seen a benchmark comparing the two approaches to motivate the much more complex and labour intensive alternative ':)

alexcnwyon Ask HN: Successful one-person online businesses?

Very cool and super inspiring, well done!

Can you give any detail on how you're getting customers... you mention learning google campaign management, is all your traffic via google or have you found any other distribution channels that work (e.g. posting on Fiverr forums yourself kind of thing?)

alexcnwyon Ask HN: What's a promising area to work on?

I’m building a tool to speed up data entry. It highlights OCR‘d words with low confidence for human review. It uses templates for capturing structured data and a fast interface for capturing unstructured data.

It grew out of consulting projects and is almost ready for beta testing if anyone is interested in playing with it and giving me some feedback - my email is in my profile :)

alexcnwyon Ask HN: What's a promising area to work on?

> OCR doesn't need a cloud vs. a beefy computer - I do a lot of that too.

Can you elaborate on this?

alexcnwyon Ask HN: How do you share/organize knowledge at work and life?

I use an outliner called Dynalist.

It’s an alternative to Workflowy. Workflowy development stalled a while ago so I switched but it seems to be active again so I might switch back.

An infinitely zoomable hierarchy works really well for my OCD :)

alexcnwyon Principal Component Analysis

Excellent explanation!

I find this visualization really helpful when I need to explain PCA: http://setosa.io/ev/principal-component-analysis/

alexcnwyon Deep Learning: Our Miraculous Year 1990-1991

1989/1990 was also when convolutional networks first started working with LeCun’s breakthrough paper on digit recognition.

Incredible to think how much amazing research was happening back then and wonder what research is being done now that will change our lives in the next 30 years.

alexcnwyon Show HN: Using ML to detect key changes to 2020 Candidates' Websites

Very cool idea.

It's not clear where you're using ML from the site - URLs seem chronological, not ranked on "relevance/importance" and I can't see any relevance/importance indicators.

I'm curious to hear some more detail on how you're encoding your visual diff model?

BTW I ran into 2 issues: 1. You can't zoom out on your image diff slider 2. I got this error when I returned to the site after closing it: {"crossDomain" : true, "method" : "GET", "url" : "https : //api.fluxguard.com/publi c/site/7f646558-f754-447a -b627-9b5202c8a1f2/page?l imit=10&publicAccount=cam paignmonitor"} Please contact us if this is error is happening frequently for you.