Core principle of Machine Learning 

There of course are many, but for someone coming from computer science, and, software engineering, where the environment is relatively clean and certain (deterministic), it usually is a leap to understand that Machine Learning (and other elements of #AI) are not. 

Machine learning, is based on probability theory and deals with stochastic (non-deterministic) elements all the time. Nearly all activities in machine learning, require the ability to factor and more importantly, represent and reason with uncertainty. 

To that end, when designing a system, it is recommended to use a simple but uncertain (with some non-deterministic aspects)  rule, rather than a complex but certain rule. 

For example, having a simple but uncertain  rule saying “most birds fly”, is easier and more effective than a certain rule such as “Birds can fly, except flightless species, or those who are sick, or babies, etc.”

As one starts getting deeper in Machine Learning, a trip down memory lane around Probability distribution, expectation, variance, and covariance won’t hurt. 

HoloPortation – Limits of Human Kind

When it comes to AI and the limits of human kind, what better example that shows the art of the possible than what Microsoft is doing with special awareness and HoloLens and other sensors.

And not only can this replay time and allow you to have a ‘living memory’ but it also is mobile.

I do believe we are living in the great time ever! 🙂

Neural Networks

Of course you heard of Neural Networks! In the context of #AI they are all the buzz of course.

You might have heard of some such as DFF (Deep Feed Forward) or RNN (Recurrent neural networks)? Or perhaps you meant Recursive neural networks? Irrespective, it can be quite messy as you can see below and it would be somewhat important to have some understanding of the differences.

neuralnetworks

And in case you are thinking, well what good or use is all this? Here is one example ( MarI/O – Machine Learning for Video Games) that shows how a computer learned to play Mario using DeepMind and a Neural network.

MarI/O uses something called NEAT (neural evolution of augmenting topologies) and is written in Lua (which is very similar to .NET) and runs in BizHalk which is a emulator for games and their various platforms (and not to be confused with BizTalk). You can checkout the code for this here.

Fjodor also has outlined a (very) brief outline on what some of these are and what they mean. If you just want to get a quick basic understand it is a great read, with of course links back to original research papers (and deeper reads) if that is your cup of tea.

Happy reading! 🙂

Object and scene detection with #AI

Continuing the previous #ArtificialIntelligence theme. Wanted to see what and how does Amazon’s rekognition work and different from the #AI offerings from the others, such as Microsoft.

Here is a #ProjectMurphy image’s confidence score. I am glad to see that there is a 99% confidence that this is a person.

Object and Scene detection

The request POST is quite simple:

{
 "method": "POST",
 "path": "/",
 "region": "us-west-2",
 "headers": {
 "Content-Type": "application/x-amz-json-1.1",
 "X-Amz-Date": "Thu, 01 Dec 2016 22:21:01 GMT",
 "X-Amz-Target": "com.amazonaws.rekognitionservice.RekognitionService.DetectLabels"
 },
 "contentString": {
 "Attributes": [
 "ALL"
 ],
 "Image": {
 "Bytes": "..."
 }
 }
 }

And so is the response:

{
 "Labels": [
 {
 "Confidence": 99.2780990600586,
 "Name": "People"
 },
 {
 "Confidence": 99.2780990600586,
 "Name": "Person"
 },
 {
 "Confidence": 99.27307891845703,
 "Name": "Human"
 },
 {
 "Confidence": 73.7669448852539,
 "Name": "Flyer"
 },
 {
 "Confidence": 73.7669448852539,
 "Name": "Poster"
 },
 {
 "Confidence": 68.23612213134765,
 "Name": "Art"
 },
 {
 "Confidence": 58.291263580322266,
 "Name": "Brochure"
 },
 {
 "Confidence": 55.91957092285156,
 "Name": "Modern Art"
 },
 {
 "Confidence": 53.9996223449707,
 "Name": "Blossom"
 },
 {
 "Confidence": 53.9996223449707,
 "Name": "Flora"
 },
 {
 "Confidence": 53.9996223449707,
 "Name": "Flower"
 },
 {
 "Confidence": 53.9996223449707,
 "Name": "Petal"
 },
 {
 "Confidence": 53.9996223449707,
 "Name": "Plant"
 },
 {
 "Confidence": 50.69965744018555,
 "Name": "Face"
 },
 {
 "Confidence": 50.69965744018555,
 "Name": "Selfie"
 }
 ]
}

Here is what the facial analysis shows;

Facial Analysis

However how does it handle something a little more complex perhaps?

Object and Scene detection

And finally, what of the comparison? I think there might be some more work to be done on that front.

Face Comparison capture

Here is the response:

{
 "FaceMatches": [
 {
 "Face": {
 "BoundingBox": {
 "Height": 0.3878205120563507,
 "Left": 0.2371794879436493,
 "Top": 0.22435897588729858,
 "Width": 0.3878205120563507
 },
 "Confidence": 99.79533386230469
 },
 "Similarity": 0
 }
 ],
 "SourceImageFace": {
 "BoundingBox": {
 "Height": 0.209781214594841,
 "Left": 0.4188888967037201,
 "Top": 0.13127413392066955,
 "Width": 0.18111111223697662
 },
 "Confidence": 99.99442291259765
 }
}

Playing with #AI

So, been spending a lot of time recently around many things related to Artificial Intelligence (#AI).  More on that some day. 🙂

Was curious about yesterdays Amazon’s announcement to jump on this bandwagon. Of course Microsoft and others have been there. I don’t know to what extend has Amazon been working on this, but given Alexa has been out for a couple of years, I know they have had rich pickings of tuning this further.

I thought Polly (like the parrot?) was quite different from the things I have seen from others. This is a text-to-speech, where it renders the inputted text into various dialects and you can have a few outputs for those too. It supports a few dialects (for the synthesized speech) and one can use it using a simple API (the Android example shows it is not very complex to consume, of course you still need to think about the overall design and elements of Software Engineering, latency, limits, bandwidth, etc.). Should you desire you can customize it using pronunciation Lexicons that allow one to tweak this.

Here are a few examples, of course none of them are me, and hence the “cold”.

Australian (Male):

Indian (Female):

Italian (Male):

US/American (Male):

Of course if you play with it, it is easy to pick up the patterns and what is being changed, versus not. But kudos to the team on this. I think it will help accelerate the adoption of #AI.