Developing Practical AI Applications - Patrick Bangert

Data Transformers Podcast
Developing Practical AI Applications – Patrick Bangert
Play Episode
Pause Episode


Mute/Unmute Episode
Rewind 10 Seconds
1x
Fast Forward 30 seconds
00:00
/

00:30:40

Apple Podcasts

Google Podcasts

Spotify

Stitcher
RSS Feed


Share






Link


Embed

Developing Practical AI Applications – Patrick Bangert

‘
/>



Apple Podcasts



Google Podcasts



Spotify



Stitcher

Episode Title : Developing Practical AI Applications – Patrick Bangert

Episode Summary: The episode focuses on developing practical applications using AI. Patrick Bangert, as the head of AI Engg and AI services at Samsung SDS, is in charge of including AI in almost all Samsung applications that are deployed on Samsung phones. If anyone is interested in learning the various phases of developing AI applications, this is the episode. Patrick discusses the avrious phases such as developing models, training the models, and deploying the models. The episode also goes over the proper characteristics of data for developing good ethical and explainable AI applications.

Topics discussed in this episode:

Types of AI practical applications (02:00) : Samsung SDS (Samsung Data Services) within Samsung. SDS is It services company; AI Engineering – Distributed Training, Auto ML, Algorithmic questions; AI Services –  Data science and AI projects for specific use cases. Natural language models for talking to your phone or Facial recognition for identifying. Textual recognition so you can answer that question.

Cost of training AI models (06:00) : Most of AI companies create AI products for other businesses and not for individuals. The models are not created for individuals as it is very expensive. Example is GPT3 model as the new state of the art language model from Google. The cost to train the model was $5 million. Only large companies can spend that amount of money.

AI Diagnostic models for Covid (10:00) (POTENTIAL FOR AUDIOGRAM). Ptarick’s team took the challenge of fast tracking Covid results using X-rays. You can take an X-ray of a chest to find accuracy of Covid tests. Hospital donated X-ray images (15,000 images); 6 doctors to assess X-rays; Lot of effort to label the data.

Relevant and Representative Data (14:50) – ANOTHER AUDIOGRAM – What problem are you trying to solve. What accuracy needs to be there? Raw dataset – 15,000 images. Data needs to be relevant. Data needs to be representative; Representation of the situation and Representation of the problem; (2) Label the data; Doctors may not be accurate all the time; May be accurate 80%. Labelling data is the domain expert’s job and not the scientist’s job

Ethical AI (21:00) – Patrick is on the board of AI Ethics journal; Some bad examples:  UK – AI was used to grade students; Facial recognition software was developed using caucasians but used on Africam people; Problems with Ethical AI are two-fold (1) Not a representative sample (2) Used on incorrect applications. Typically a data set problem

Explainable AI (23:30) – One of AI drawbacks is that it is a black box; The price for automation, reduction in costs is that it is not inherently explainable.

Why is it so difficult to explain AI? (26:20) – Neural network is about multiplying matrices with vectors.It is difficult to explain how this 1,000 coefficient equation came up with a prediction. Humans simplify assumptions. Neural networks will not simplify assumptions and compare details with details.

Resources mentioned in this episode:

Podcast website: https://DataTransformersPodcast.Com

Data Transformers Podcast

Listen Now!

Join Peggy and Ramesh as they explore the exciting world of Data Management, Data Analytics, Data Governance, Data Privacy, Data Security, Artificial Intelligence, Cloud Computing, Internet Of Things.

Apple Podcasts



Google Podcasts



Spotify



Stitcher